Predictive Model Fusion: A Modular Approach to Big, Unstructured Data

Model fusion methods, or more generally ensemble methods, are a useful tool for prediction. Combining predictions from a set of models smoothes out biases and reduces variances of predictions from individual models, and hence, the combined predictions typically outperform those from individual models. In many situations individual predictions are arithmetically averaged with equal weights; however, in the presence of correlated models the fusion process is required to account for association between models. Otherwise, the naively averaged predictions will be suboptimal. This chapter describes optimal model fusion principles and illustrates the potential pitfalls of naive fusion in the presence of correlated models for binary data. An efficient algorithm for correlated model fusion is detailed and applied to algorithms mining social media information to predict civil unrest.