Trend or no trend : a novel nonparametric method for classifying time series

Abstract

In supervised classification, one attempts to learn a model of how objects map to labels by selecting the best model from some model space. The choice of model space encodes assumptions about the problem. We propose a setting for model specification and selection in supervised learning based on a latent source model. In this setting, we specify the model by a small collection of unknown latent sources and posit that there is a stochastic model relating latent sources and observations. With this setting in mind, we propose a nonparametric classification method that is entirely unaware of the structure of these latent sources. Instead, our method relies on the data as a proxy for the unknown latent sources. We perform classification by computing the conditional class probabilities for an observation based on our stochastic model. This approach has an appealing and natural interpretation - that an observation belongs to a certain class if it sufficiently resembles other examples of that class. We extend this approach to the problem of online time series classification. In the binary case, we derive an estimator for online signal detection and an associated implementation that is simple, efficient, and scalable. We demonstrate the merit of our approach by applying it to the task of detecting trending topics on Twitter. Using a small sample of Tweets, our method can detect trends before Twitter does 79% of the time, with a mean early advantage of 1.43 hours, while maintaining a 95% true positive rate and a 4% false positive rate. In addition, our method provides the flexibility to perform well under a variety of tradeoffs between types of error and relative detection time.