System and method for dynamic profiling of users in one-to-one applications and for validating user rules

11249/00 ΠΙΠΊΝ7ΐ "7ΠΝ*? 7Τ1Ν" ϋ^ΠΠΙΠΠ 7ΐΰ Tri ^3173 Πϊ?3ΐϊ7 ΓΏΊΟΠΙ Πϋ’ϋ) ωητιωη

SYSTEM AND METHOD FOR DYNAMIC PROFILING OF USERS IN ONE-TO-ONE APPLICATIONS AND FOR VALIDATING USER RULES WO 99/26180 PCT/US98/24339

SYSTEM AND METHOD FOR DYNAMIC PROFILING OF USERS IN ONE-TO-ONE APPLICATIONS AND FOR VALIDATING USER RULES

FIELD OF THE INVENTION

The present invention relates to a system and method for dynamic profiling of a user in one-to-one marketing applications . 5

BACKGROUND INFORMATION

Many organizations collect historical data about every transaction that every customer performs with that organization. Such historical transactional data is useful in 10 various one-to-one marketing applications, such as, e.g., shopping assistant application and dynamic Web site content presentation. A number of problems have been encountered in these marketing applications. One such problem relates to the creation of highly pertinent and comprehensible individual 15 user profiles that are derived from the historical transactional data. In addition, it is also important to have the ability to utilize these user profiles when the marketing application obtains a current status of the user. If the user profiles are generated in a highly relevant and comprehensible 20 manner with respect to a specific user, the applications would Le able to understand that user's needs better and more efficiently serve that user.

There are two basic types of user profiles that can be generated - a "static" profile and a "dynamic" profile. 25 The static profile contains all of the factual information of the user including, for example, demographic data (e.g., age, sex, address), psychographic data (e.g., personality traits and habits), purchasing preferences (e.g., what does the user purchase in an average week), etc. Static profiles are 30 generated using conventional methods that are known to those of ordinary skill in the art.

Dynamic profiling information includes specific rules describing the user's behavior. For example, such rules may include: "whenever user X travels to France, user X often 35 buys expensive wines in Paris" or "when user Y shops on a WO 99/26180 PCT/US98/24339 weekend and did not buy any groceries for at least 3 days, user Y usually purchases a large amount of groceries." These rules can be generated with transactional data for each user using various rule generation methods that are generally known 5 to those of ordinary skill in the art. For example, one such conventional rule generation method is implemented in a rule learning system which generates behavior rules for individual customers. (See T. Fawcett et al., "Combining Data Mining and Machine Learning for Effective User Profiling", Proceedings of 10 the KDD'96 Conference, 1996, pp. 8-13).

In order to obtain an extensive understanding of the user, it is desirable to build both static and dynamic profiles for that user. Although the generation of static profiles is generally straight-forward, generating dynamic 15 profiles for a large number of users may present potential problems. Many transactional systems (e.g., airline reservations systems, credit card transactional systems and/or Web site management systems) generate a various number of transactions for each user. For example, some systems and/or 20 applications may only generate a dozen transactions per each user, which may not be enough to construct a statistically significant and reliable set of rules for a specific user.

Even if there are enough transactions to construct a statistically significant set of rules, these rules should 25 still be verified for their pertinence to the user. Since there can be a large number of users, and since the rules generated for each user may not be reliable, there is a problem of verifying a large set of generated rules for the users. For example, in a typical system facilitating 5 30 million users and providing about 100 rules per user, approximately 500 million rules would have to be either stored or processed. Generally, many of these rules are either not useful or insignificant. Thus, due to the amount of these generated rules, a rule validation process becomes 35 considerably complicated. Furthermore, checking the usefulness of these rules "by hand" becomes practically impossible. WO 99/26180 PCT/US98/24339

Conventional systems have not successfully provided detailed solutions to constructing reliable dynamic profiles for the users. One such system (described in T. Fawcett et al., "Combining Data Mining and Machine Learning for Effective 5 User Profiling", Proceedings of the KDD'96 Conference, 1996) provides a limited generation of user's dynamic profiles. However, this conventional system does not provide a comprehensive method and system for analyzing a large number of dynamic rules, and thus does not provide adequate 10 assistance for the user.

SUMMARY OF THE INVENTION

The system and method according to the present invention generates dynamic profiles and, thereafter, 15 transforms the dynamic profiles for various users into aggregate rules. In particular, "similar" individual rules are compressed into a smaller number of aggregated rules. Because the total number of aggregate rules is substantially smaller than the total number of individual rules for all of 20 the users, the aggregate rules can be examined manually by a human expert. This expert examines these aggregated rules and selects only rules based on the expert's preferences. Only the individual rules that correspond to the aggregated rules selected by the human expert are retained in the user's 25 profiles. Since the selected aggregate rules were selected by the human expert, a creation of more accurate dynamic profiles is further assured. The system and method according to the present invention thus provide a more useful set of individual rules for each user. 30 The dynamic profiles generated with the system and method according to the present invention can be used in various systems (e.g., Personal Shopping Assistant and Personal Intelligent Digital Assistant) to provide better recommendations to the users as to which products and services 35 each individual user should utilize.· Accordingly, the user would be more satisfied with these systems and the suggestions that these systems provide to the user. In addition, Dynamic WO 99/26180 PCT/US98/24339

Web Content Presentation systems can include the system and method according to the present invention because the users will be provided with better quality profiles to facilitate the provision of more pertinent Web pages to the user visiting 5 a particular Web site. Fraud detection systems may also include the system and method according to the present invention, thus providing higher quality user profiles which may facilitate better fraud detection. Other applications for the system and method according to the present invention are 10 also conceivable to those of ordinary skill in the art.

In addition, the system and method according to the present invention utilizing the above-described rule compression method is not limited to a construction of pertinent dynamic profiles, and can be used in a vast variety 15 of applications (e.g., construction of high quality association rules in data mining applications). Other usages of the system and method according to the present invention are also conceivable to one having ordinary skill in the art.

In another embodiment of the system and method 20 according to the present invention, the user rules are validated using a processing device. These user rules are retrieved from a storage device. The user rules are then separated into at least one subset of a user set. Then, it is determined if particular rules of the at least one subset is 25 one of acceptable, unacceptable and undecided based on a defined criteria. If the particular rules of at least one subset are acceptable, the particular rules of the at least one subset are provided to a corresponding user.

When constructing good dynamic profiles, validation 30 is an important consideration. It is possible to construct dynamic profiles for individual customers using any existing data mining methods. For example, an exemplary user rule may indicate that whenever a user buys milk in the evening, this user also buys onions. It is difficult to ascertain if this 35 rule adequately describe the user's behavior. In fact, it may be a statistical coincidence. Therefore, it is preferable to 4 WO 99/26180 PCT/US98/24339 allow the user (or a human expert) to examine groups of the user rules. BRIEF DESCRIPTION OF THE DRAWINGS 5 Fig. 1 shows a top level diagram of a process for generating user profiles.

Fig. 2 shows a flow diagram for generating static and dynamic user profiles.

Fig. 3 shows a flow diagram of a process for 10 compressing dynamic rules, generating aggregate rules, validating the aggregate rules and creating user profiles.

Fig. 4 shows a detailed flow diagram of an exemplary rule compression process according to the present invention.

Fig. 5 shows a detailed flow diagram of an exemplary 15 cluster compression process according to the present invention .

Fig. 6a shows an exemplary system for generating user profiles according to the present invention.

Fig. 6b shows a first system for generating the user 20 profiles according to the present invention as illustrated in Fig. 6a.

Fig. 6c shows a second system for generating the user profiles according to the present invention as illustrated in Fig. 6a. 25 Fig. 7 shows a block diagram of an exemplary

Personal Intelligent Digital Assistant system according to the present invention.

Fig. 8 shows a flow diagram of another embodiment of the process according to the present invention in which 30 individual user rules are selectively validated using a selective validation module.

Fig. 9 shows a flow diagram of an exemplary embodiment of a process executed by the selective validation module (illustrated in Fig.8). 35 Fig. 10 shows a flow diagram of an exemplary procedure to generate the attribute hierarchy and to provide a cluster operation. WO 99/26180 PCT/US98/24339

Fig. 11 shows a detailed illustration of an exemplary procedure to generate "Cut" data.

Fig. 12 shows an exemplary procedure for grouping subsets using the "Cut" data. 5 Fig. 13 shows an exemplary attribute hierarchy which can be utilized with this embodiment of the process and system according to the present invention.

Fig. 14 shows an exemplary illustration of a first level extension and a second level extensions of an exemplary 10 node/group illustrated in Fig. 13.

Fig. 15 shows an exemplary implementation of the process and system according to this embodiment of the present invention as illustrated in Figs. 8 and 9.

15 DETAILED DESCRIPTION OF THE INVENTION

In many customer-related applications (e.g., banking, credit card, Internet marketing applications, etc.), user profiles for each user (or customer) are generated to better understand the user (i.e., user's purchasing trends, 20 business travel locations, types of favorite restaurants, etc.). A flow diagram of an exemplary process for building user profiles is illustrated in Fig. 1. In particular, information regarding, e.g., the user's past purchasing history is retrieved in step 1. In step 2, user profiles are 25 built, and the process completion is signaled in step 3. User profiles can preferably be generated using static profiles and dynamic profiles. A more detailed flow diagram of the process of building user profiles (represented in Fig. 1 by step 2) is illustrated in Fig. 2. The static profile includes user 30 static characteristics (e.g., name of the user, address, telephone number, date of birth, sex, income, etc.). The static profile is built in step 10 using methods known to one having ordinary skill in the art. After the static profile is built, this static profile is stored in a separate file based 35 on the data obtained from the CUST and TRANS files, as discussed below. The "CUST" file has the following format: 6 WO 99/26180 PCT/US98/24339 CUST (Cust_ID, Aj, A2 ... AJ , where Cust_ID is a user identifier that provides an index value for locating a specific user in the CUST file. A1, A2 5 . . . are fields describing the characteristics of the user (e.g., sex, income, education, etc.).

The dynamic profile is built in step 15. A dynamic profile consists of rules (or patterns) characterizing a user's behavior, e.g., "if user X shops in the evening on 10 weekdays and purchases diapers, user X also buys beer", "if user X shops on weekdays, user X usually buys a small number of items", "if user X travels to New York on business, user X prefers to have lunches at expensive seafood restaurants."

The rules are derived from a set of transactions pertaining to 15 a particular user. These transactions may be, for example, credit card transactions, airline reservations and Web site visit transactions, and are stored in the "TRANS" file which has the following format: 20 TRANS (Trans_ID, Cust_ID, Cj, C2, ... Cn) where Trans_ID corresponds to a unique index key that identifies the transaction being performed by the user.

Fields Clf C2, ... Cn identify a particular transaction (e.g., 25 date of transaction, time of transaction, amount spent, location of the transaction, etc.). The field "Cust_ID" corresponds to an index key pointing to a particular user having a respective record in the CUST file. Thus, the user performing a particular transaction can be identified. 30 Other file formats can also be utilized, as can be understood by those having ordinary skill in the art. For example, the user-specific information can also be stored in several files rather than in a single CUST file (thus, the CUST file can be normalized by splitting the CUST file into several smaller 35 files). Using different file formats does not affect the operability of the system and process according to the present 7 WO 99/26180 PCT/US98/24339 invention. After the dynamic profile for a particular user is generated, this dynamic profile is validated in step 20.

After the validation of the dynamic profile, the static and dynamic profiles are combined to form a combined 5 user profile in step 25. The following exemplary information can be obtained from the TRANS file to be provided into the static profile when the static and dynamic profiles (the CUST and TRANS files) are combined: a) an average transaction amount for user X; b) user X's favorite brand of beer is, 10 e.g., Heineken; c) user X shops mostly on week-ends.

While it is relatively uncomplicated to construct user static profiles, it is much more difficult to construct quality dynamic profiles. Rules provided in the dynamic profile are generated for each user. Because a user may 15 perform only a small number of transactions, the corresponding rules generated may be statistically insignificant, unreliable and irrelevant. In many systems (e.g, airline reservations systems, credit card transactional systems, or Web site usage systems) , it is possible to have from as little as a few dozen 20 to a few hundred transactions per each user. The rules generated with such amounts of data are often ineffective and insignificant.

The total number of generated rules can also be very large. For example, in a system serving 5 million customers 25 and generating an average of 100 rules per user, a total number of generated rules can reach 500 million. Many of the 500 million generated rules are of questionable quality and usefulness. In order to filter the rules having such undesirable characteristics, a human expert must decide which 30 dynamic rules should be stored and which dynamic rules should be discarded. It would be impossible for the human expert to manually check the usefulness of all 500 million rules.

Quality dynamic profiles are generated by validating dynamic rules generated using various rule induction methods. 35 Ultimately, however, the human expert validates the machine generated rules to determine their "usefulness" in various systems. Since most of the systems generate too many rules to WO 99/26180 PCT/DS98/24339 be manually examined by human experts, the system and method according to the present invention facilitates compressing individual rules into "aggregated" rules. After the individual rules are compressed into the aggregated rules, the 5 aggregated rules are evaluated by a human expert who selects only the rules that the expert believes are pertinent for the user. In addition, it is possible (in some applications) that the respective user can be such a human expert (and examining only the rules that are pertinent to the respective user). 10 A. Dynamic Profile Construction Procedure

It can be assumed that user-specific rules have been already created using methods known to those having ordinary skill in the art. For example, individual user rules can be 15 generated using an induction software system (e.g., "CART"

Breiman et al., 1984; C4.5, Quinlan, 1993; or RL, Clearwater & Provost, 1990) . The structure of these rules has, preferably, the following form: 20 Ciieilail λ Ci20i2ai2 Λ · · · Λ Cik0ikaik =3 C1diai (1) where Cu, Ci2, ..., Cik, Cx are fields from the TRANS file, aix, ai2, ..., alk, ax are constants, and are relational operators (e.g., , "<", etc.). In addition, each rule 25 is assigned to a user defined by the Cust_ID (user identifier) from the CUST file.

Next, it is important to remove "useless" individual rules from the total number of rules. A process to remove these useless individual rules is shown in Fig. 3. In step 30 30, individual rules are provided for processing. In step 35 several "similar" individual rules (of the form (1)) are compressed into one aggregated rule of the form: λ Ai20i2bi2 λ ... λ A1j^ijbij λ 35 Λ ^-i2®i2^i2 Λ ... Λ Cxkdikaik =» (2) 9 WO 99/26180 PCT/US98/24339 where Aix, ..., Ai;j are the fields in the CUST file, b^, ..., bi:j are constants, and Θ i5 are relational operators (e.g., " = ", etc.). For each individual rule of the form (1), the aggregated rule of the form (2) is formed after the 5 individual rules are compressed. The newly aggregated rules (formed in step 40) can be, e.g., fuzzy rules, and the operators Θshould also be, e.g., fuzzy operators. For example, several of the individual rules that are similar (generally pertaining to different users) can be compressed 10 into one aggregated rule pertaining to the same subject matter that can be applicable to several users. For example, if several rules have the form: IF Shopping_time = "evening" and Day_of_week = "weekday" and Purchase = "diapers" THEN Purchase = 15 · "beer", and it is known that most of the users corresponding to this rule are males, then these rules can be compressed into the aggregated rule having the following form: 20 IF Sex = "Male" and Shopping_time = "evening" and

Day_of_week = "weekday" and Purchase = "diapers" THEN Purchase = "beer".

Additional fields (e.g., Sex, etc.), unlike other 25 fields in the above exemplary rule, are fields from the CUST file. Individual rules relating to different users can be compressed into the same aggregated rule for a group of users. Thus, the rule compression can preferably be implemented for different users. The number of aggregated rules (of the form 30 (2)) generated by the compression algorithm should be much smaller than the initial number of individual rules. Then, in step 45, the aggregated rules can be validated (one by one) by the human expert (including a particular user) to determine which rules are appropriate for that user. After the user 35 validates the aggregated rules, this user selects the set of preferred aggregated rules in step 50. Only the individual rules corresponding to the aggregated rules selected in step 10 WO 99/26180 PCT/US98/24339 50 are retained in the user's dynamic profile (step 55) to provide validated individual rules (step 60) to the user. B. Rule Compression Process 5 Fig. 4 illustrates a detailed flow diagram of an exemplary rule compression process (starting from step 35 in Fig. 3). Two individual rules of the form (1) are referred to as "similar" rules if they differ from each other only in the values of their respective constants ai3. Thus, similar rules 10 should have the same number of terms, the same fields C , and the same comparison operators 0i;j. Similar rules can be mapped into the (k+1) dimensional space defined by Dom(Cix) x...x Dom(Cik) x DomiCj, where Dorn is a domain (or range of values) of the field C, with a rule having the form (1) being mapped 15 into the points (i.e., au, ai2, ..., aik, ax) . This set of points is generated by similar rules. For example, the rule "if user X shops in the evening on weekdays and purchases diapers, user X also buys beer" can be written as: IF (Shopping_time = "evening" and Day_of_week = 20 "weekday" and Purchase = "diapers") THEN Purchase = "beer".

This sample rule would be mapped into the corresponding vector ("evening", "weekday", "diapers", "beer") of the 4-dimensional space of attributes (variables):Shopping_time, Day_of_week, 25 Purchase and another Purchase.

The exemplary rule compression process (described below in detail) then generates rules (e.g., fuzzy rules of the form (2)). These fuzzy rules utilize fuzzy linguistic variables for the fields from the CUST and TRANS files, which 30 are generally known to those having ordinary skill in the art.

Each fuzzy linguistic variable has a corresponding identifier (e.g., Income, Transaction_Amount, etc.), each being capable of providing a range of values (e.g., natural numbers between 0 and 1,000,000), a set of terms (e.g., "low", "medium", 35 "high", etc.), and a membership function that assigns a membership value (e.g., between 0 and 1) to each value from the domain of the fuzzy linguistic variable for each range of 11 WO 99/26180 PCT/US98/24339 10 15 20 25 30 35 values. In addition, the non-ordered fields in the CUST and TRANS files (e.g., "Product_Purchased") have assigned classification hierarchies; for example, the field "Product_Purchased" can include standard classification hierarchies used in marketing. Thus, UPCs, e.g., can be grouped into brands, brands can be grouped into product categories, etc.

The following exemplary inputs are provided to the Rule Compression Process: a. Individual rules from users' dynamic profiles. b. Fuzzy linguistic variables for all fields in the CUST and TRANS files. c. Hierarchical classifications for non-ordered fields.

Exemplary outputs generated by the Rule Compression Process are a set of (preferably) fuzzy aggregated rules having the form (2) .

The steps of the exemplary rule compression process shall now be described in detail with reference to Fig. 4. In step 160, all the individual rules of the form (1) are grouped into sets of similar rules (i.e., rules having the same structure) . The maximal number of such similar groups is 4~, where n is the number of fields in the TRANS file. For example, if n = 10, then there can be at most 1 x 220 similar groups. However, this number is typically much smaller in practice. Each set of similar rules forms a set of points in a k-dimensional space generated by the individual rules described above. In step 165, a group of clusters of the generated points is determined using any of the cluster computation methods known to those of ordinary skill in the art. In step 170, starting from the first cluster of the group of clusters determined in Step 165, an approximate rule for that cluster is determined in Step 180. The approximate rule is determined as a function of the points in the cluster. 12 4 '> WO 99/26180 PCT/US98/24339

For example, a point in the cluster may be the "center" of the cluster. The center can be identified as a point that minimizes the sum of distances from a particular point to other points in the cluster. For example, given 5 Cluster Cx = (clX, ci2, ... cik) , the center of this cluster is the point that minimizes the expression: min Σ d(x, c_.) c. xeClust. 2 2

The center of the cluster can also be determined using other 10 methods, such as, e.g., selecting the most "representative" point in that cluster.

In step 185, the next cluster is selected and the procedure described with respect to step 180 is repeated. In step 175, it is determined whether all of the clusters in the 15 group of clusters have been evaluated. As an illustration, if cluster Clustj contains three 3-dimensional points (0,0,1), (0,1,0), (1,0,1), corresponding to the vertices of a equilateral triangle, then the center of this cluster Cx is the center of the triangle, i.e., the point (^,^,½). Other 20 approaches to defining the center of a cluster can be used.

This exemplary rule compression process does not depend on any specific method for defining any center of a cluster C\.

Given the set of rules (1) corresponding to the cluster with the center C\ = (c.;, ci2, ... cik) , the 25 corresponding aggregated rule has the form: Λ Ci20i2Cl2 A ... A >-lk0ikCik CiACi (3) which is a "representative" rule for the cluster of similar 30 rules. For example, if the center of the cluster is (^,^,½), the following rule is generated: C, = K a C, = K => C3 = fc.

Also, for totally ordered fields CXJ, standard deviations σχί of the points in that cluster are calculated for that field. For 13 WO 99/26180 PCT/US98/24339 unordered categorical fields , a measure of the "deviation" of the points is determined in that cluster along the j-th dimension from οχί (by using the hierarchical classification for that field). 5 In step 190, a total number of clusters generated in

Step 165 is provided to the user. In step 195, the user is asked if there are too many of the generated clusters for manually inspecting the aggregated rules (i.e., the number of generated clusters is greater than a predetermined number). 10 If so, the generated clusters are compressed using a cluster compression process described in step 205. Thereafter, there is a smaller number of clusters (and corresponding aggregation rules per cluster). The user is asked again, in step 195, if there are too many generated clusters for the manual 15 inspection of aggregated rules. If the number of clusters is smaller than the predetermined number, for each cluster Cx obtained in step 165 or in step 205, a set of users corresponding to the points for that cluster is identified in step 210. Each point in a cluster corresponds to a first 20 representative rule from the dynamic profile of the user, so that all of the users corresponding to the dynamic profile rules from that cluster can be identified. For example, CUST-IDx is defined as a set of values Cust_IDi:j corresponding to the users corresponding to the rules of cluster Οχ. A set 25 of records ("CUST/') from the CUST file corresponding to the users of that cluster is identified (i.e., having user ID values from the set CUST_IDJ . Thus, CUSTX = { r j CUST(r) and r.Cust_ID € CUST_IDx }.

The set of records CUSTX form a set of points in m- 30 dimensional space (where m is the number of fields in the CUST file). These points are separated into clusters using the same techniques as described in step 165. For each resulting cluster CUSTi;j, a center is located as explained below. The set of points belonging to that cluster is approximated with a 35 logical statement having the form: Λ A^H.jb^j Λ · · · Λ Am0xjmbxjm (4) 14 < WO 99/26180 PCT/US98/24339 to form a corresponding condition in step 215, where A± are the fields of the CUST file, 0,,, are relational operators (e.g., " = ", , etc.) and bi;jl are constants. One way to construct the condition (4) would be by finding the center b13 5 = (bxjl, ..., bijra) of the cluster CUST υ as described in step 180, and substituting the values of bi;jl into the condition (4) (also setting all the relational operators to be " = ") .

Another way to construct this condition (4) is described in A. Motro, "Using Integrity Constraints to Provide Intentional 10 Answers to Relational Queries", Proceedings of the 15th

International Conference on Very Large Databases, 1989, pp. 237-246, and C. Shum et al., "Implicit Representation of Extensional Answers", Proceedings of the 2nd International Conference on Expert Database Systems, 1988. 15 In step 220, the first and second representative rules are augmented (i.e., expression (4) is augmented with expression (3)). The resulting rule is: Λ A20ij2bij2 A . . . A A^^i^bjjn, A Cil0ilcil λ 20 ci2ei2ci2 λ ... λ clkelkclk = c^e^Ci. (5)

For example, assume that the center of a cluster is a rule: "if a user shops in the evening on weekdays and buys diapers, the user also buys beer" (i.e., IF Shopping_time = 25 "evening" and Day_of_week = "weekday" and Purchase = "diapers" THEN Purchase = "beer"). Also, assume that most of the users in that cluster are men, thus forming the expression (4) where "Sex" = "Male". Accordingly, the augmented rule is "if a male user shops in the-evening on weekdays and buys diapers, the 30 user also buys beer" (i.e., IF "Sex" = "Male" and "Shopping_time" = "evening" and "Day_of_week" = "weekday" and "Purchase" = "diapers" THEN "Purchase" = "beer").

Then, in step 225, the rules of the form (5) generated in step 220 are converted into fuzzy aggregated 35 rules. In particular, each field At and Ci:j in the form (5) is mapped into a corresponding fuzzy linguistic variable associated with that field. In addition, all of the terms in 15 WO 99/26180 PCT/US98/24339 • the expression (5) are converted into appropriate fuzzy expressions. For example, assume that a non-fuzzy term Ax = 20 corresponds to a fuzzy linguistic variable also denoted as A,.

Further assume that the term set for A3 is either low or high, 5 and that there is a membership function that assigns the membership value (e.g., between 0 and 1) to each value from the domain of fuzzy term A3 for each value from the term set. Then, it can be determined for which term (i.e., low or high) the membership value 20 is higher, and a corresponding term is 10 assigned. If the membership value is higher for the term "low", then the expression A3 = 20 is replaced by A2 = LOW.

In step 230, the set of aggregated fuzzy rules generated by the rule compression process is shown to the human expert who selects only the meaningful and useful rules 15 from this set according to user desired criteria. C. Cluster Compression Process

Fig. 5 shows an exemplary cluster compression process as discussed above with respect to step 205 20 illustrated in Fig. 4. As an initial matter, it is assumed that, e.g., clusters Clust2 and Clust2 are determined in step 165. Since Clust! and Clust2 can be generated by dissimilar rules, the rules from each of these clusters Clust, and Clust2 can be very different (or similar). Therefore, it is 25 important to determine whether two different clusters are substantially similar to each other so that they can be merged. In particular, the distance between two aggregated rules of the form (3) corresponding to the centers of these clusters is determined to ascertain whether these different 30 clusters are substantially similar. As an example, the following two aggregated rules corresponding to the center of Clustj and Clust2 are considered:

Cx = a A C2 < b =» C4 = c, and 35 C2 = d Λ C3 = e => C4 = g 16 ./ WO 99/26180 PCT/US98/24339

It may be also assumed that the domains of attributes C2 and C, are discrete and ordered. These rules have different structure and therefore are different. In order to calculate the distance between these rules, we first have to bring these 5 rules into the same 4-dimensional space of attributes C3, CL, C;, and C4. This can be done by replacing these rules with the rules = a A C2 = zAc3 = x = C„=c (6) 10

Cx = d A C2 = y A C3 = e => C„ = g (7) where x and y are uniformly distributed random variables 15 ' ranging over the domains Dom(C3) and Dom(C2) of attributes C3 and C2 respectively and z is a uniformly distributed random variable ranging over the domain of Dom(C2) from its smallest element to b. This procedure can also be performed using actual distribution in the data for corresponding attributes 20 of x and y variables. It should be noted that the term C2 < b (the first aggregated rule described above) should be replaced with C2 = z in rule (6) . In addition, term C3 = x is provided into the first aggregated rule and term C2 = y is provided into the second aggregated rule. It is also assumed that, e.g., 25 random variables x, y, and z are uniformly distributed over their respective domains.

If constants are substituted for the variables x, y, and z, the terms of the aggregated rules (6) and (7) will contain only equalities and constants. Thus, these aggregated 30 rules (with the above-described substitutions) will have respective points in the same 4-dimensional space. If the distance between these two points can be calculated for fixed values of variables x, y, and z -- diClus^ (x,z), Clust2 (y) ) (i.e., if all the attributes are numeric, then the distance 35 can be a Euclidean distance; if some of the attributes are categorical and unordered, the distance can be calculated in terms of how far the nodes are in the aggregation hierarchy 17 WO 99/26180 PCT/US98/24339 defined for that attribute) -- then the distance between clusters Clustj and Clust, is equal to: d (Cl us t., Cl us t,) =-

Dom (C2) * Dorn (C3) * (b-min (Dom (C2) ) JJ d[ Clus^ (χ, ζ) , Clust2(y} ) xeDom {C3) , ye Dom ( C7) , z<b since it can be assumed that the domains of attributes C2 and 5 C3 are discrete. If these domains were continuous, integration would have been used instead.

In general, let cx = (c1L, c12, ... clk) and c2 = (c21, c22, ..., c2m) be the centers of two clusters Clust, and Clust2 as calculated in steps 170 through 185 illustrated in Fig. 4, 10 where c3 and c2 are vectors with different dimensions (because different rules can have different numbers of terms). The rules corresponding to the centers of these two clusters are extended with, e.g., dummy attributes and dummy random variables that form a union of the attributes for clusters 15 Clustj and Clust2. Assuming that the dummy variables are uniformly distributed over their domains, the distances between the two rules for fixed values of random variables can be calculated. Thereafter, the random variables are either integrated (for continuous random variables) or summed (for 20 discrete random variable) over different values of these random variables. Thus, the distance between clusters can be determined using the system and method according to the present invention.

Once the distance between the two clusters is 25 determined, the clusters can be merged as follows. In order to perform this operation, the size of the cluster should be determined as a part of the Cluster Compression process. The size of the cluster is the measure of how far the points of the cluster are apart from each other. This size can be 30 determined, e.g., using the following formula: 18 WO 99/26180 PCT/US98/24339 size (Clust) 1 £ cf(x,c) (Clustj xeClust where c is the center of the cluster. Other measurements can also be used by those having ordinary skill in the art.

The flow diagram in Fig. 5 illustrates an exemplary 5 process for compressing clusters. In particular, two clusters

Clusti and Clust2 are selected in step 250. There are a number of ways to determine which clusters should be selected in step 250. The simplest way to select the clusters is in an arbitrary manner. In step 260, the distance between the 10 clusters is determined, as discussed above. In step 265, the respective size of each cluster is determined. In step 270, a check is performed to determine if the distance between the clusters {d(Clust1, Clust2) } is smaller than the sizes of these clusters (e.g., to determine if these two clusters are "close 15 enough" to each other). If so, the clusters should be merged into one cluster in step 275; otherwise, the clusters are maintained as separate clusters. In particular, an inquiry as to whether two clusters are "close enough" can be computed in the following manner, e.g.: 2 * d(Clust,, Clust 2) -----<a 2 0 size(Clustj) -t-size (Clust 2) (θ where a is a predetermined threshold value. The two clusters should be merged by forming a new cluster consisting of points from Clust1 and Clust2 if condition (8) occurs. Steps 250-275 should be repeated until there are no more clusters left that 25 can be merged (see step 255).

In deciding which clusters Clustj and Clust2 should be chosen in step 250 of the cluster compression process, distances between, e.g., all pairs of clusters can be calculated and condition (8) can be checked to ascertain which 30 clusters should be merged. Other methods to select the clusters for compression can also be used. Furthermore, the 19 WO 99/26180 PCT/US98/24339 distance between all the pairs of clusters does not necessarily have to be calculated.

The system according to the present invention can be used in a Personal Shopping Assistant (PSA), a Personal 5 Intelligent Digital Assistant (PIDA), and in a dynamic Web content presentation system, described below. A Personal Shopping Assistant (PSA) system according to the present invention provides recommendations on the products and services that its users should consider 10 purchasing (including, e.g., suggestions for purchasing at a specific source, and at a particular price). An exemplary embodiment of the PSA system according to the present invention is shown in Fig. 6a. In particular, the system includes a User Transaction Collection and Recording Unit (or 15 · module) 115, a Past Purchasing History Storage Unit (or module) 120, a User Profile Generation module 110, a State-of-the-World module 150, a User Estimated Purchasing Needs module 140, a Purchasing Recommendations module 145, and the State-of-the-User module 160. 20 The User Transaction Collection and Recording Unit 115 collects most of the shopping transactions performed by the user (e.g. 80-90% of all the purchases made by the user). The User Transaction Collection and Recording Unit 115 can be implemented as a "smart card," or as a smart Point of Sales 25 register that records individual items purchased by the user. Alternatively, the user himself can record this information (as part of the User Transaction Collection and Recording Unit 115) using some transaction recording systems such as Quicken or Microsoft's Money. 30 When the user purchases one or more products, the User

Transaction Collection and Recording Unit 115 records and transmits this information to the Purchasing History Storage Unit 120 where this information is stored as part of the purchasing history of the user. The Purchasing History 35 Storage Unit 120 can be implemented, e.g., as a database that

records transactions performed by various users in the TRANS file, as described above. 20 WO 99/26180 PCT/US98/24339

Information stored by the Purchasing History Storage Unit 120 is provided to User Estimated Purchasing Needs module 140. In order to estimate the user's purchasing needs, pertinent static and dynamic profiles of the user should be 5 constructed based on the past purchasing histories obtained from the Purchasing History Storage Unit 120, which is performed by the User Profile Generation module 110. Static profiles include the user's demographic information (e.g., age, sex, marital status), particular preferences (e.g., user 10 prefers a particular brand of beer), and certain purchasing decisions (e.g., the user bought a particular automobile in a particular month). Dynamic profiles include a set of rules (e.g., "if a user goes to France, the user often buys perfumes in Paris", "if user Y visits a Web site from the site Z in the 15 evening, user Y does not spend a predetermined amount of time at site Z", etc .) .

In addition, the PSA system maintains information on the current State of the World using the State-of-the-World module 150, which records information, e.g., on a broad range 20 of products and services offered by various suppliers and on promotions and discounts run for these products and services. Also, the PSA system includes the State-of-the-User module 160 that maintains information about the user obtained from the Purchasing History Storage Unit 120 (e.g., the user will be in 25 New York on June 28, 1995 because she purchased an airline ticket for that date and destination) and various external information (e.g., the date, time, and the user's location, if available).

By knowing the purchasing history of a user 30 (provided from the Purchasing History Storage Unit 120), the user's profile (provided from the User Profile Generation module 110), and the external information about the user (provided from the State-of-the-User module 160), the PSA system estimates the user's future purchasing needs using the 35 User Estimated Purchasing Needs module 140. This Estimated

Purchasing Needs module 140 may match the rules specifying which products the user will buy (and when) with the user's 21 WO 99/26180 PCT/US98/24339 purchasing history. As a result, a set of products that the user should consider buying is produced.

Once future purchasing needs are estimated in Step 140, the PSA system will match these needs against a broad 5 range of products and services offered by various suppliers and on the promotions and discounts run for these products and services. This matching process is performed by the

Purchasing Recommendation module 145 using conventional methods that are known to those of ordinary skill in the art. 10 For example, if the user needs to buy a pair of jeans within the next two months, the Purchasing Recommendations module 145 selects the merchants selling jeans, e.g, the cheapest pair of jeans that fits the use's requirements (considering the promotions offered within the next two months) by matching to 15 the user profile (i.e., the user's purchasing needs). Once the Purchasing Recommendations module 145 matches the user's purchasing needs against the products and services, the Purchasing Recommendations module 145 provides purchasing recommendations to the user. 20 For example, based on the past purchasing history of a particular user, the PSA service may ascertain that whenever user X goes to France, user X often buys perfume in Paris.

This rule is stored as a part of the user profile using the User Profile Generation module 110. In addition, the 25 Purchasing History Storage Unit 120 of the PSA service may receive information that the user has purchased a ticket to Paris, and in a substantially same time period, the State-of-the-World Unit 150 of the PSA service also receives information that, e.g., Christian Dior has launched a new line 30 of perfumes that is similar to the brands previously purchased by user X. In addition, the State-of-the-World Unit 150 may · also receive information that the duty-free shop at Charles de Gaulle airport is having a sale on these new perfumes (the price being very competitive). Using the above-described 35 exemplary information, the PSA service (using the User

Estimated Purchasing Needs module 140) estimates that user X

may want to buy these perfumes and sends a message to user X 22 WO 99/26180 PCT/US98/24339

<img img-format="tif" img-content="drawing" file="IL136103AD00021.tif" id="idf0001" />

. '* (via the Purchasing Recommendation module 145) to consider purchasing the new perfume at the duty-free shop at Charles de Gaulle airport.

The success of the PSA service depends primarily on 5 accurate predictions by the PSA service of users' future needs. If the user finds, e.g., 50% of the PSA suggestions useful, the user will probably be satisfied with the PSA service. However, if the user finds, e.g., only 10% of the suggestions to be useful, the user will, most likely, reject 10 this service. As indicated above, in order to make predictions of the user's future needs more accurate, it is important to build reliable user profiles. The present invention provides a method and system for generating better dynamic profiles and, therefore, providing more accurate 15 predictions of the users' future needs.

The PSA system illustrated in Fig. 6a can be implemented using a first exemplary system shown in Fig. 6b and a second exemplary system shown in Fig. 6c. The first exemplary system of Fig. 6b provides that the User Transaction 20 Collection and Recording Unit 115 is stored on the client side. All other modules from Fig. 6a are stored on the server side and are connected to the User Transaction Collection and Recording Unit 115 via a Telecommunication Medium 130 (e.g., a telephone line or a wireless communication medium). In the 25 first exemplary system, individual user purchasing histories and static and dynamic profiles of these users are stored on the server at a central location (e.g. a database), and the method and system according to the present invention (as described above) generates improved dynamic profiles, and thus 30 provides better estimated purchasing needs of the users.

The second exemplary system of Fig. 6c provides that the User Transaction Collection and Recording Unit 115, the User's Profile Generation Module 110, the Purchasing History Storage Unit 120, the State-of-the-World module 150, the 35 State-of-the-User module 160, and the User Estimated

Purchasing Needs module 140 are stored on the client side, while the State-of-the-World module 150 and the Purchasing 23 WO 99/26180 PCT/US98/24339

Recommendations module 145 are stored on the server side. In the second exemplary system, the user dynamic profiles are validated in Step 20 of Figure 2 by the user (since these profiles are stored on the client side and are available to 5 the user for checking and validating). Once module 140 estimates user purchasing needs, these estimated user purchasing needs are transmitted via the Telecommunication Medium 130 (e.g., a telephone line or a wireless communication medium) to the server, where the estimated user purchasing 10 needs are matched by the Purchasing Recommendation module 145 to various products and services offered by various suppliers (that are stored on the server side). The resulting purchasing recommendations are transmitted back to the client side via the telecommunication medium 130 for the user's 15 consideration.

The PSA service can also be used in a Personal Intelligent Digital Assistant (PIDA) service as illustrated in Fig. 7. Each user subscribing to this additional service is provided with a Personal Digital Assistant (PDA) (e.g., the

20 remote device 350 or the User Transaction Collection and Recording Unit 115), which is connected to the PSA system (e.g., a general purpose computer 300). The PDA remote device(s) 350 (which includes, e.g., a PDA processor 360, a PDA I/O port 365 and a PDA input device 355) and the PSA 25 system(s) 300 (which includes, e.g., a display device 310, a storage device, a PSA processor 320, a PSA/ I/O port 325 and a PSA input device 305) form a client-server architecture, in which the PDA remote device is a client and the PSA system is a server. The PSA system, using the Past Purchasing History 30 Storage Unit 120 (e.g., a storage device 315) and the User

Profile Generation module 110, the State-of-the-World module 150, the State-of-the-User module 160, the User Estimated Purchasing Needs module 140 and the Purchasing Recommendations module 145 (executed by, e.g., a processor 320) estimates

35 users' future needs and behavior as described above. The PDA device accumulates additional information on the user's current state, such as the user's location information, 24 WO 99/26180 PCT/US98/24339 preferences, and desires {e.g., the user is hungry now and wants to eat). This additional information is transmitted from the PDA device to the PSA system via the telecommunication medium 130 (e.g., a wireless network, fiber- 5 optics communication system, telephone wired system, etc.) to be stored using the State-of-the-User module 160 (e.g., in the storage device 315) as part of the user's state and is used subsequently for estimating the user's purchasing needs.

For example, in order to illustrate how the PIDA 10 service operates, assume that it is Tuesday, 11:30 am and that user X is driving in his car on 1-87 in the Albany region on business, and that he indicated through his PDA device 350 that he wants to have lunch. The PDA device (350) records the current state of user X as "Tuesday, 11:30 am, presently 15 driving in user X's car on 1-87 in the Albany region, travel purpose is business, wants to have lunch." This information is sent from the PDA device 350 to the PSA system 300 via telecommunication medium 130. Based on user X's past purchasing history, the PIDA service recognizes that whenever 20 user X is traveling on business, he likes to have- light lunches at good quality restaurants and that he-generally likes sea food. By examining user X's personal profile, and by matching the dynamic rule which provides that "whenever user X travels on business, he prefers light lunches at good 25 quality restaurants", with user X's current state (user X is currently traveling on business), the PSA system 300 can predict that user X prefers a lunch at a good quality restaurant and he wants to eat light food. Next, the State-of-the-world module 150 of the PSA system 300 searches for 30 highly rated seafood restaurants in the Albany region. If the PSA system 300 finds any such restaurant, user X is provided with restaurant choices (e.g., if more than one restaurant is located) by contacting user X's PDA device 350. If the PSA system 300 does not find first choice restaurants conforming 35 to the user X's preferences, the PSA system 300 provides second choice restaurants to user X. 25 WO 99/26180 PCT/US98/24339

User needs are estimated based on purchasing history, the user's static and dynamic profiles and the current "state" of the user (sent to the PSA system from the PDA device). When the needs of the user are estimated (e.g. 5 the user wants to buy a perfume in Paris, or wants to eat at a good seafood restaurant in the Albany region), they are matched with the current state of the "world." If the PIDA service finds good matches (e.g., Christian Dior perfumes are on sale at Charles de Gaulle airport in Paris, or that there 10 is a good seafood restaurant in the Albany region serving special lunches and located very close to the user's current route), purchase recommendations are provided to the customer based on these matches. These recommendations are sent back from the PSA server 300 to the PDA device 350 via a 15 telecommunication medium 130 (e.g., via e-mail or through another intelligent user interface).

The PIDA service incorporating the system and method according to the present invention can be used for notifying the users about various purchasing opportunities, both time 20 sensitive (e.g., a particular sale will start next week) and spatial (e.g., if you need a new sweater, and sweaters you would probably like are on sale at the store near your work).

The system and method according to the present invention can also be incorporated in a Web site system. In 25 conventional systems, when· a user visits a particular Web site, the user usually sees the same contents, regardless of who the user is. Using the system and method according to the present invention (i.e., individual profiles for respective users), the dynamic Web content o: the Web site presented to 30 the user can be varied to conform -o the dynamic profile of the user visiting the Web site. Furthermore, dynamic profile construction methods can also be used in fraud detection systems. In particular, a set o£ fraud detection rules can be dynamically generated for each user. 35 It should be noted that the use of the above- described rule compression process and the cluster compression process according to the present invention is not limited to a 26 WO 99/26180 PCT/US98/24339 construction of user profiles. For example, these process can also be used for computing useful association rules in data mining applications, or in general compressing large sets of rules generated by data mining algorithms. 5 D. Selective Validation Procedure

Another embodiment of the present invention for providing a selective validation of individual user rules is shown in Fig. 8. In particular, user rules for all individual 10 users (e.g., customers) are provided to a selective validation module/arrangement (step 375). The selective validation module/arrangement can be preferably executed by a central computing device illustrated in Figs. 6a and 6b, or executed by the processor 320 of the general purpose computer 300 15 . illustrated in Fig. 7. The individual user rules may be stored in the storage device 315. It is also possible to provide the selective validation module/arrangement in the remote unit 350 illustrated in Fig. 7. In step 380, the selective validation module/arrangement receives still 20 unvalidated user rules and outputs at least one set of selectively validated individual user rules (step 390). In addition, the selective validation module/arrangement can optionally include the process illustrated in Fig. 3. In an exemplary embodiment of the present invention, this selective 25 validation procedure allows the human expert to select particular subsets of individual user rules and characterize these subsets as "Good" subsets, "Bad" subsets and/or "Undecided" subsets. A flow chart representation of an exemplary 30 embodiment of a process executed by the selective validation module (or an exemplary steps executed by the selective validation arrangement) described above is illustrated in Fig. 9. According to the present invention, a "Good_Rules" set is provided to maintain (e.g., store) all sets of individual user 35 rules which were selected by the human expert as rules which are usable for a particular user. A "Bad_Rules" set is 27 WO 99/26180 PCT/US98/24339 provided to store all sets of individual user rules selected by the human expert to be unusable for that user.

As shown in Fig.9, (in step 400) each of the "Good_Rules" and "Bad_Rules" sets are initialized, e.g., to be 5 empty or null sets. In step 405, all user rules are combined to form Set S. Set S initially contains all related (e.g., similar) subsets of the unvalidated individual user rules for all users. These similar subsets may be grouped in a similar manner as discussed above with reference to Fig. 4, or using a 10 filtering and/or clustering operator as discussed below. In step 410, the user rules in Set S (or subsets in Set S) can be displayed. The human expert examines the subsets of "related" rules from Set S (e.g., one rule or one set at a time), and selects which subsets (or which rules) in Set S are "good", 15 "bad" and/or neither (step 415). These subsets can also be examined automatically by a system (e.g., the processor 320 implementing an expert system or an artificial intelligence system) using a predetermined criteria. If a particular subset in Set S is selected to be usable, the particular 20 subset is marked as "good"; if this subset is selected to be unusable, it is marked as "bad"; if the human expert (or the system) cannot determine if the particular subset is usable or not, such subset is marked as "undecided" (step 420). In step 425, the subsets which are marked as "good" are moved from Set 25 S to the Good_Rules set, and the subsets which are marked as "bad" are moved from Set S to Bad_Rules set.

In step 430, a decision is made (e.g., automatically via the processor 320 or by the human expert) if the processing of the - selective validation module/arrangement is 30 completed, and, if so, initiates a completion process according to this embodiment of the present invention. There can be numerous conditions to indicate to the selective validation module/arrangement according to the present invention that the completion process should be initiated. 35 For example, the following exemplary conditions may prompt the selective validation module/arrangement to stop processing: 28 , "v WO 99/26180 PCT/US98/24339

Set S can became empty (i.e., all subsets of rules are moved from Set S to "Good_Rules" set and/or to "Bad_Rules" set). If this is the case, all subsets of rules are marked with their appropriate 5 designation (i.e., "good" or "bad"); the number of subsets in Set S is less than a predetermined number; the ratio of the rules in Set S with respect to all of the existing rules is less than predetermined 10 value; and the user decides to stop the process (e.g., a desired number of rules has already been classified or marked).

Other stopping criteria may be used for initiating the 15 completion process according to the present invention.

If it is determined that the processing of the completion process according to the present invention should be initiated, the rules from the Good_Rules set is assigned to one or more corresponding users (step 435), Good_Rules set 20 and/or undecided subsets can be displayed (step 440), and the execution of the process according to the present invention is stopped. If, however, it is determined that the completion process should not be initiated (i.e., the subsets should be regrouped), the remaining rules in Set S (i.e., the subsets 25 marked as "undecided") are grouped or regrouped to generate a new Set S (step 445), and this new Set S is provided to the human expert (i.e., looped back to step 410) so that the rules within new Set S may be reclassified using the process and/or the arrangement according to the present invention (i.e., 30 looped again starting with step 410).

It should be noted that if a particular subset Set S is marked as "undecided", this subset is then further analyzed by either splitting it into smaller subsets using techniques described below or optionally regrouping this particular 35 subset with other related sets from Set S as also described below. 29 WO 99/26180 PCT/US98/24339

According to an exemplary embodiment of the process according to the present invention, the rules in Set S which

were marked as "undecided" are grouped to generate a new Set S according to the following exemplary methods: 5 - A predetermined number of the remaining subsets (which can also be a single subset) contained in Set S are selected and merged together to form new subsets. The above-described remaining sets can be selected by the human expert or according to some predetermined selection 10 criterion (e.g., the size of individual sets of rules should be smaller than a predetermined value).

One or more subsets are selected from Set S. For each of these subsets, at least one of the following exemplary 15 "partitioning" operators is applied to the selected subsets: a filtering operator and/or a cluster/grouping operator (which are as described below). Other "partitioning" operators can also be implemented. The terms - "clustering operator" and "grouping operator 20 refer to identical operations and shall be utilized interchangeably below. In a particular embodiment of the present invention, subsets in Set S (obtained using the cluster operator with a particular "cut" operator) can be re-grouped based on a different "cut" operator, which may 25 depend from the previous cut and/or can be based on other parameters or criteria. For example, these subsets can be merged back into a single set of rules and the cluster operator is then applied to this subset again (but with a different "cut" parameter). Other operators can also be 30 used to regroup the subsets in Set S. I. Filtering Operators

An exemplary filtering operator receives a subset of rules and splits this subset into at least 2 subsets: one 35 subset contains rules which pass a predetermined selection criteria of the filter, and another subset contains rules which do not. In particular, this selection criteria may be 30 WO 99/26180 PCT/US98/24339 specified using a data mining query (or a pattern template). The data mining query describes a class of patterns in general terms .

Data mining queries are described in publications - 5 T. Imielinski et al., "DataMine: Application Programming

Interface and Query Language for Database Mining", Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, August 1996; J. Han et al., "DMQL: A Data Mining Query Language for relational Databases", Proceedings 10 of the SIGMOD Workshop in Research Issues on Data Mining and Knowledge Discovery, Montreal, June 1996; and W. Shen et al., "Metaqueries for Data Mining," Advances in Knowledge Discovery and Data Mining, chap. 15, AAAI Press, 1996. Any pattern description language or any data mining query language can be 15 used to specify patterns and data mining queries. For example, article by T. Imielinski et al., "DataMine: Application Programming Interface and Query Language for Database Mining," Proceedings of the Second International Conference on Knowledge Discovery and Data Mining, August 1996 20 introduced "M-SQL" for association rule discovery which is based on software query language ("SQL") modified with additional data mining operators. However, the exemplary embodiment of the data mining query does not depend on any specific language. 25 For the following exemplary request, "Find all rules in customer purchase data specifying which product categories the customers with children of various ages are buying", M-SQL query is as follows: 30 SELECT * 35

FROM

WHERE

Mine (Cus tome r Pure ha seDa ta) R R.Body&lt;{(Children^*), (ChildrenAgeLess6=*), (ChildrenAge6tol2=*), (ChildrenAgeMorel2=*)} and {(Children^*)}&lt;R.Body and R.Consequent IN {(CategorySweets=*), (CategoryCereal =★), (CategoryFruit=*)} and R.Confidence&gt;=0.5 and R.Support&gt;=0.01. 31 WO 99/26180 PCT/US98/24339

This data mining query discovers association rules if and only if they satisfy certain criteria. First, the association rules must include the fields Children, ChildrenAgeLess6, ChildrenAgre6tol2, ChildrenAgeMorel2 of the table 5 CustomerPurchaseData in the body of the rule. Second, the attribute Children must necessarily be present (this is specified by R. Body). Third, the discovered patterns must have one of the fields CategorySweets, CategoryCereal or CategoryFruit as a consequent of the rule (specified by 10 R. Consequent) . Finally, the discovered patterns must satisfy certain thresholds measuring statistical significance (i.e., R.Confidence and R. Support) .

Thus, this exemplary data mining query specifies a set of patterns. The set of these exemplary patterns may 15 indicate: the extent to which families with children younger than six years old buy sweets, the extent to which families with children older than 12 years old buy sweets, 20 - the extent to which families with children older than 12 years old buy fruit, etc..

Therefore, the pattern specified by the association rule:

Children = YES and ChildrenAgeLess6 = YES --&gt; CategorySweets = YES (0.01, 0.55) 25 noted above is also one of the patterns specified by the data mining query.

Pattern Templates are described in M. Klemmettinen et al., "Finding Interesting Rules for Large Sets of Discovered Association Rules", Proceedings of the Third 30 International Conference on Information and Knowledge

Management, December, 1994. For example, a pattern template may be provided as follows:

Children and ChildrenAge * --&gt; Category(0.01,0.5) where ChildrenAge and Category are generalizations of 35 attributes. Thus, if ChildrenAge specifies the set of attributes {ChildrenAgeLessS, ChildrenAge6tol2,

ChildrenAgeMorel2} and Category specifies the set of 32 WO 99/26180 PCT/US98/24339 attributes {CategorySweeCs, CategoryCereal, CategoryFruit}, then this pattern template specifies the same patterns as the above-described data mining query. 5 II· Clustering Operator

The clustering operator receives, as input, a subset of rules and an attribute hierarchy of this subset . In particular, the attribute hierarchy can be formed using the procedure described below with reference with Fig. 10. An 10 exemplary attribute hierarchy is illustrated in Fig. 13. All of the fields (i.e., attributes) of the attribute hierarchy are provided at the bottom of the attribute hierarchy. These fields are portions of the transaction file TRANS(Trans_ID, Cust_ID, Cx, ... Cn) as described above, without the fields 15 Trans_ID and Cust_ID. In the exemplary hierarchy illustrated in Fig. 13, n = 13. A top portion of Fig. 10 shows an exemplary procedure to generate the attribute hierarchy. In step 450, grouping data of a particular subset of rules is determined by 20 combining the fields of the TRANS file (e.g., a table) into groups (e.g., fields Cl, C2, C3 illustrated in Fig. 13 are combined into group NI, fields C4 and C5 into group N2, etc.) . In step 455, these groups are further combined into larger groups, and so on. For example and as shown in Fig. 13, 25 groups N2 and N3 are combined into group N4, groups NI and N4 are combined into group N5, fields C9 and CIO are combined into group N6, group N7 and field Cll are combined into group N8, groups N6 and N8 are combined into N9, and groups N5 and N9 are combined into N10. As a result, the attribute 30 hierarchy is generated (step 455), with attributes of the TRANS transaction file being its leaves. It should be noted that a tree which defines this attribute hierarchy (shown in Fig. 13) does not have to be balanced, i.e., all path lengths from the root node to the leaves do not have to be equal. 35 The attribute hierarchy may include one or more (e.g., two) levels of nodes below the descendent leaves of the attribute hierarchy (i.e., the fields of the TRANS transaction 33 WO 99/26180 PCT/US98/24339

file). A first level consists of a pair of attributes - field and a relational operator. The relational operator may include exemplary operators such as " = ", "&lt;", , etc. A second level is below the first level and consists of three 5 attributes - field, relational operator and sets of values which the field attribute can be compared to (e.g., predetermined values, one or more intervals, etc.). For example, the second level can be (C3, =, a) (i.e., field C3 uses the relational operator to be compared to variable 10 "a") , (C5, &lt;, 20) (i.e., field C5, via the relational operator is compared to number 20), (C8, =, [60, 80]) (i.e., field C8, via the relational operator " = " is compared to a range between 60 and 80), etc. Fig. 14 shows an exemplary illustration of the first and second level 15 extensions of node N7. In particular, the first level of field C12 is a leaf 540, which contains field C12 and a relational operator . Below leaf 540, a lowest leaf of field C12 (leaf 550) is provided with field C12, the relational operator "&lt;" and a comparison value "20". In 20 addition, the first level of field C13 is a leaf 545, which contains field C13 and a relational operator "=". Below leaf 545, a lowest leaf of field C13 (leaf 555) is provided with field C13, the relational operator "=" and a comparison range "[60, 80]". These leaves are only provided for illustrative 25 purposes, and it should be understood that other combinations of field to relational operators to comparison values/ranges are possible. These hierarchies don't necessarily have to include the same number of extensions/leaves. For example, field C12 may have two extensions, field C4 may have one 30 extension, field C5 can have no extensions and field C6 can have four extensions .

After the attribute hierarchy is generated in step 455 (shown in Fig. 10), "Cut" data is generated with respect to the attribute hierarchy (step 460) by providing a "Cut" in 35 the attribute hierarchy. "Cut" in the attribute hierarchy is defined as a set of nodes of the tree such that a union of all descendant leaves of the nodes which were identified in the 34 WO 99/26180 PCT/US98/24339 cut consists of all the fields of TRANS transaction file (i.e., C!,...,Cn). An exemplary cut is shown in Fig. 13 which includes the following groups/fields - Cl, C2, C3, N4 , N6 , Cll and N7. In addition,, the "Cut" is not limited to the nodes of 5 shown in Fig. 13, and can also include one or two levels below the field levels (shown in Fig. 14). Fig. 11 shows a detailed illustration of step 460 in which "Cut" data is generated. In step 480, the "Cut" is provided to the attribute hierarchy.

If the "Cut" is properly specified (e.g., all of the leaves of 10 the attribute hierarchy are above the "Cut", leaves being the lowest level of the attribute hierarchy) in step 485, or if the human expert (or the system) indicates that the "Cut" is unacceptable (step 490), a different "Cut" is created using similar techniques as described above for providing the 15 original cut (step 497) and the procedure is restarted at step 485 with this newly created "Cut". Otherwise, "Cut" data is generated as a function of the "Cut" (step 495) and can be stored in memory for a possible future use.

After the "Cut" data is generated (step 460 in Fig. 20 10), subsets of the user rules are grouped using "Cut" data and the hierarchy data (step 465), and these grouped subsets are placed into Set S (step 470) to be provided to the human expert.

Thus, the clustering operator consists of steps 460- 25 470. As indicated above, the following data is provided as input to the clustering operator: a) initial set of user rules, b) an attribute hierarchy as described above, and c) the "Cut". The output of the clustering operator is Set S which includes subsets of rules. These subsets are mutually 30 exclusive and collectively exhaustive (e.g., a union of the subsets is equal to all of the rules in Set S).

Fig. 12 shows an exemplary procedure for grouping subsets of the user rules using the "Cut" data as described for step 465 above (Fig. 10). In particular, all user rules 35 are combined from a number of subsets of Set S to form Set A.

In step 505, another set (i.e., a Cluster Working Set B) is initialized (e.g., to be an empty set or a null set). A new 35 WO 99/26180 PCT/US98/24339 rule is then retrieved from Set A (step 510). In step 515, if there are no more rules in Set A to be analyzed or regrouped (e.g., Set A has no more rules or is a null set), the exemplary procedure shown in Fig. 12 is completed. Otherwise, 5 in step 520, it is determined if the new rule corresponds to a class of any existing cluster subset in Set B. If that is the case, the new rule is moved into a "matched" subset in Set B (step 525) and the procedure is directed to step 510.

Otherwise, a new cluster subset is created in Set B (step 10 530), the new rule is moved to the new cluster subset in Set B (step 535), and then the procedure is directed to step 510.

Using the "Cut", two rules are provided to the same class if and only if they have the same structure with respect to the "Cut". In particular, the rules should have the same 15 number of attributes and these attributes, e.g., can be grouped in pairs so that two attributes in the same pair have the same ancestor in the "Cut". For example, the rules:

Cl = 5 and C4 &lt; 6 and C9 &gt; 8 = C12 = 8 20 and

Cl &gt; 3 and C6 = 5 and CIO &lt; 2 =&gt; C13 &lt; 7

are equivalent because fields C4 and C6 (shown in Fig. 13) have group N4 as an ancestor in the "Cut", rules C9 and CIO 25 have group N6 as an ancestor in the "Cut", and rules C12 and C13 have group N7 as an ancestor in the "Cut". It should be noted that the user rules in the same cluster are "equivalent". As such, a new rule retrieved from Set A can be compared with any rule (or a specific rule) in the related 30 cluster subset in Set B in step 520 shown in Fig. 12. In addition and as shown in Fig. 13, the rules: C2 = 4 and C9 = 5 = C12 = 8 and 3 5 C2 = 8 and Cll = 3 =· C12 = 6 36 WO 99/26180 PCT/US98/24339 are not equivalent because fields C9 and Cll do not have a common ancestor in the "Cut". Accordingly, using the procedure shown in Figs. 10 and 12, the subsets of rules of the generated clusters are provided into the set of regrouped 5 rules generated in step 445 of Fig. 9.

After Set S is split into a subset of clusters, one or more statistics may be generated for each cluster. These statistics may be, e.g., the number of rules per cluster. 10 - if a component of the rule is an attribute, the ranges of values that such attribute can assume. For example, if the attribute is "Age = a", then it may be preferable to collect statistics on the maximum and minimal values for the age in the rules for that cluster, in addition to the 15 - average value and standard deviation for that age. for different nodes/groups, how many rules correspond to different attributes for each node/group. For example, for group N6 shown in Fig. 13, it is possible to maintain the number of rules with attribute C9 and the number of 20 rules with attribute CIO. centers of clusters (calculated, e.g., with the method described above and illustrated in Figs. 4 and 5). These centers can be reported to the human expert.

These exemplary statistics may be utilized by the human expert 25 in step 410 (shown in Fig. 9) to determine which sets of rules the human expert may select for a manual examination.

This completes the description of Fig. 9 and the way rules are examined by the human expert.

If the human expert determines that the clustering 30 of rules based on a particular "Cut" is unsatisfactory, the rules may be regrouped in step 445 using a different "Cut".

For example, this different cut would be a finer cut which generated a larger number of clusters (which are smaller in size). This can be done by merging back the clusters of rules 35 obtained with the previous "Cut" (in step 445), returning to step 410 where the human expert marks all the merged rules as 37 WO 99/26180 PCT/US98/24339 "undecided", and then, in step 445 again, re-cluster rules based on the different (e.g., finer) "Cut".

The process and system according to the present invention can be implemented using, e.g., a graphical user 5 interface ("GUI") which enables the human expert to communicate with the validation system according to the present invention. Using this GUI, the human expert selects a number of operators from a graphical menu of the GUI.

Exemplary operators provided on this graphical menu may 10 include a "Filtering" operator, a "Clustering" operator and a "Browsing" operator (e.g., allows the human expert to examine sets of rules generated by the "Clustering" operator or another operator). Other operators can also be included in the graphical menu of the GUI. 15 Fig. 15 shows an exemplary flow of the process and system according to this embodiment of the present invention. In particular, the user (e.g., human expert) can select the "Filtering" operator from the graphical menu and apply this operator to Set S (step 600). As a part of the filtering 20 operator, the human expert may specify a data mining query which selects "Good", "Bad" or "undecided" rules from Set S. Then, in step 605, "Good" rules are moved from Set S· to "Good_Rules" set, and "Bad" rules are moved from Set S to "Bad_Rules" set which is, preferably, automatically saved by 25 the system (e.g., the processor) into a memory device. In step 615, the system may mark the user rules which were determined by the user (or automatically by the system) as "undecided". In step 620, the human expert may apply the remaining "undecided" rules through another filter to again 30 obtain "Good", "Bad" and "undecided" rules (which can be determined using another user-specified data mining query) from the rest of the rules. After the second "Filtering" operator is applied, the system may move "Good" rules from Set S to "Good_Rules" set, and "Bad" rules from Set S to 35 "Bad_Rules" set (step 625). In step 640, the human expert may

decide to cluster the remaining "undecided" rules in Set S using the "Clustering" operator (which the human expert 38 WO 99/26180 PCT/US98/24339 selects from the graphical menu). The "Clustering" operator generates many sets of rules that the user may decide to examine using a graphical browser by selecting a "Browsing" operator from the graphic menu (step 645). The "Browsing"

5 operator allows the user (e.g., the human expert) to examine the clusters of generated user rules by analyzing the statistics (described above) for these clusters. This process of selecting operators (from the graphical menu of available operators) can continue until, e.g., all the rules in Set S 10 have been validated or until the human expert decides to stop the processing of the validation procedure based on at least one of the above-described stopping criteria.

The human expert may apply a number of (e.g., four) operations in sequence (e.g., two filtering operators, one 15 clustering operator, and one browsing operator). This process can also be performed in parallel (e.g., the human expert may decide to perform two filtering operations in parallel and then combine their results).

In another embodiment of the present invention, 20 while the human expert proceeds deeper into an validation process (i.e., performs more iterations of steps 410-430 and 445 shown in Fig. 9), the process steps may be recorded using the GUI interface. 39