In the previous week we had an excellent data mining conference in Warsaw – Rough Sets and Current Trends in Computing (RSCTC). Several months ago, TunedIT had organized the Discovery Challenge for RSCTC: analysis of genetic data for medical purposes. Now, there was a challenge session where the winners presented their solutions to general public. Everyone was really curious how they did it and many questions followed after their talks, so they had no choice but to lift the curtain on their secret tricks. If anyone still wants to learn more, I recommend looking into the challenge paper – to be found here or in conference proceedings (pp. 4-19). We’ll also post shortly an interview with one of the winners, so stay tuned!
Apart from the contest, the conference brought many interesting presentations. First of all, there were four invited keynote talks given by prominent researchers, professors: Roman Słowiński, Sankar Pal, Rakesh Agrawal and Katia Sycara.
Rakesh Agrawal is the head of Microsoft Search Labs, responsible for the development of Microsoft’s Bing search engine. In his talk, Search and Data: The Virtuous Cycle, he sketched what kinds of data mining problems they face when trying to make Bing more “intelligent”, so that search results contain exactly the pages that the user is looking for. It appears that one of the toughest problems is to discover real intentions of the user: what is he really looking for? Search engine knows only the query string, usually very short (1-2 words,often misspelled), say “Ireland”, and must guess what the user expects: travel guide for a tourist or geographical facts about the country? Another problem is that many words have several different meanings: if the user writes “polish” does it mean a verb, “to polish”, or an adjective, “Polish”? Yet another problem: how to deal with numbers in a smart way? The query “$200 camera” gives few sensible results if treated literally – better try “$199 camera” 🙂
Many more issues of this kind must be dealt with. Add that the algorithms must dig through petabytes of data in a matter of seconds, and you’ll have no doubts that guys in Microsoft Search Labs never complain about boring assignments. BTW, I must confirm from own experience that data size and performance requirements are critical factors to make data mining fun. With small data and no performance difficulties, data mining is just an interesting thing to do. When performance begins to play a role, you discover that 95% of your fantastic algorithms just don’t catch up and you’ve got to turn all the bright ideas (and software) upside down.
Another talk which I really enjoyed – Emergent Dynamics of Information Propagation in Large Networks – was delivered by Katia Sycara from Carnegie Mellon University. It’s interesting to observe how large networks of “agents”, for example people, share information among themselves on a peer-to-peer basis, like through gossiping, and how the information fills the whole network at some point in time or – conversely – suddenly disappears. It’s important that we can predict evolution of such processes, because in real world the “information” distributed may be an infectious disease whose spread should be stopped as soon as possible; or an operator’s request that must be distributed to all computers in a large decentralized network, in a shortest possible time.
Which outcome is observed depends on different parameters of the network: how many connections there are between agents, what’s the topology (uniform connections? separated clusters?), how keen the agents are to pass the gossip further on. But what’s the most interesting is that Read more of this post