CUBS: Multivariate Sequence Classification Using Bounded Z-score with Sampling

Abstract

Multivariate temporal sequence classification is an important and challenging task. Several attempts to address this problem exist, but none provide a full solution. In this paper we present CUBS: Classification Using Bounded Z-Score with Sampling. CUBS uses item set mining to produce frequent subsequences, and then selects among them the statistically significant subsequences to compose a classification model. We introduce an improved item set mining algorithm that solves the short sequence bias present in many item set mining algorithms. Unfortunately, the z-score normalization hinders pruning. We provide a bound on the z-score to address this issue. Calculation of the z-score normalization requires knowledge of some statistical values of the data gathered using a small sample of the database. The sampling causes a distortion in the values. We analyze this distortion and correct it. We evaluate CUBS for accuracy and scalability on a synthetic dataset and on two real world dataset. The results demonstrate how short subsequence bias is solved in the mining, and show how our bound and sampling technique enable speedup.