Event2vec: Learning Representations of Events on Temporal Sequences

Abstract

Sequential data containing series of events with timestamps is commonly used to record status of things in all aspects of life, and is referred to as temporal event sequences. Learning vector representations is a fundamental task of temporal event sequence mining as it is inevitable for further analysis. Temporal event sequences differ from symbol sequences and numerical time series in that each entry is along with a corresponding time stamp and that the entries are usually sparse in time. Therefore, methods either on symbolic sequences such as word2vec, or on numerical time series such as pattern discovery perform unsatisfactorily. In this paper, we propose an algorithm called event2vec that solves these problems. We first present Event Connection Graph to summarize events while taking time into consideration. Then, we conducts a training Sample Generator to get clean and endless data. Finally, we feed these data to embedding neural network to get learned vectors. Experiments on real temporal event sequence data in medical area demonstrate the effectiveness and efficiency of the proposed method. The procedure is totally unsupervised without the help of expert knowledge. Thus can be used to improve the quality of health-care without any additional burden.