Over the last few years, many organizations

Transcription

1 Evaluating and Selecting Mining Tools by Mike Ferguson Over the last few years, any organizations have invested oney developing data warehouses and data arts. They have ipleented systeatic schees for extracting data fro a variety of operational sources, purchased new tools to help with data cleaning and collating, designed new database structures to suarize and store the data, and purchased new hardware for the database s. The justification for these efforts has varied widely, particularly in the degree of focus on specific business questions to be addressed. While subject-oriented data arts are usually designed to eet a well-defined business need, enterprise-wide data warehousing efforts have often had less specific goals. Soe took the If you build it, they will coe approach. Generally, regardless of the original otivation for investing in warehousing, that prophecy will coe true. In this era of coputer inforation systes, ost organizations are drowning in data while starving for real inforation. Build a warehouse or a data art offering data in a for that perits easy analysis, and there is likely to be an upsurge of interest in data analysis. Business analysts want to find new ways to increase profitability, budget watchers want to axiize the return on their investent in infrastructure. Such is the cliate that has led to the widespread current interest in data ining, also known as database exploration, inforation discovery, or knowledge discovery. This article is an overview of data ining, addressing: A fraework for understanding the eleents of data ining The atheatical and statistical techniques that data ining tools are based on Soe applications of these data ining techniques A set of evaluation criteria for selecting data ining products Planning considerations when introducing today s (still iature) data ining technology Knowing What to Look for No atter how uch integrated and suarized data it contains, every data warehouse ust include a set of reporting tools that can turn the data into useful inforation. analysis products offer functions ranging fro basic query and reporting capabilities to sophisticated developent tools for creating online analytical processing, or OLAP, applications. In practice, different tools are aied at different warehouse users. The ost powerful tools are typically used by inforation object producers, while visual reporting and data delivery tools are designed for inforation consuers. Despite their any differences, however, ost of these analysis tools share one key characteristic: they depend on their users to guide the investigation. They require a predefined starting point a hypothesis, a query, a procedure, or a progra that dictates the nature of the data analysis to be carried out by the tool. Naturally, a tool (assuing it is easy to use) does provide its user the facility to ore easily explore the contents of a database and test out various hypotheses. However, the tools theselves do not address the need to uncover previously unknown business facts. Yet this is a very real requireent. Thanks to our coputers, we have ore data than we can handle. As a result, any areas of decision aking territory that could have potentially enorous value to a business still reain totally unexplored because no one knows of their existence. Today, data ining technology is taking the first step into this unknown territory. In general, the goal of data ining is to bring the practice of inforation processing closer to providing the real answers organizations are seeking fro their data. Coercial users hope that data ining will suppleent huan anageent insights, allowing the business to ake ore proactive, knowledgedriven decisions in their quest to reain copetitive. A Mining Fraework Figure 1 illustrates a siple fraework that will help explain what is involved in data ining. Based in part on an analysis by Two Crows Corporation, 1 the fraework depicts the following five ajor eleents: the business proble, data ining approaches, specific techniques, algoriths or ipleentations, and tools or products. The business proble. This, of course, ust be the starting point in any consideration of data ining technology. For exaple, you would like to answer questions like, Which custoers are ost likely to default, and why? ining tools seek to address two key business requireents:

2 Description discovering patterns, associations and clusters of inforation (such as custoer buying patterns). Prediction using those patterns to predict future trends and behaviors (such as the likelihood of a custoer defaulting on a loan). ining approaches. These are the classes or categories of data ining ethods used for description and/or prediction. There are six ajor classes: Classification (developing profiles of groups of ites in ters of their attributes) Regression (establishing a relationship between a series of ites, for the purpose of forecasting) Ipleentation algorith Figure 1. A data ining fraework 1 Business proble ining approach ining technique 1 1 Mining product Tie series (like regression, but using the additional properties of tied inforation) Clustering (segenting ites that exhibit consistent behavior or characteristics into subsets or clusters) Association analysis (recognizing that the presence of one set of ites iplies the presence of another set) Sequence discovery (recognizing that of one set of ites is followed by another set) Classification, regression, and tie series are priarily used for prediction, while clustering, association, and sequence discovery are ore appropriate for describing relationships that exist in the data. ining techniques. These are the basis of all data ining. Over the last thirty years, coputers have inspired researchers to suppleent the forer anual ethods of statistical analysis with any new coputational techniques for discovering patterns in data. (See the sidebar for a list of popular data ining techniques used in today s products, together with soe exaples of typical business applications to which they ight be applied.) Algoriths. Although each technique represents a specific type of analysis, products ipleent any varieties of the techniques. The design of the data ining algorith(s) can ake a significant difference to product perforance and scalability. This is crucial when the product is required to ine a very large database; never assue linear scalability of any data ining tool as data volues grow. ining products range fro stand-alone desktop-based tools to data ining s ipleented on high-perforance client/ or parallel processing coputers, exploiting the power of ultiprocessors. Soe products extract saples of inforation and ine the saples locally on the desktop while others ine the data directly in the warehouse databases. Fraework Relationships Turning to the relationships aong the eleents of the fraework, Figure 1 includes soe data odeling notations indicating the cardinality ( 1 or ) of each relationship. For those less failiar with data odeling notations, here is what Figure 1 says: A business proble ay be studied using ore than one class of odel, and (of course) any odeling approach can be applied to any different business probles. More than one technique can be used for any class of odel, and any given technique can be used for ore than one class of odel. There is usually ore than one way of ipleenting any given technique. ining tools ay support ore than one of the techniques, and (of course) each technique is supported by ore than one vendor s product.

3 For any given technique (taking slight liberties with the odeling notation), a particular product supports a particular ipleentation algorith. Evaluation Criteria for Mining Products To help in selecting data ining tools, we now propose a set of evaluation criteria. These are classified into nine ajor areas, as follows: 1. Product Architecture 2. Warehouse/OLAP Integration 3. Perforance 4. Function 5. Presentation 6. Sources 7. Preparation ining client tool or Web browser Extract/ saple LAN WAN or WWW SQL or DML Tier 1 Tier 2 Tier 3 Figure 2. Possible data ining configurations Mining Mining or file Single Hardware Server SQL SQL or DML or file 8. Environent 9. Adinistration Each area is described briefly below, and represented by a set of evaluation criteria in Table 1, A Product Evaluation Checklist. Following this review of the nine areas, we discuss two iportant product requireents rule conversion and data visualization in a bit ore depth. 1. Product Architecture When evaluating any software product, we should understand its operational and connectivity requireents, how its data (and etadata) is stored, and whether run-tie perforance and scalability requireents have been taken into account in its design. Possible two-tier and three-tier data ining architectures are depicted in Figure 2. Sapling vs. Direct Access. Tools ay use sapling techniques to process a representative subset of data or process the data by accessing a warehouse directly. Sapling supports client-based analysis, while Figure 3. ining parallelis direct access is a -based approach. The ore data to be ined, the ore we would tend to favor tools that perfor their analysis in place. Direct access tools that read the data using the native SQL of the database are likely to be the ost scalable as data volues rise. On the other hand, a sapling approach will usually offer ore opportunity for data cleaning and other data preparation activities (see Preparation below). But if the data is already stored in a well-designed data warehouse, these activities ay ulti threaded ining objects - 1 per processor Mining Server A P I Parallel DBMS or file A P I have already been carried out. We return to this subject in our discussion of planning considerations, in the last part of this article. 2. Warehouse/OLAP Integration This criterion deals with how well the products integrate with other coponents of a data warehousing or data analysis architecture. Several fors of integration are possible:

4 Mining Techniques The ost coonly used techniques in data ining, together with exaples of potential applications are: Decision Trees A decision tree is a tree-shaped structure that visually describes a set of rules (conditions) that caused a decision to be ade, such as a decision to purchase a product. Fro decision trees we can generate rules for the autoatic classification of a set of data, for exaple, to segent a custoer database. Specific decision tree ethods include Classification and Regression Trees (CART) and Chi Square Autoatic Interaction Detection (CHAID). Potential applications for decision trees include: Medical diagnosis: What are the factors that affect kidney transplant survival rates? Retail analysis: What are the categories of results fro a custoer survey? Decision trees are often cobined with neural networks to explain why a neural network has reached a particular conclusion. They can also be cobined with rule induction (see below) to generate logic for repetitive decision aking. For exaple, to help telesales staff confine their sales of ortgages to those custoers and prospects who are identified as good risks because their profile atches one autoatically ined fro the data. Genetic Algoriths Genetic algoriths are optiization techniques that can be used to iprove other data ining algoriths so that they derive the best odel for a given set of data. The resulting odel is then applied to the data to uncover hidden patterns or to ake predictions. Genetic algoriths are best suited to segentation/clustering applications but can be applied to ost situations involving learning. Note the use of the word optial in the following questions, which is an indicator that genetic algoriths ay be the technique to use to answer these types of questions. Potential applications include: Direct ailing: What is the optial profile of the ideal custoer who is likely to invest ore than 25,000 in utual funds this year? Risk analysis: What is the optial profile of a low risk hoe loan prospect who earns less than $75,000 and wants to live in San Jose? Medical diagnosis: What is the optial treatent plan for a particular diagnosis? Retail analysis: What is the optial store layout for a particular location? Neural Networks These are non-linear predictive odels that learn how to detect a pattern to atch a particular profile through a training process that involves iterative learning, using a set of data that describes what you want the to find. They are so naed because, on a superficial level, their operation resebles that of the huan brain. Potential applications include: Direct ailing: Who will respond to this ailing? Risk analysis: Which prospective custoers are a good credit risk? Medical diagnosis: What disease is this person likely to contract? Retail analysis: What product is this custoer likely to purchase? Neural networks are appropriate for clustering, sequencing, and prediction probles. Their ain drawback is their opaqueness they do not explain why they have reached a particular conclusion. But in soe applications (like atching handwriting saples) this is not iportant. Predictive Modeling A variety of techniques can be used to identify patterns, which can then be used to predict the future. These include classical linear regression and its derivative, logistic regression analysis, and further extensions such as Generalized Additive Models (GAM) and Multivariate Adaptive Regression Splines (MARS). These techniques are ost often used to predict the odds of a particular outcoe, based upon the observed data. For exaple: Direct ailing: Which geographic areas are ost likely to respond to a new ailing initiative? Retail analysis: Which custoers are ost likely to be interested in a new product? Rule Induction Rule induction is the process of extracting useful If...then... rules fro data, based on statistical significance. Potential applications include:

5 Direct ailing: Will the response to a ailing capaign be greater than 5%? Risk analysis: Will a one year return on investent of a stock be greater, equal to, or less than 10%? Risk Manageent: What are the good risk criteria and thresholds that ust be et before offering a custoer a ortgage? Fuzzy Logic Fuzzy logic handles iprecise concepts (like sall, big, young, old, high, low) and is ore flexible than other techniques. It provides the notion of a fuzzy set rather than clear dearcation boundaries; for exaple, rather than 0 or 1 there are also 0.9, 0.85, 0.93, 0.21, 0.05 etc. A potential application ight be direct ailing: Who ight be a likely person to ail in our new capaign? K-Nearest Neighbor (k-nn) This technique places an object of interest into a class or group by exaining its attributes and grouping it with others whose attributes are closest to it. k-nn is a classic technique for discovering associations and sequences when the data attributes are nueric. With non-nueric attributes or variables, it is a lot harder to apply this technique, because of the difficulty of defining a etric that can be used to quantify the distance between a pair of non-nueric values. Published APIs open up the possibility of interoperation with any other tools. Next, consider etadata integration. Where and how is a product s etadata stored, and whether it can be launched fro an inforation directory, for exaple. Rule conversion is another useful feature. If a ining tool can convert the rules it has discovered about the business into SQL or 3GL code, this code can then be reused by other decision support tools or applications. See the later section on this topic. Ideally, an interface with an OLAP tool would perit a two-way exchange of inforation, supporting an iterative style of data analysis. The OLAP tool could pass selected subsets of a data warehouse to the ining tool to be ined for patterns that are not easily exposed using analytical tools. Then, if the data ining tool can pass its findings as SQL queries directly to the OLAP tool, the OLAP tool can be used to verify the findings in a larger database. 3. Perforance ining is inherently coputationally expensive and can involve large aounts of data. To support interactive data ining activity, a tool ust exploit perforance enhancing technology. Does the tool exploit ultiprocessors, for exaple, by running ultiple instances of itself, or by handling the processing for a ining technique like a neural network in parallel (see Figure 3). 4. Function Next, consider the extent of the data ining function supported by a product to what degree does it support the full range of potential data ining activities? Users generally need to have a range of techniques available to use in a data ining solution how any different data ining approaches and techniques are supported in the product? Does it have a set of canned rule odels and applications that can aid productivity? Application developent facilities and agent technology that enables event- or tierdriven ining are also worth looking for. Clearly the ore techniques supported by a product the ore types of ining applications can be supported. Furtherore, support for ultiple techniques allows users to look at the sae proble fro any different angles, and so iprove the thoroughness of an investigation. If a pattern is not detected using one for of ining, another technique can be used to see if it uncovers anything. Equally, if different techniques autoatically coe to the sae conclusion over pattern discoveries, then user confidence grows in the accuracy of the find. 5. Presentation This evaluation criterion deals with the user interface. Especially iportant is the support for data visualization, by providing different ways to view the ined data. We discuss this area further in a later section see Visualization is Key. Also, a tool s usability will be influenced by its interfaces with workflow or groupware products, which can allow us to tie together the separate stages of a data ining process and disseinate the findings. 6. Sources Earlier (under Architecture) we addressed the issue of whether the tool supports direct access to data or extracts a saple for ining. Here we

6 evaluate the level of support for various possible data sources. Of course, for any potential user of a data ining product, the specific data sources of interest will depend on the existing database environent. For soe applications, it ay also be iportant to be able to include external data sources in the ining process. 7. Preparation A variety of data preparation activities are norally needed before applying the actual data ining techniques, especially when the source data coes fro operational databases. A data ining tool ay provide support for: cleansing, such as copleting issing data fields, identifying inconsistent data in different data sources, or resolving referential integrity violations. description, supplying etadata such as row or value counts or distributions. transforation, adding new derived values, cobining continuous attribute values into ranges, or replacing categorical attributes by a series of binary (yes/no) attributes. The need for such transforations will depend upon the tool and the technique(s) to be applied. sapling, as required for training or odel building. pruning, identifying dependent, independent, and correlated coluns or variables. Typically, data preparation is the ost tie-consuing aspect of data ining. Also, the preparation effort increases with the nuber of data sources to be consolidated. Anything that a tool can do to support this process will speed up the overall process of developing a data ining odel. Product evaluation criteria 1. Product Architecture Client/ architecture storage anipulation Connectivity to data 2. Warehouse/OLAP Integration Interoperability Metadata Rule conversion Connectivity to other tools 3. Perforance Scalability Parallelis 4. Function Mining approaches Mining techniques Canned applications Canned rule odels Application developent 5. Presentation visualization Workflow interface Groupware interface GUI 6. Sources Warehouse data Operational data External data Multi-database 7. Preparation cleansing description transforation sapling pruning 8. Environent Platfors Size constraints 9. Adinistration Central adinistration Security Table 1. A product evaluation checklist Features to consider 2-tier or 3-tier? Flat file or relational database? Extract sapling, direct access to database, or both? ODBC, DBMS gateway, or ORB? Server-resident ining objects supported? Published APIs exist? Docuent ining objects in an inforation directory? Launch ining objects fro a directory? Convert ined rules to SQL or 3GL code? Pass rules directly to OLAP tools? Receive data for ining fro OLAP tools? Support for ultiple user access (ining )? Multi-threaded ining? Support for ining very large databases? Architecture exploits ultiprocessors, parallel s? Algoriths designed for parallel operation? Support for ajor prediction and description approaches? How any data ining techniques are supported? For exaple, pre-built Risk Manageent applications For exaple, Credit Model, Debit Model, Fraud Detection Model, Credit Risk Model, Profitability Model, Attrition Model Application developent capability? Support for tier-based or event-based ining triggers? Support for graphs, aps, tables, rotation, etc? Support for a ining process (saple, report, etc.)? Support for disseinating output (e-ail, publish/subscribe)? Toolbar custoization etc.? Direct access to data warehouse? Access to data in operational databases? Access to external data (deographics, etc.)? Multiple data sources supported directly, or via a gateway? Identify issing, inconsistent or incorrect data? Generates etadata fro source data? Perfors data conversions, transforations? Assists with sapling process? Assists with pruning and selection of independent variables? Client platfors supported? Server platfors supported? Maxiu saple size? Maxiu nuber of rows or records? Maxiu nuber of fields? Maxiu nuber of values for a discrete field? Maxiu range for a nueric field? Support for versioning of data ining objects? Usage onitoring and reporting? User access authorization? Private, group, and public ining objects?

7 8. Environent This category deals with the client and platfors supported, and any liitations on the sizes of data that can be ined. 9. Adinistration There is a significant difference between a tool designed for a single user and one intended for enterprisewide use, or even departental sharing. Shared tools will include central support for user security and authorization, usage onitoring and reporting, and for saving and reusing ultiple versions of data ining objects such as data saples, odels, and rules. In addition to the features discussed above, one ust also (as with all products) consider the vendor. ining is a coplex and relatively new field; any sall copanies are offering new products. Does the vendor have the resources to aintain the product s evolution and growth? Is training available? Rule Conversion Support for rule conversion is an iportant requireent for the integration of data ining tools with OLAP and other data analysis applications. The ability to convert ined rules into SQL queries and 3GL code supports two requireents of inforation processing: segentation. Autoatically converting the rules that define ined patterns into SQL or 3GL allows the to be used iediately for exaple, to segent a custoer base for custoized arketing. Rule reuse. Mined rules, when expressed in the for of SQL queries, becoe available for reuse by other OLAP applications. Rules converted to 3GL can be stored (for exaple, on a rules ) as standard procedures or objects, and reused by application developent tools. This is an area in which OLAP and data ining products often work well together, the ining tool generating the SQL fro the rules discovered and passing the SQL to an OLAP product for execution. Alternatively, converting ined rules into 3GL code allows these sae rules to be included in applications. This capability can then perit soe systes to be brought forward to the front office. For exaple, suppose that the front-end of a ortgage issuing application norally siply a data gathering stage could be extended to check the custoer input against previously ined rules that qualify custoers as a good risk. Telephone operators and counter clerks could then be epowered to authorize certain safe loans without referring the application fors to back office staff. This exaple illustrates how data ining can lead to a variety of benefits, including: Increased expertise ebodied in application software Correct reuse of previously established business rules More propt custoer service Risk anageent, rather than risk avoidance. Visualization Is Key No atter which type of data ining application is planned, or developent tools used, a rich set of data visualization capabilities are an iportant prerequisite. Mining is a cobination of two concepts: Autoatic pattern discovery Pattern visualization. No atter how effective the data ining tool ay be discovering patterns, those patterns are viewed, evaluated, and acted upon through the tool s user interface. The tool user has the power to abandon a line of inquiry, to regenerate the inquiry with adjusted paraeters, or to ake a strategic anageent decision based on the inforation presented. Clearly, the influence of the visualization coponent of the data ining tool is paraount. The success of the visualization activity relies heavily on the huan user s ability to access and coprehend the results of the analysis. Therefore: There ust be a natural way for users to visualize the results of the data ining activity. Visualization ust be intuitive and integrate sealessly into the user s current environent. Even the best set of rules or the ost coplete table of data ay reveal ore inforation when visualized with color, shape, position, size, orientation, relief, or texture. To illustrate this, an exaple of segentation is shown in Figure 4. The figure shows deographic data reporting the levels of response to a arketing prootion, for states and local areas. The height of each state indicates its response rating and ebedded coluns identify specific localities. [Editor s note: Unfortunately, black and white reproduction reduces the effectiveness of the original colored iage.] Planning the Introduction of Mining Our nine-part checklist proposed a set of evaluation criteria for data ining products. In an ideal world one data ining tool would satisfy all these requireents. However life is not like that at least, not

8 yet. Copared with expectations, early releases of any of the new ining tools are liited in functionality and not well docuented. In this part of our overview, we review soe of the practical considerations surrounding the introduction and use of today s data ining tools. An Ipleentation Fraework ining, like any other technology, should be introduced in an architected anner to ensure true integration. A defined technical architecture is a blueprint for a true endto-end inforation delivery solution that includes data ining technology. Figure 5 shows the Base Associates International Warehouse Technical Architecture. The Access coponent of the above technical architecture coprises several related technologies that address data access and analysis: Query, reporting and analysis tools Multi-diensional DBMS OLAP clients Relational OLAP clients DSS application developent tools ining technology is the ost recent addition to this coponent, although the overall level of function, and the degree of integration aong the current tools in this arketplace is still soewhat liited, as we discuss below. Early Product Liitations It is, of course, a fact of business that the aturity and scope of any software product in both its current and its future releases depends on arket forces, and on the scale of the vendor s operations and the range of environents the vendor can support. A Brief Look at the Marketplace Many of the products in the arketplace are based on a particular ain technique. Soe support ultiple techniques. Below is a list of techniques and soe of the products in the arketplace that use the. Decision Tree Soe tools that use an easy-tounderstand graphical categorization of knowledge (decision-aking logic) are: KnowledgeSeeker fro Angoss Software XpertRule Profiler fro Attar Software Business Miner fro Business Objects Cleentine fro Integral Solutions Ltd. Quadstone fro DecisionHouse (TreeHouse) Neural Network Soe tools that support parallel interconnecting coputing units learning to do tasks by a training process on available data are: Marksan fro HNC Software IBM s Intelligent Miner RedBrick s Mine Cleentine fro Integral Solutions Ltd. SAS Suite fro SAS Institute Predictive Modeling Soe tools that use identified patterns to predict the future are: Discovery Server fro Pilot Software IBM s Intelligent Miner Rule Discovery Soe tools that support autoatic identification and testing of hypotheses to generate rules and patterns are: Surveyor fro Distilleries Mineset fro Silicon Graphics IBM s Intelligent Miner Multiple Techniques Soe tools that support ultiple ining techniques including verification of stated hypotheses by pattern validation (top down), forensic analysis, discovery of unknown facts by rule induction and neural networks (botto up) are: IBM s Intelligent Miner Cleentine fro Integral Solutions Decisionhouse fro Quadstone IDIS fro Inforation Discovery SAS Suite fro SAS Institute Unfortunately, any of today s data ining tools tend to be available only for a liited nuber of database environents and platfors. In practice, tools offering liited platfor support are likely to be used only for specific, focused applications. While data ining is undoubtedly iportant to the enterprise, top anageent are unlikely to change their underlying database and operating systes infrastructure to deploy it. For these reasons, it is likely that any organizations will need to deploy several different ining tools to satisfy the full range of possible data ining applications, and there will

9 be a strong interest in web-enabled ining s. Integration is another weak area. ining technology is currently represented by a plethora of data ining tools, any of which are soewhat stand-alone in their first releases: Specialist freestanding tools fro sall-scale independent vendors (such as Cleentine fro Integral Solutions, Decisionhouse fro Quadstone, and XpertRule fro Attar Software). Specialist freestanding tools fro large ulti-functional software vendors (such as Intelligent Miner fro IBM, and Mineset fro Silicon Graphics). Integrated tools fro decision support vendors (such as Discovery Server fro Pilot Software, and SAS Syste fro SAS Institute). In the long ter, as we ove into the world of distributed software coponents, lack of interoperability is a severe liitation. But one short-ter advantage of freestanding tools over integrated tool suites is that copanies can get started with data ining without having to invest in a coplete technology or pay the price of a coplete suite. On the other hand, an integrated suite favors developers already using that suite for other developent work. They can probably adopt the data ining coponents quickly, while continuing to use other failiar coponents for related requireents such as data visualization. While soe tool integration is starting to appear (for exaple, Red Brick has introduced Red Brick Mine, which cobines its star schea data warehouse with data ining software fro Mind Corporation), we have yet to see a widespread erger of data ining and decision-support software. Currently, -based Figure 4. visualization exaple ining client tool or Web browser Extract/ saple LAN WAN or WWW Tier 1 Tier 2 Tier 3 Figure 5. A data warehouse syste SQL or DML ining tools with APIs (like IBM s Intelligent Miner, for exaple) hold the ost proise in this area. Mergers, acquisitions, and arketing agreeents will inevitably lead to further integration. Quality and Warehouse Access One of the biggest probles in data ining is the quality of the data. Unlike ordinary SQL queries, data ining techniques are highly sensitive to issing or inconsistent Mining Mining or file Single Hardware Server SQL SQL or DML or file data. This is particularly true for techniques that use sapling, which is a proble when starting with operational databases; ining saples are likely to contain issing or inaccurate data. One advantage of ining a data warehouse is that the tools are working with data of a ore unifor and predictable quality, because the data has already been cleansed as it was taken fro operational systes.

10 As we pointed out in the evaluation fraework (under Warehouse/OLAP Integration), to ake the best use of data ining techniques, tools need to integrate with a data warehouse (to ensure quality of data within a consolidated database) as well as with other interactive business analysis tools (to ensure visual coprehension of data ining results). However, the liitations of any of today s data ining tools noted above predoinantly standalone tools, supporting a narrow range of platfors liits their ability to interact with already established data warehouses. This in turn andates a ining approach involving extraction and iporting of data onto a different platfor for analysis. Deciding What to Mine The yth that ining cannot be done at the atoic level is not an arguent for data sapling. Mining detailed atoic-level transaction data ay take longer, but it usually eans that the analysis is ore exhaustive, leading to ore accurate knowledge discovery. When atoic level data is stored in an enterprise data warehouse, with suaries and soe detail data being available at the data art level, it is often the case that data ining is best deployed against the warehouse. On the other hand, there is a trend towards data art developent rather than enterprise data warehouse developent, because of the shorter tie scales involved and the faster return on investent. It ight see, therefore, that data ining projects would have to wait for the enterprise level warehouses to be deployed. Actually, this is not the case. ining can and does work well on suary data and/or atoic transactional data in a data art. However, because of the possibility that data suarization has obscured key aspects of a pattern, interoperability with OLAP tools becoes very iportant. Interoperability allows an analyst to develop an iterative approach to data ining and pattern verification, as we discussed earlier under Warehouse/ OLAP Integration. Future Directions While data ining technology is still in its infancy, there is no doubt where it is headed. The direction is clearly towards ining s, ining agents, and ining objects. The industry trend toward thin client architectures will pressure ining vendors to ake their products available fro web browsers and fro groupware frontends like Lotus Notes. The actual data ining algoriths, on the other hand, will ove as close to the data as possible. Given the trend toward universal (or objectrelational) DBMS s (such as Inforix s Universal Server, and IBM s DB2 Universal base), it would be logical to expect that ining algoriths will eventually becoe an integral part of the DBMS itself, supported by query optiization technology, and exploiting the DBMS s support for parallelis. That developent will not aterialize overnight. In the eantie, there will be fierce copetition aong the vendors of existing data ining tools, and an inevitable shakeout. Greater tool integration is inevitable, either through acquisition (vendor takeovers), consortia (arketing agreeents) or design (integrated design or developent agreeents). Multiplatfor support, and the evolution of key features including ining s, ulti-threading, scalability, published APIs, and web visualization are all going to becoe critical over the next two years. So before you dig into your gold ine of data looking to strike it rich, evaluate the data ining and DBMS arket carefully. References 1. Mining: Products, Applications, and Technologies, a research report by Two Crows Corporation. For ore inforation see