Behavioral sequence prediction for evolving data stream

Abstract

Behavioral pattern prediction has many applications, ranging from consumer buying behavior analysis, web surfing prediction to network attack prediction. The traditional behavioral prediction technique works mainly on a fixed dataset. But recent advances in digital technology generates a huge amount of data which contributes to data stream. Data evolves over time due to the concept drift. Stream-based classification also needs to evolve over time. Our goal is not to predict a single action/behavior, but a sequence of actions that can occur later depending on the previous actions. We call this problem “Behavioral Pattern Extrapolation”. In our research, we exploited a stream mining based technique along with Markovian model, where we used an incremental and ensemble based technique for predicting a set of future actions. We have experimented using a number of benchmark datasets and shown the effectiveness of our approach.