A Latent Source Model for Online Time Series Classfication

We study a binary classification problem with infinite time series having more than two labels (”event” and ”nonevent” or ”trending” and ”non-trending”). We want to predict the label of the time series given some training data set. Intuitively, the longer we wait, the longer the time series we can observe so that the prediction is more accurate. However, in many applications, such as predicting which topic will go popular in a social network or revealing an imminent market crash, making a prediction as early as possible is highly valuable.Motivated by these applications, we look into a latent source model which is a nonparametric model to predict the binary status of a time series. Our main assumption is that these time series only have a few ways to reach the binary status such as Twitter topic going trending online. The latent source model naturally leads to a weighted majority voting as the classification rule without knowing the latent source structure. In the project: 1. We will investigate the theoretical performance guarantees of the latent source model; 2. We will implement the model by programming language C; 3. We will investigate the strategy to estimate the values of different parameters; 4. We will test our implementation and use the model to predict which news topics on Twitter will go viral to become trends and analyze the results.