Combining Disparate Information for Machine Learning.

Abstract

This thesis considers information fusion for four different types of machine learning problems: anomaly detection, information retrieval, collaborative filtering and structure learning for time series, and focuses on a common theme -- the benefit to combining disparate information resulting in improved algorithm performance.

In this dissertation, several new algorithms and applications to real-world datasets are presented. In Chapter II, a novel approach called Pareto Depth Analysis (PDA) is proposed for combining different dissimilarity metrics for anomaly detection. PDA is applied to video-based anomaly detection of pedestrian trajectories. Following a similar idea, in Chapter III we propose to use a similar Pareto Front method for a multiple-query information retrieval problem when different queries represent different semantic concepts. Pareto Front information retrieval is applied to multiple query image retrieval. In Chapter IV, we extend a recently proposed collaborative retrieval approach to incorporate complementary social network information, an approach we call Social Collaborative Retrieval (SCR). SCR is applied to a music recommendation system that combines both user history and friendship network information to improve recall and weighted recall performance. In Chapter V, we propose a framework that combines time series data at different time scales and offsets for more accurate estimation of multiple precision matrices. We propose a general fused graphical lasso approach to jointly estimate these precision matrices. The framework is applied to modeling financial time series data.