Mining Jumping Emerging Patterns by Streaming Feature Selection

One of the main challenges for making an accurate classifier based on mining emerging pattern is extraction of a minimal set of strong emerging patterns from a high-dimensional dataset. This problem is harder when features are generated dynamically and so the entire feature space is unavailable. In this scheme, features are obtained one by one instead of having all features available before learning starts. In this paper, we propose mining Jumping Emerging Patterns by Streaming Feature selection (JEPSF for short) using a dynamic border-differential algorithm where builds a new border of the jumping emerging pattern space based on an old border of the jumping emerging pattern space for new coming features. This framework completely avoids going back to the most initial step to build a new border of the jumping emerging pattern space. This algorithm helps reducing number of irrelevant emerging patterns, what in turns brings significant advantages to the presented algorithm. Thus the number of jumping emerging patterns is reduced and strong emerging pattern is also extracted. We experimentally represent effectiveness of the proposed approach against other state-of-the-art methods, in terms of predictive accuracy.