Traditional Data-Mining Tools Are Too Complex for Nonexperts

Many companies store information in data warehouses in an attempt to derive value from the massive amount of transactional data collected. Companies want answers to key business questions, such as: Who are our most valuable customers? How can we design and target promotions to increase sales? Which customers are most likely to leave in the future and how can these customers be retained? Corporations could get some answers to these questions by taking a retrospective look at already collected data. Relational databases, online analytical processing (OLAP), and statistical tools are used for historical data analysis, which assists users in answering questions such as: How much did sales increase in the eastern region over the past two quarters? These tools have become invaluable because they have sophisticated graphical user interfaces that lead the person with the question directly to the answer, making historical information accessible to the average businessperson.

By 1996, several generic data-mining tools were available to predict future behavior, but, to be used effectively, they required scarce and expensive artificial intelligence and machine-learning expertise. Predictive models based on traditional data-mining tools pose two significant limitations: they require knowledge in machine-learning technology, a skill the average business user does not possess; and, for every query the business user poses, a specific model responsive to the query has to be built.

Building separate models in response to each query is time consuming and requires input from data-mining experts. Continuum recognized that the lengthy process of hiring experts and building specific models was too slow for this age of rapid data exchange. By the time a question has been asked and then simulated, the answer is often irrelevant, thereby thwarting the company's ability to gain a competitive advantage.

Continuum Intends to Predict Customer Behavior with Future Database

Using machine-learning, statistical, and visualization techniques to discover and present knowledge in a form that could be accessed quickly and is easily comprehensible, however, required an innovative approach.

Generic data-mining tools were available topredict future behavior, but they required scarce
and expensive artificial intelligence and machine-learning expertise.

Continuum planned to create a unique application to build simulators using data-driven, machine-learning technology. The simulators would be used to simulate what customer behavior would be like months into the future by extrapolating from the behavior of existing customers. That information would then be used to create a predictive database in the same format as a historical database.

The business user would be able to extract, display, and analyze information from this future database using the same familiar relational, OLAP, and statistical tools that are currently used for extracting information from historical databases. Continuum's simulator also would be able to respond to questions that involve dynamic situations, unlike conventional data-mining models, which are only capable of analyzing problems involving static situations. The future database could answer a question such as: What would happen if a company lowered its prices? To answer the question, a company would start its projection with the original customer base and would project results both with and without the hypothetical event of lowering prices. Then the company could compare the two projected futures, identify the customers who would respond to the lower prices, and determine what total revenues and costs would be with and without the action.

Continuum planned to investigate extending current technology to handle huge amounts of customer data found in data warehouses to enable the construction of simulators for creating a future database of those customer records. The future database would contain future records projected from current records using projective visualization (PV), a new and unique application of machine learning that is able to analyze huge amounts of warehoused data, even when the records in the warehouse are incomplete. Unlike conventional machine-learning techniques, which only provide an answer to the specific question for which they are created, PV could be used to answer many questions, just like historical databases.

Prior to the ATP project, Continuum had already received the enthusiastic support of companies intrigued with the potential impact the future database could have on their businesses, but the nature of the research was still too high risk to attract funding. To demonstrate the applicability of the future database technology, Continuum planned to work with Switchboard.com to provide training sets of data and test domains for the application of the future database. Because Links2Go was the most promising test domain, Continuum decided to work with it rather than with Switchboard's. However, Continuum's relationship with Switchboard ended amicably.

Predictive Simulation Tools Promise Powerful, New Data-Analysis Capabilities

When Continuum approached ATP for funding to continue its research and development in 1997, ATP recognized that the company's machine-learning techniques would advance the slow and costly practice of building simulation models to analyze customer behavior. Continuum could tap into the explosive use of the Internet, and the increasing volume of data being generated and warehoused in numerous web sites, to address new opportunities that were impossible with the data-analysis methods then available.

The market for data-analysis tools was growing rapidly, and the commercialization of Continuum's predictive data models also had potential for market spillover.

Continuum organized and classified tens of millions of web pages solely on predicted user interest.

Continuum Validates Projective Visualization

The principal technical goal of this project was to develop a software tool that would enable the creation of databases that could predict future behavior. The following requirements had to be fulfilled to validate the use of PV:

1. Increase the quantity of training data.
   Results:The company made significant progress in its ability to load and analyze large training sets of data in a reasonable time. At the beginning of the project, Continuum was able to analyze 600,000 training instances on a machine with 1 gigabyte (GB) of random access memory (RAM). By the end of the project, it was able to analyze 7 million instances per GB of RAM and had built systems with as many as 40 million training instances (almost a hundred fold increase).
2. Increase the quality of training data.
   Results: The company developed numerous automated and semi-automated techniques for enriching data sources and for fusing additional data sources, allowing systems with greater accuracy to be built for fixed training sets.
3. Identify the most effective machine-learning techniques for specific problem types.
   Results: Continuum examined a large number of alternative machine-learning techniques and integrated the best ones into a unified tool set that can be applied to a wide variety of problems.
4. Develop effective user interfaces for complex information.
   Results: The company developed and deployed interfaces for visualization and manipulation of the system during the course of development.
5. Validate results on selected test domains.
   Results: Continuum monitored software performance by applying the predictive technology to selected test domains to examine its applicability and to receive feedback on the system's functionality.
   

Continuum achieved these technical requirements and developed a software tool that uses PV to project the effects of an action on the basis of prior behavior. Not only was Continuum able to develop the necessary tools for a future database, it also organized and classified tens of millions of web pages solely on predicted user interest.

Continuum Refocuses on Content Classification

In one of the initial test domains, referred to as Links2Go, the results of the company's work far exceeded any of the other test domains. Continuum realized that its technology also could be used to project user interest in sets of topics and pages in the context of an online research tool, which would allow the classification of related web data. Continuum was so impressed with the capabilities of the research tool that it set out to further develop and commercialize the technology for the classification of web content.

The Links2Go test site was well received because it had the ability to automatically organize web content and to provide end users with highly relevant links pertaining to the topic of their query. Typical search directories rely heavily on human editors; these directories have the ability to manually classify several million pages and are organized on the basis of the editor's subjectivity. In contrast, the Links2Go directory automatically organized 70 million web pages by topic, thus providing a significantly higher number of relevant retrievals. Furthermore, competing search engines typically refresh their web pages every 30 days; during this lag time, they could potentially send web users to expired uniform resource locators (URLs). The Links2Go directory was refreshed overnight, thus ensuring that users were viewing the most current versions of the web pages in the directory. With no marketing or public relations effort, the Links2Go web site experienced more than one million distinct visitors per month.

TopicalNet used the predictive technologydeveloped as a result of the ATP project to
provide businesses with a software solution to classify massive amounts of related, electronically
stored data into easily accessible topics.

The tool was able to read a document and assign it to a topic on the basis of knowledge gained from the topical directory. The tool also could build a taxonomy and decide where the document should fit within that categorization scheme.

Test Domain Leads to Compelling Business Opportunities

By the midpoint of the project in 1999, Continuum executives decided that the commercial consequences of the technology emerging from the ATP project were so significant that they formed a new company. In August 1999, Continuum Software, Inc., became Links2Go.com, Inc. The new company decided to focus on the Links2Go web site after it closed a $4 million investment deal with the venture capital firm, Bertelsmann, Inc. This investment financing enabled Links2Go to begin to aggressively commercialize the initial outcomes of its research. Links2Go not only provided end users with a powerful topical search and directory tool, it also offered businesses a variety of unique options such as targeted web advertising.

The Links2Go directory gave businesses the opportunity to advertise and receive competitive intelligence. Businesses were able to target their web advertisements to all links and keyword searches on topics related to their particular business areas. The businesses would receive substantial access to their specific markets by targeting web users with an interest in topics pertaining to the companies' scope of business. Links2Go also offered statistics on a company's web site traffic and how the site compared with competitors' sites.

Links2Go approached several companies in the WWW market segment that could benefit from its technology. Links2Go proposed to allow these companies to use Links2Go's vertical technology for six months at no cost, with payment to begin after the trial period. After the six months was up, however, Links2Go had difficulty collecting payment for its tool. It continued to market its services, with limited success. When new Chief Executive Officer Ray Kingman joined the organization in October 2000, the company began to change its focus to further expand the classifier technology developed during the ATP-funded project. At this time, the company was renamed TopicalNet, Inc., to commemorate a new step in its path to commercialization.

TopicalNet used the predictive technology developed as a result of the ATP project to provide businesses with a software solution to classify massive amounts of related, electronically stored data into easily accessible topics. The technology can classify information from the Internet, corporate intranets, and extranets. Content classification is an ever-increasing need among businesses and end users; in fact, it is estimated that 80 percent of the content within an enterprise is unstructured. As information continues to be created and stored, content classification is essential in quickly obtaining accurate, relevant information. By 2004, the content classification and web analytics industry is predicted to grow to an estimated $2.2 billion. TopicalNet's technology has the potential to significantly impact this new and expanding market.

Conclusion

The ATP grant offered Continuum (later renamed Links2Go and then TopicalNet) the means to explore the possibilities of a predictive modeling tool, which led to the development and commercialization of the company's content classification technology. ATP's support permitted the company to examine the benefits of this technology without assuming the total risk. The project's successful completion led to additional funding and commercial viability in the content application and web analytics industry. TopicalNet's future is bright as it makes inroads in this growing market

Research and data for Status Report Status Report 97-01-0087 were collected during April - June 2002.

Return to Table of Contents or go to next section of Status Report No. 3.

Date created: April 4, 2006
Last updated: July 11, 2006

ATP website comments: webmaster-atp@nist.gov  / Technical ATP inquiries: InfoCoord.ATP@nist.gov.

NIST is an agency of the U.S. Commerce Department
Privacy policy / Security Notice / Accessibility Statement / Disclaimer / Freedom of Information Act (FOIA) /
No Fear Act Policy / NIST Information Quallity Standards / ExpectMore.gov (performance of federal programs)