Viewing multi-dimensional data through hierarchical visualization

Description of equivalent WO 03090167 A2

VIEWING MULTI -DIMENSIONAL DATA THROUGH HIERARCHICAL VISUALIZATION

0002CROSS-REFERENCE TO RELATED APPLICATIONS This application claims the priority of U.S. application no. (Attorney Docket No. 66209) filed March 28, 2003 and entitled "VIEWING MULTI -DIMENSIONAL DATA THROUGH HIERARCHICAL VISUALIZATION", and the benefit of the following co-pending provisional applications: (a) Serial No. 60/374,064, filed April 19, 2002 and entitled "PROCESSING MIXED NUMERIC AND/OR NON-NUMERIC DATA" ;

0003(b) Serial No. 60/374,020, filed April 19, 2002 and entitled "AUTOMATIC NEURAL-NET MODEL GENERATION AND MAINTENANCE";

0004(c) Serial No. 60/374,024, filed April 19, 2002 and entitled "VIEWING MULTI -DIMENSIONAL DATA THROUGH HIERARCHICAL VISUALIZATION";

0005(d) Serial No. 60/374,041, filed April 19, 2002 and entitled "METHOD AND APPARATUS FOR DISCOVERING EVOLUTIONARY

CHANGES WITHIN A SYSTEM";

0007(e) Serial No. 60/373,977, filed April 19, 2002 and entitled "AUTOMATIC MODEL MAINTENANCE THROUGH LOCAL NETS"; and (f) Serial No. 60/373,780, filed April 19, 2002 and entitled "USING NEURAL NETWORKS FOR DATA MINING".

TECHNICAL FIELD

0009This application relates to processing of multi- dimensional data. In particular, the application relates to analyzing multi -dimensional data through hierarchical visualization .

0010DESCRIPTION OF RELATED ART As use of computers and other information and communication appliances proliferate in the current information age, data frequently is collected from numerous sources. Further, large amounts of data exist in many databases. Much of the data is collected for archiving purposes only and therefore, in many instances, are stored without organization. Sifting through the morass of multidimensional, and perhaps unorganized, data (which may include non-numerical data) to extract useful information for a specific purpose may be a substantial challenge. For example, business concerns are finding an increasing need, in order to remain competitive in their business market, to effectively analyze and extract useful information from data they and/or others have collected and use the extracted information to improve operation of the business. This, however, often may be a daunting task. Since collected data typically have many fields (in other words, many dimensions), it is difficult to make sense out of a large body of the multi -dimensional collected data.

0011Obtaining answers to a number of questions, such as the following, regarding the multi-dimensional data may facilitate the objective of understanding the data:

0012(i) Where is a data point located relative to other data points? (ii) What is the environment around a data point? (iii) For time series data, how did the system evolve over time? (iv) How can a system represented by the data be searched? (v) How can a system represented by the data be optimized?

0013Distance between data points may provide valuable information relevant to questions (such as the above) regarding the data points. The distance between two data points, in many cases, can be computed easily. Even for non-numerical data, distance or dissimilarity can also be defined and determined.

0014A data set containing a large number of data points may, however, present a problem. Although pair-wise distances can be computed for each and every pair of data points in the set, the resulting collection of distances also is a large multi-dimensional data set, and is not necessarily easier to interpret than the data points. Statistics may be used to glean some information from the data set, such as by using market basket analysis to find association rules. Statistical analysis typically only provide, however, information regarding some characteristics of the data set as a whole and leave unanswered questions concerning specific data points, such as questions (i) - (v) above.

0015Clustering can be used to separate data points in a data set into clusters from which inter-pattern relationships may be inferred. Data points inside the same cluster are generally considered to be, although they are not necessarily, close to each other. For example, such an assumption is misleading for data points away from the center of a cluster. An example is illustrated in FIG. 1. Point A, which is in Cluster 1, is actually much closer to point B in Cluster 2 than to point C in Cluster 1. Therefore, although points A and C are in the same cluster, it may be that points A and B are more similar.

0016In an analysis of a large data set, a good visualization of the data may play an important role for providing an understanding of the data. FIG. 1 shows an example in which visualization may be a good tool for understanding data. Humans often gather most of their information through visualization. Typically, the amount of information gathered by the visual sense is several times more than those gathered by the other four senses combined. Humans also possess far superior capability in identifying interesting features of an image than conventional computers. Therefore, in the analysis of a large multi-dimensional data set, good visualization of the multi-dimensional data contributes substantially towards an understanding of the data.

0017Conventional visualization techniques have been used to help the understanding of two-dimensional (2-D) and three-dimensional (3-D) data. For example, the process of analyzing such data, traditionally, is to plot the data points, in order to determine if the data can be described by a linear, parabolic or some other functional form which fits the data and can be used in a parametric model of the data. A plot-based data analysis technique often is used to handle data gathered from physics, chemistry or other experiments, in which the plot is used as a visualization of the data and plays a substantial role in the discovery of relationships between the data points. A plot-based data analysis technique may be well suited if the data are 2-D or 3-D, such as from well-defined experiments with only one or two variable parameters.

0018When data is collected from business operations or social environments, typically, many parameters are tracked simultaneously, and thus large data sets of high (more than three) dimensions are collected. Data points in a high- dimension data set cannot be plotted directly through conventional information media (for example, computer- controlled output devices, such as printers, plotters, displays, etc . ) .

0019It also has been proposed to provide a single view (referred herein as "one-shot") of the whole multi- dimensional data set, by using a dimension-reduction technique. A single reduced-dimension view may not provide, however, a sufficient visualization for^' achieving a desired result, such as separation of data points of different classes for a particular region of the multidimensional data space.

0020Another known technique for analyzing multidimensional data sets includes selecting two or three fields from the data set and plotting the data points based on the selected fields, and then repeating the selecting and plotting steps for another subset of fields. By examining the plots of corresponding subsets of the fields

0021(i.e. dimensions), some inter-pattern relationship information may be inferred. However, plots of selected dimensions only provide side views (as opposed to views from arbitrary angles) and it is not an easy task to understand the data distribution based on an examination of multiple side views. For example, most people have difficulty understanding a three-view blue print of a moderately complex machine part.

0022Therefore, relationships between data points in a multi-dimensional data set cannot be easily understood through the above-described and conventional visualization techniques . There is a great need for visualization methodologies that provide intelligible views of multi-dimensional data as a whole as well as details of the data. SUMMARY

0023The present application provides a method for hierarchical visualization of multi-dimensional data. In one embodiment, the method includes (a) applying a first dimension-reduction process to a multi-dimensional data set to obtain a first visualization, (b) selecting a subset of the multi-dimensional data set associated with a selected region of the dimension-reduced first visualization, and (c) applying a second dimension-reduction process to the selected subset of the multi-dimensional data set to obtain at least one additional visualization.

0024The additional visualization typically is at a higher level of detail than the first visualization. A sequence of more detailed visualizations may be obtained by repeating (b) and (c) for a further subset (i.e., selecting a subset of the subset) , until a sufficiently detailed, dimension-reduced visualization having a desired level of separation of points is obtained. For example, a further subset associated with a subregion of the selected region may be selected, and then (b) and (c) are repeated for the further subset . The subregion associated with the further subset may include a mixed portion. A view of the whole or part of this multi-dimensional data set may be obtained at "an angle of view" associated with the sufficiently detailed visualization which may be advantageous for a specific purpose. The first visualization may have one angle of view, and the additional visualization may have a different angle of view.

0025The second dimension-reduction process may apply the same dimension-reduction technique as used in the first dimension-reduction process. Alternatively, the first dimension-reduction process and the second dimension- reduction process may apply respective different dimension- reduction techniques.

0026The dimension-reduction process may include applying a continuous dimension-reduction technique to obtain a sequence of dimension-reduced visualizations. The method may further include selecting any two data points in the multi-dimensional data set for distance estimation. If the two points appear to be far apart in any one of the visualizations, they are far apart in the original multidimensional space. In particular, the largest one of the reduced-dimension distances is a lower bound estimate of the actual distance in the original space, if PCA (principal component analysis) is used as the dimension- reduction technique.

0027The method may further include utilizing a hierarchical cluster tree to automate the generation of hierarchical visualizations by generating a visualization for each node of the cluster tree on an as-needed basis, e.g., if the cluster still contains mixed-class patterns for a classification problem. The multi-dimensional data set, which may include non- numeric data, may be preprocessed into numerical form prior to dimension reduction.

0028The additional visualization may be consulted with multi-dimensional data from a test set (or other additional multi-dimensional data, such as recently collected data) by applying a mapping corresponding to the second dimension- reduction process.

0029The resulting visualizations are useful in many task areas such as classification, categorization, process monitoring, search and optimization, etc., both in terms of preprocessing for a more quantitative method and validation of results from another method.

0030For example, the method may be applied to classify a multi-dimensional data set according to one or more features associated with the data set. As another example, a multi-dimensional data set may be collected from a production process, and the method is applied to obtain information for predicting product properties.

0031Further, the multi-dimensional data set may correspond to data collected from a system. The method may be applied to obtain information for diagnosing a problem in the system or for predicting a problem, before the problem develops in the system. Alternatively (or additionally) , the method may be applied to obtain information for optimizing or searching the system.

BRIEF DESCRIPTION OF THE DRAWINGS

0033The features of the present application can be more readily understood from the following detailed description with reference to the accompanying drawings wherein: FIG. 1 shows a diagram of clusters in an exemplary two-dimensional data space;

0034FIG. 2A shows a flow chart of a method, according to one embodiment of the present disclosure, for hierarchical visualization of multi-dimensional data; FIG. 2B shows a flow chart of a method for hierarchical visualization of multi-dimensional data, according to another embodiment ;

0035FIG. 3A shows a diagram of a visualization of an exemplary multi-dimensional data set; FIG. 3B shows a diagram of a dimension-reduced visualization of a region of the visualization shown in FIG. 3A;

0036FIG. 3C shows a diagram of a dimension-reduced visualization of a subregion of the visualization shown in FIG. 3B;

0037FIG. 4A shows a diagram of a visualization of a second exemplary multi-dimensional data set;

0038FIG. 4B shows a diagram of a dimension-reduced visualization of a region of the visualization shown in FIG. 4A;

0039FIG. 4C shows a diagram of a dimension-reduced visualization of a subregion of the visualization shown in FIG. 4B;

0040FIG. 4D shows a diagram of a dimension-reduced visualization of a subregion of the visualization shown in

0041FIG. 4C; and

0042FIG. 4E shows a diagram of a consultation of the visualization shown in FIG. 4D, with data from a test set. DETAILED DESCRIPTION

0043The present disclosure provides tools (in the form of methodologies and systems) for analyzing a multi- dimensional data set through hierarchical visualization

0044(herein referred as "hierarchical visualization methodologies") .

0045The hierarchical visualization methodologies include a sequence of two or more dimension-reduced visualizations at respective levels of detail. Data points (e.g., a subset) in a region of a visualization can be separated out and a new visualization of the subset of data can be generated through a dimension-reduction technique. The process of separating out a portion of the data points and generating a visualization of the group of data points through dimension-reduction may be repeated for a sub-region of the new visualization (i.e., subset of the subset) until a desired level of detail (e.g., separation of points) is obtained. The sequence of visualizations are organized into a hierarchy (discussed below) and forms an instance of hierarchical visualization. Each visualization in the hierarchy has an associated angle of view and may provide increasingly better determination power for a specific task. Although a visualization in the hierarchy may be based on a subset of the data in the data set, the whole data set may be viewed through an angle of view associated with the particular dimension-reduced visualization, which in some cases may provide a better overall view (of the entire data set) for specific purposes. The hierarchical visualization methodologies may be applied to task domains in which one-shot visualization are applied conventionally. Use of hierarchical visualization may provide new information, which may lead to better determination capability than a one-shot technique. Hierarchical visualization can provide the information in a desired level of detail and its computational complexity is not much more -than a one-shot visualization since only regions of interest (usually, mixed regions, e.g., regions in which points in different categories are mixed together) are recursively processed.

0046Through hierarchical visualization, questions such as the sample questions (i) - (v) above may be answered (or at least hints to an answer may be evident) . The first three questions [(i)-(iii)] can be answered quite straightforwardly through hierarchical visualization. Searching usually involves finding close points around a selected cue, which may also be assisted by looking at the neighboring points. At least hints to an answer to the last question [ (v) ] may be obtained by using sensitivity analysis with hierarchical visualization. Sensitivity analysis involves changing the value of one field while holding the values of the other fields and monitoring changes to the data points.

0047A non-exhaustive list of dimension-reduction techniques which may be used includes the linear principal component analysis (PCA) through the Karhunen-Loeve (K-L) transform, neural-net implementations of PCA, self- organizing map (SOM) , the auto-associative mapping technique, the generative topographic mapping (GTM) , the nonlinear variance-conserving (NLVC) mapping and the equalized orthogonal mapping (EOM) , which are described in commonly-assigned United States Patents Nos . 5,734,796, 6,134,537 and 6,212,509, incorporated herein in their entirety by reference, as well as Sammon' s nonlinear mapping and its neural-net implementation, described in J. W. Sammon Jr, "A nonlinear mapping for data structure analysis," IEEE Trans . Comput . , vol. C-18, pp. 401-409, 1969, and the distance ratio constrained (DRC) mapping, described in Y. H. Pao and Z. Meng, "Visualization and the understanding of multidimensional data," Eng. Applicat . Artif . Intell . , vol. 11, pp. 659-667, 1998.

0048Hierarchical visualization may be an effective technique to obtain inter-pattern position relationship information of patterns in high dimensional data sets. When a continuous dimension-reduc ion technique is used, if two data points appear to be far apart in any one of the visualizations in the hierarchy, they are far apart in the original data space. In particular, as discussed below, if PCA is used as the dimension-reduction technique, the greatest distance (between two selected data points) in the visualizations in a hierarchy is deemed the lower bound of the true distance between the two data points in the original data space .

0049Data points in a visualization may be decorated/ annotated with properties, characteristics, features, etc. A sequence of visualizations facilitates separation of data points of different classes, properties, characteristics, features, etc., for a particular region of the multi -dimensional data space.

0050A method for hierarchical visualization of multi- dimensional data, in accordance with one embodiment, will be described with reference to FIG. 2A. A first dimension- reduction process is applied to a multi-dimensional data set to obtain a first visualization (step S21) . A region of the dimension-reduced first visualization is selected and a subset of the multi-dimensional data set associated with the selected region of the dimension-reduced first visualization is determined (step S22) . A second dimension-reduction process is applied to the subset of data to obtain an additional visualization (step S23) . If the additional dimension-reduced visualization is not sufficiently detailed (step S24) , a further subset, associated with a subregion of the additional dimension- reduced visualization, is selected (step S25) , and the dimension-reduction process is applied to the further subset to obtain a more detailed visualization (step S23) . The second dimension-reduction process may apply the same dimension-reduction technique as used in the first dimension-reduction process. Alternatively, the first dimension-reduction process and the second dimension- reduction process may apply respectively different dimension-reduction techniques.

0051Acceptable visualization of multi-dimensional data may be obtained through dimension-reduction techniques. Although the data in many instances appear to be high dimension, it is likely that many fields are not totally independent of each other. Changes in many fields may be attributed to changes to a few parameters, which may not even be present in the data set.

0052For example, when the Federal Reserve Board changes the interest rate, the interest rate change in most instances affects many business activities which results in changes to many fields of many data sets. In these exemplary circumstances, the cause for the changes to many data fields is merely a single Federal Reserve Board decision. Typically, there is no provision for annotating the collected data with the cause of data changes (for example, the Federal Reserve Board decision) . Dimension-reduction techniques are typically used to extract deciding parts from high dimension data. Reduced dimension data may then be plotted. While almost any dimension-reduction technique can be used to obtain a visualization of the data, a single generated view is not necessarily suitable for a selected purpose (for example, classification) . The deciding part of the high dimensional data in many instances cannot be fully represented through a single 2-D or 3-D visualization. Furthermore, a single view of the whole data set is not likely to produce a correct emphasis of a desired portion of the deciding part of multi-dimensional data which causes the data points to separate into desired classes. To mitigate the dilemma described in the paragraph above, hierarchical visualization methodologies are provided by the present disclosure. The hierarchical visualization methodologies, according to one embodiment, start with a. single view of the whole data set by using an appropriate dimension-reduction technique. At a high level, while some portions of the map may already bear the desired characteristics (e.g., separation of points in different categories in a classification problem) , it is likely that the resulting mapping for one or more other regions are not yet satisfactory (e.g., points in different categories are still mixed together) . For a mixed region, a subset of data points which are in the region may be singled out and a new visualization of the subset may be generated. It is different from zooming-in in the original visualization since the parameters in the new visualization are tuned to the data points in the .selected region only.

0053If more than one region in any generated visualization are to be reprocessed, sequences of visualizations may be generated from that point. The set of visualizations at different levels of detail form the hierarchical visualization. However, although a visualization in the hierarchy is based on a subset of data, the whole data set may be viewed through an angle of view associated with the visualization, which in some cases provides a better overall view for specific purposes.

0054The process of singling out a subset of interest and generating a new visualization of the subset may be repeated for a region of the new visualization until a desired level of detail with sufficient separation of points is obtained. To obtain finer inter-pattern relationship representation, a continuous dimension- reduction methodology such as the K-L transform or EOM may be used. When a continuous dimension-reduction technique is used to generate the visualizations, if two data points appear to be far apart in any one of the visualizations in the hierarchy, they are indeed far apart in the original data space, although the reverse is not generally true. Therefore, if linear PCA is used in the continuous dimension-reduction method, the largest distance between any two selected points shown in all of the visualizations in the hierarchy may be used as a lower bound of the true distance between them in the original space. The proof of the above proposition is quite straightforward. For points pi, p2 in the data set, the coordinates in the original space are [ n, x_±2 , ■■- , <img file="WO03090167A2_D0001.tif" /> and [-2i; x₂2; ••• X2π-/ respectively, and the coordinates in the transformed space are [y_11# y_{12 l} ..., yml and [y₂χ , V22, ••■, y_2n ■ The transformed space may be generated using the K-L transform which yields a visualization that provides the largest distance d between the two points and a sequence of the y coordinates corresponds to a descending order of the principal components. Since K-L transform is distance invariant, Equation (1) as follows is true (π is the number of dimensions in the original space) : <img file="WO03090167A2_D0002.tif" />

0055A distance d' between the two points on the reduced- dimension visualization is expressed in Equation (2) as follows (m is the number of dimensions in the reduced space) :

0056' ^ ^(y₂ι - yn ² + (y 22 - ynf + - - + (y2_m - yι,n)² (2 )

0057Since m &lt; n, it follows that d'&lt;d, i.e. d' is a lower bound of d .

0058While the mixed regions may be manually selected according to one embodiment, hierarchical visualizations may be generated automatically, according to another embodiment, by utilizing a hierarchical cluster tree and generating a visualization for each node of the cluster tree on an as-needed basis (for example, if the cluster still contains mixed-class patterns for a classification problem) . An advantage is that unattended processing can be carried out. However, because the configuration of the hierarchical clustering may be less than optimal, a greater-than-necessary number of visualizations for regions of less interest may be generated in some instances, while less-than-enough visualizations are generated in some other instances. Cluster tree and other clustering (and decision tree) techniques are discussed in commonly-assigned U.S. Provisional Applications No. 60/374,020, entitled "AUTOMATIC NEURAL-NET MODEL GENERATION AND MAINTENANCE", NO. 60/374,041, entitled "METHOD AND APPARATUS FOR DISCOVERING EVOLUTIONARY CHANGES WITHIN A SYSTEM", and No.

005960/373,977, entitled "AUTOMATIC MODEL MAINTENANCE THROUGH LOCAL NETS".

0060In many instances the multi-dimensional data set is the data of interest and is the ultimate subject of the visualization task. However, in other instances, an initial data set (for example, historical data which were collected from a system or process which is now monitored) may be used off-line for identifying, through the hierarchical visualization methodologies, a suitable reduced-dimension visualization with associated angle of view. The visualization with associated angle of view may then be consulted on-line with additional data which is collected from the process or system of interest (see, for example, FIG. 2B) . The consultation^' may be used, for example, to diagnose or predict a problem in the process or system, or to optimize the process or system.

0061Compared with a one-shot visualization technique, hierarchical visualization has the clear advantage of providing more detailed information on demand. Furthermore, hierarchical visualization is not much more computationally intensive than the underlying dimension- reduction method which may also be used in one-shot visualization, since only the mixed regions are to be recursively processed.

0062As with a one-shot visualization technique, hierarchical visualization may be applied to many tasks in areas such as classification, categorization, process monitoring and optimization, etc. Hierarchical visualization is especially suited for qualitative estimation and determination. In particular, the advantage of easy estimation of the distribution of solutions from a visualization makes it very useful in selecting ranges of operation for subsequent more quantitative processing methods . When a proposed solution point is plotted on the visualization map, its stability may also be quickly estimated from the characteristics of its neighboring points. Additional applications of visualization methodologies are described and/or suggested in commonly- assigned United States Patents Nos. 5,734,796, 6,134,537 and 6,212,509, which are incorporated herein in their entirety by reference.

0063Some examples are discussed below. The examples are provide to aid in an understanding of the subject matter of this disclosure but are not intended to, and should not be construed to, limit in any way the claims which follow thereafter.

0064Compound Formation Example The first example concerns a problem of predicting compound formation with three ingredient elements. Five attributes of each element form the 15-featured data set. The five attributes are electro-negativity, number of valence elections, Mendeleev number, melting temperature and Zunger radius .

0065FIG. 3A shows a specific 2-D visualization of the data set. In the map of FIG. 3A, compound forming patterns (referred herein as "formers") are denoted with open circles. The non-formers are denoted with solid black circles. Some areas are dominated by formers or by non- formers. If an unknown pattern falls into one of these areas, it is easy to determine whether the new pattern is a former or a non-former. However, there are also mixed areas. A gray circle, indicating a new pattern, is located in a mixed area. Whether it is a former or a non-former cannot be determined yet. Therefore, the single visualization, as illustrated in FIG. 3A, is not enough to solve the classification problem. In order to make a determination regarding the gray circle pattern, hierarchical visualization is applied.

0066The patterns that are inside the rectangular box around the new pattern shown in FIG. 3A are selected and a new visualization is generated, which is shown in FIG. 3B.

0067Although FIG. 3B shows some further separation of the patterns into the two classes, the new pattern is still in a mixed area, which is delimited by a rectangular box.

0068A recursive application of hierarchical visualization on data points that are inside the rectangular box shown in FIG. 3B generates a new visualization illustrated in FIG. 3C. From FIG. 3C, it is clear that the new pattern falls inside the area of non-formers and it is more likely to be a non-former. The actual class of the pattern is indeed non-compound forming.

0069News Example

0070The second example involves classification of selected stories from a library of news stories. The data set is generated by randomly selecting stories out of three topical categories from the news stories library. The three topic categories are "crude," "earn" and "ship." These three topics were selected since (a) they are among the few topics with over a hundred stories in the library and (b) they seem to be relatively distant in concept and hence one might expect them to be separable when visualized. The stories were separated into a training set and a test set . Since the original data are not in numerical form, they may be preprocessed using signpost transformation methodologies, discussed in commonly- assigned U.S. Provisional Patent Application No. 60/374,064 entitled "PROCESSING MIXED NUMERIC AND/OR NON-NUMERIC DATA", which is incorporated herein by reference. The visualization obtained using the stories in the training set is illustrated in FIG. 4A. The black circles represent stories of topic "earn", and the gray and open circles represent stories of topics "crude" and "ship", respectively. FIG. 4A shows two distinct groups, with stories of all three topics in each group. Within each group, some separation of stories based on topics may be observed. The small group represents all stories with only a title and no body words. The high-level visualization of FIG. 4A apparently emphasized a difference between whether a story contains words in its body or not. However, the goal is to distinguish stories of different topics and therefore hierarchical visualization is applied.

0071FIGS. 4B through 4D show three successive levels of hierarchical visualization, each generated using patterns in a selected region of the previous visualization. Better separation of stories of different topics may be achieved as a more localized region was selected. FIG. 4B shows a new visualization based on patterns that are in a rectangular box shown in FIG. 4A. Stories of topics

0072"crude" and "ship" and some of "earn" are still mixed in the visualization shown in FIG. 4B .

0073FIG. 4C shows a new visualization, at a higher level of detail, of patterns that are in a rectangular box shown in FIG. 4B . Separation of stories based on topics is more evident. However, stories of topic "ship" and some of topics "crude" and "earn" are still shown to be close to each other.

0074FIG. 4D shows a new visualization, at an even higher level of detail, of patterns inside a rectangular box shown in FIG. 4C. Separation of stories based on topics is well established with only a few exceptions.

0075The visualization shown in FIG. 4D was consulted with data from the test set. FIG. 4E is a visualization of the patterns in the test data set using the same mapping as in FIG. 4D. Separation of stories of different topics similar to that in FIG. 4D is evident. It may be seen that although the mapping in FIG. 4D is based on only a portion of the data set, the particular angle of view associated with the visualization is actually better for the purpose of predicting the topic of a story then a view based on all the patterns. The same type of separation of stories of different topics as shown in FIG. 4D may again be observed with new stories.

0076The examples discussed above demonstrate that the advantages of easy estimation of distribution of solutions and stability of a solution which are provided through hierarchical visualization may be expanded into previously undeterminable problems .

0077For example, much of the morass of data in the current information age is mixed data which includes non-numeric data. The analysis of mixed data may be very difficult, particularly when the data are from different sources. Very different non-numeric data patterns may have similar meaning and/or effect. As discussed above, hierarchical visualization may be used, after mixed data are preprocessed through signpost transformation into numerical form, to obtain a more intelligible, reduced-dimension visualization of the data.

0078As another example, process (or system) monitoring often is a difficult task which involves examination of multi-dimensional data (such as collected from a process or system in an enterprise environment) . As demonstrated by the compound formation example, the hierarchical visualization methodologies may be applied to make predictions regarding properties of products produced through a process or system. Furthermore, in many instances, it would be very useful for making adjustments to, for example, setpoints of the process or system, when (or ideally before) a problem arises in the process or system, by diagnosing or predicting such problem based on examination of multi-dimensional process data samples. Hierarchical visualization may be used (e.g., offline) with historical process data to identify a dimension-reduced visualization, with associated angle of view, of the process data which permits ready diagnosis or prediction of problems. The visualization with the associated angle of view may then be deployed in an on-line system for consultation with real time (or pseudo real time) process data . Hierarchical visualization also may be useful in other phases of a business cycle in which multi-dimensional data are analyzed, including design, engineering, development, marketing, sales, training, support and customer support. In addition, analysis of multi-dimensional data in a variety of industries, including retail, technology, automotive, financial services, healthcare and medical, eBusiness, etc., may be aided through use of the hierarchical visualization methodologies.

0079The hierarchical visualization methodologies of this disclosure may be incorporated in, for example, software- implemented systems, and may be embodied in one or more computer programs or software modules stored on a conventional program storage device or computer readable medium, and/or transmitted via a computer network or other transmission medium. The hierarchical visualization methodologies may be integrated with graphical modeling and/or presentation programs which provide perspective selectivity (for example, rotationally) through user interaction. Thus, the user can view the reduced-dimension (e.g., 3-D) visualization of the multi-dimensional data set from a user selected perspective.

0080The hierarchical visualization methodologies may be integrated with predictive analysis systems and software. For example, a predictive analysis tool may be applied to multi-dimensional data collected from the hotel industry. A hotel business may use the predictive analysis tool to obtain predictions regarding occupancy rates as a function of assorted factors (such as price, amenities, time of year, etc.) . It is generally desirable to visualize the predicted occupancy rates in the context of the data. However, since the data includes many factors which may collectively affect the number of hotel customers, it is not possible (or at least practical) to present the multidimensional data in a single multi-dimensional view. It should be evident in view of the discussion above that the hierarchical visualization methodologies may be applied to identify a reduced-dimension visualization which would be suitable for studying the predicted occupancy rate. The reduced-dimension visualization of the hotel industry data may facilitate, for example, pseudo real-time selection of suitable prices by the hotel business which allow the hotel to achieve full occupancy while optimizing profitability. The above specific embodiments are illustrative, and many variations can be introduced on these embodiments without departing from the spirit of the disclosure or from the scope of the appended claims. Elements and/or features of different illustrative embodiments may be combined with each other and/or substituted for each other within the scope of this disclosure and appended claims.

0081Additional variations may be apparent to one of ordinary skill in the art from reading the following U.S. applications, which are incorporated herein in their entireties by reference:

0082(a) Serial No. 60/374,064, filed April 19, 2002 and entitled "PROCESSING MIXED NUMERIC AND/OR NON-NUMERIC DATA" ;

0083(b) Serial No. 60/374,020, filed April 19, 2002 and entitled "AUTOMATIC NEURAL-NET MODEL GENERATION AND MAINTENANCE";

0084(c) Serial No. 60/374,024, filed April 19, 2002 and entitled "VIEWING MULTI -DIMENSIONAL DATA THROUGH

HIERARCHICAL VISUALIZATION";

0086(d) Serial No. 60/374,041, filed April 19, 2002 and entitled "METHOD AND APPARATUS FOR DISCOVERING EVOLUTIONARY CHANGES WITHIN A SYSTEM"; (e) Serial No. 60/373,977, filed April 19, 2002 and entitled "AUTOMATIC MODEL MAINTENANCE THROUGH LOCAL NETS";

0087(f) Serial No. 60/373,780, filed April 19, 2002 and entitled "USING NEURAL NETWORKS FOR DATA MINING"; and

0088(g) Serial No. (Attorney Docket No. 66209) filed March 28, 2003 and entitled "VIEWING MULTI -DIMENSIONAL DATA

0089THROUGH HIERARCHICAL VISUALIZATION".