Early classification of time series using multi-objective optimization techniques

Abstract

In early classification of time series the objective is to build models which are able to make class-predictions for time series as accurately and as early as possible, when only a part of the series is available. It is logical to think that accuracy and earliness are conflicting objectives, since the more we wait, more data points from the series are available, and it is easier to make accurate class-predictions. Considering this, the problem can be very naturally formulated as a multi-objective optimization problem, and solved as such. However, the solutions proposed in the literature up to now, reduce the problem into a single-objective problem by combining both objectives somehow. In this paper, we present a novel multi-objective formulation of the problem of early classification, and we design a solution using multi-objective optimization techniques. This method will provide a variety of solutions which find different trade-offs between both objectives, allowing the user to select the most suitable solution a-posteriori, depending on the accuracy and earliness requirements of the problem at hand. To prove the usefulness of our proposal, we carry out an extensive experimentation process using 45 benchmark databases and we present a case study in the financial domain.