Sequence Prediction using Partially-Ordered Episode Rules

Predicting the next event (or symbol) of a sequence has many applications such as network event prediction, webpage prefetching, activity recommendation and stock price prediction. An effective approach for event prediction that has the benefit of being interpretable, is to extract episode rules to then apply them for prediction. This paper improves upon this approach by using a novel type of episode rules called partially-ordered episode rules. These rules are more general than standard episode rules as they loosen the ordering constraint on events in each rule antecedent and consequent. Hence, a partially-ordered episode rule can replace multiple standard episode rules. Experiments on several benchmark datasets show that partially-ordered episode rules can provide more accurate predictions than standard episode rules.

I. Introduction

Predicting the future is key to decision-making. To perform accurate predictions, it is necessary to consider previous observations and their temporal relationships with possible upcoming observations. Temporal observations are typically represented as either time series or discrete sequences [1]. A time series is an ordered list of numbers, generally measured at a fixed time interval. Predicting the next observation of a time series is typically viewed as a problem of finding a function that closely fit the data points. This can be done with techniques such as least square linear regression or more complex techniques. On the other hand, a discrete sequence is an ordered list of symbols. In this paper, the focus is on discrete sequences as their prediction is useful in many domains. For instance, it can be used to predict the next word that a user will type on a phone, the next error that will occur in a network, the next purchase of a customer, and the next location where someone will drive [2]. Because the nature of discrete sequences is much different than time series, different techniques are used for predicting the next symbol of a discrete sequence than for time series. One of the most accurate techniques for event prediction is artificial neural networks. However, a major drawback is that they mostly operate as black-boxes. Thus, a user is often unable to understand the reasons why an event is predicted. But developing explainable models is often critical for decision-makers in the industry. This is for example the case for network fault management, where network technicians wish to not only predict network errors but understand the relationships between complex network events to prevent errors [3].