Classification of Temporal and Sequential Data Using Bag-of-Subsequences Features

Classifying sequential data is an important problem in machine learning with applications in time series, sensor streams, and image analysis. The ordered structure of sequential data presents a difficulty for the standard classification models, which has motivated the task of generating features for vector-based discriminative models. Shapelet methods, which have been extensively studied in this topic, identify class-specific subsequence patterns and exploit dissimilarity from them. Their approach merits the classification performance and the visual interpretability of the discriminative patterns. In this paper, we aim to improve on the robustness of the pattern-based discriminative model by considering the sequential data instances a "bag" of characteristic patterns. We propose a data cleaning method for generating small clusters of class-specific patterns and a vector representation which features the distance between clustered patterns and the sequential instance seen as a bag-of-subsequences. We further employ max-margin learning and feature selection to extract few class-characterizing patterns. We present numerical experiments using time series and silhouette image data and compared the performances of the proposed methods with other time series classification models. Our visual analysis of the extracted features showed that the proposed method can extract characteristic subsequences of the data.