First International Workshop on Mining Multiple Information Sources

This paper demonstrates how methods borrowed from information fusion can improve the performance of a classifier by constructing (“fusing”) new features that are combinations of existing numeric features. This work is an example of local pattern analysis and fusion because it identifies potentially useful patterns (i.e., feature combinations) from a single data source. In our work, we fuse features by mapping the numeric values for each feature to a rank and then averaging these ranks. The quality of the fused features is measured with respect to how well they classify minority-class examples, which makes this method especially effective for dealing with data sets that exhibit class imbalance. This paper evaluates our combinatorial feature fusion method on ten data sets, using three learning methods. The results indicate that our method can be quite effective in improving classifier performance, although it seems to improve the performance of some learning methods more than others. General Terms Algorithms, Performance, Experimentation