Abstract: This talk covers a "sequence" of solutions to cover three of the most important problems in data mining. First, it is rarely the case that in real-world applications that data is already in feature vector format. In reality,  data is normally semi-structured, e.g., transaction database, sequences, graphs, etc. Thus, in order to use most inductive learning method, one has to first come up with predictive features from these raw data. We discuss a method called Model-based Tree (or MbT) that search for highly predictive "frequent patterns" (such as frequent subgraphs, frequent sequences, sub-graphs etc) to construct feature vectors from raw data. The method has linear scalability and can find those features that can train model with accuracy higher than benchmark results on some of the most difficult problems.  Many of the feature constructed by MbT can be found any existing approaches due to combinatorial explosion.

After data is already in feature vector format, the next important question is "which inductive model to use"? This is a non-trivial selection process given the fact that there are many many inductive learning algorithms out there and choosing the best approach isn't straightforward. We discuss a method called Random Decision Tree (or

RDT) that is remarkably simple to use, works for all three major problems in inductive learning (classification, regression, and probability estimation) but performs quite well in many situations. Its main advantage is simplicity, accuracy and efficiency. One of its application on ozone forecasting has won the ICDM06 application best paper award.

The third important scenario is that training and testing data may even come from the same distributions. There is really a lot of practical reasons on collecting the best data. In the last part of the talk, we will discuss an effective approach based on locally weighted ensembles to transfer related knowledge from a related but a different domain into target domain (examples such as using Reuters data to predict New York Times data). The source data and data set of some of the solutions are available from the speaker's homepage.

Biography: Dr. Wei Fan received his PhD in Computer Science from Columbia University in 2001 and has been working in IBM T.J.Watson Research since then. He published more than 60 papers in top data mining, machine learning and database conferences, such as KDD, SDM, ICDM, ECML/PKDD, SIGMOD, VLDB, ICDE, AAAI, etc. Dr. Fan has served as Area Chair, Senior PC of SIGKDD'06, SDM'08 and ICDM'08, sponsorship co-chair of SDM'09, as well as PC of several prestigious conferences in the area including KDD'09/08/07/05, ICDM'07/06/05/04/03, SDM'09/07/06/05/04, CIKM'08/07/06, ECML/PKDD'07'06, ICDE'04, AAAI'07, PAKDD'09/08/07, EDBT'04, WWW'09/08/07, etc. Dr. Fan was invited to speak at ICMLA'06. He served as US NSF panelist in 2007/08. His main research interests and experiences are in various areas of data mining and database systems, such as, risk analysis, high performance computing, extremely skewed distribution, cost-sensitive learning, data streams, ensemble methods, graph mining, predictive feature discovery, transfer learning, feature selection and commercial data mining systems. He is particularly interested in simple, unconventional, but effective methods to solve difficult problems. His thesis work on intrusion detection has been licensed by a start-up company since 2001.  His co-authored paper in ICDM'06 that uses "Randomized Decision Tree" to predict skewed ozone days won the best application paper award.  His co-authored paper in

KDD'97 on distributed learning system "JAM" won the runner-up best application paper award.

More information about the speaker can be found at http://www.cs.columbia.edu/~wfan