In the emerging market for data mining tools the better to understand what you’ve got in your data warehouse, new technology from a tiny French start-up is commanding the respect of the likes of Hewlett-Packard Co and Arthur Andersen & Co. Datamind SA’s product, neurOagent, is a second-generation data analysis pacakge. Unlike first-generation, rules-based software that uses a data verification model to implement system rules based on experience, neurOagent launches an automatic discovery process, using technologies like artificial intelligence, induction and neural networks to tell you what’s interesting in the data and to predict tendencies, explains Khai Mihn Pham, founder and chairman of Datamind. Ramin Mikaili, business technology analyst for Andersen Consulting’s emerging technology solutions group in Chicago, explains that rules-based products are for users who have a clear picture of a market, such as stock brokers, who implement system rules according to their experience of when they want to buy or sell stocks. That is a saturated product area that has been around 10 to 15 years, he adds. In the second generation, he said, so-called `implicit’ techniques are used where the expert doesn’t know, but wants to determine the underlying patterns in the data to add to his existing knowlege and experience. It’s an area that is just emerging.
Fraudulent
Mikaili’s group did a study that applied four technologies for implicit learning, or data-mining, to a car manufacturer’s need to sift through warranty claims to determine their validity or fraudulence. The four were neural networks, inductive reasoning, statistical techniques (regression analysis) and Datamind’s neurOagent. Not only is Datamind capable of doing implicit learning on non-linear discrete data, like car warranties, but it enables you to do explicit knowledge description as well, says Mikaili. That makes it very attractive; it’s a cross of all four tools and knowledge-based tools. Also, it’s agent-based technology, so an agent can be a neural network or a statistical technique or a knowledge-based network, so it allows you to bring in other technologies. Dr Pham’s cross-breeding of technology is what interests Hewlett-Packard as well. Their algorithms are interesting. What I find attractive is their mixing different algorithms together and taking the best of both. It’s something I hadn’t seen before, a fresh approach. I don’t know yet how much of it is advanced thinking or just combining existing techniques, but that is in itself an advance, anyway, says Thierry Costa, international programme manager for Hewlett-Packard’s data warehouse project OpenWarehouse in Cupertino. Dr Pham, a French citizen of Vietnamese origin, developed the technology in the course of acquiring his two degrees, in medicine and information systems engineering. When I started out, I thought medical knowledge could easily be put into a computer, but when I was confronted with rule-based systems, I was shocked. It was too logical, he told Computergram.
By Marsha Johnston
So, beginning in 1987, he began trying to develop a new technology by putting myself into the skin of a doctor and determining how I reason. He says he didn’t try consciously to marry neural net and rules-based technologies, but that when I showed what I’d done to the artificial intelligence professors in Paris, they said, `No, that’s neural net stuff.’ So I went to see the neural net people and they said, `No, that’s rule-based technology.’ He established the company in France in 1991 with the idea to distribute a particular technology that unified the symbolic and neurologic aspects of data analysis technology. Pham says most of the US companies with second-generation products have neural networks to predict and use rules to understand the data, but that they use two separate logic mechanisms, making it virtually impossible to connect the two. He says these companies will often say that rules allow users to make predictions, but says that in fact you have to r
ework the rules to construct an operational prediction system. The big difference is that our logic mechanism manages both explicit data and learning, he explains. Although the company has seen interest from Oracle, Informix and Sybase, it hasn’t pushed ahead with them because HP and Red Brick are the most active, Pham says. Hewlett-Packard has adopted Datamind into three of the four areas of its OpenWarehouse program: as a partner; in its consultancy, where Hewlett-Packard recommends Datamind’s product and works with the company to determine the best way for a client to use it; and integrating its product into Hewlett-Packard’s Intelligent Warehouse offering. We have agreed in principle to do it, but we haven’t started on the technology yet, Costa says, adding that testing will begin this month.
Enormous market
Content from our partners
Pham says Hewlett-Packard presents the company to its clients who have data warehouses and that the two do joint communications on data mining. All of the market players are pushing data warehousing because it’s an enormous market. They realised quickly that people needed tools to really help them understand the data. IBM has lots of people working on data mining, but has no product out yet, so Hewlett-Packard wants very much to push ahead. We have started to port the product to their machines, he said. The company has officially had offices in California, mostly administrative, since July, 1994. In early March this year it brought over the majority of its research and development team. We need strong R&D here to be able to carry out our partnerships, Pham said, adding that the company is also in the process of hiring the vice-president of marketing for Red Brick Systems Inc, the number two data warehouse company after Oracle Corp. Stewart Schuster, a co-founder of Sybase Inc, has also joined the board of directors to assist with marketing strategy, for which it has just received $3.4m in venture capital. Says Mikaili, I like to learn new things, and this was real exciting, to see the development of a new technology, it was like being around when neural networks or the theory of relativity were discovered. It’s very innovative and applicable to many of today’s problems.
