Learning from heterogeneous data by Bayesian networks
- Publication Type:
- Issue Date:
Non-i.i.d. data breaks the traditional assumption that all data points are independent and identically distributed. It is commonly seen in a wide range of application domains, such as transactional data, pattern recognition data, multimedia data, biomedical data and social media data. Two challenges of learning with such data are the existence of strong coupling relationships and mixed structures (heterogeneity) in the data. This thesis mainly focuses on learning from heterogeneous data, which refers to the non-i.i.d. data with mixed structures. To cater for the learning from such heterogeneous data, this thesis presents a number of algorithms based on Bayesian networks (BNs) that provide an effective and efficient method for representation of heterogeneous structures. A wide spectrum of non-i.i.d. data with different heterogeneity is studied. The heterogeneous data investigated in this thesis includes sequential data of unequal lengths, biomedical data mixed with time series and multivariate attributes, and social media data with both user/user friendship networks and user/item preference matrix. Specifically, for modeling a database of sequential behaviors with different lengths, latent Dirichlet hidden Markov models (LDHMMs), are designed to capture the dependent relationships in two levels (i.e., sequence-level and database-level). To learn the parameters of the model, we propose a variational EM-based algorithm. The learned model achieves substantial or comparable improvement over the-state-of-the-art models on predictive tasks, such as predicting unseen sequences and sequence classification. For learning miscellaneous data in clinical gait analysis, whose data consists of both sequential data and multivariate data, a correlated static-dynamic model (CSDM) is constructed. An EM-based framework is applied to estimate the model parameters and some intuitive knowledge can be extracted from the model as by-products. Then, for learning more complicated social media data that records both the user/user friendship networks and user/item preference (rating) matrix in social media, we propose a joint interest-social model (JISM). We approximate the lower bound of the likelihood of the observed user/user and user/item interaction data and propose an iterative approach to learn the model parameters under the variational EM framework. The learned model is then used to predict unknown ratings and generally outperforms other comparison methods. Besides the above pure BNs-based models, we also propose a hybrid approach in the context of the sequence anomaly detection problem. This is because the estimation of the parameters of pure BNs-based model usually falls into local minimums, which may further generate inaccurate results for the sequence anomaly detection. Thus, we propose a model-based feature extractor combined with a discriminative classifier (i.e., SVM) to overcome the above issue, which is theoretically proved to have better performance in terms of Bayes error. The empirical results also support our theoretical proof. To sum up, this dissertation provides a novel perspective from Bayesian networks to harness the heterogeneity of non-i.i.d. data and offers effective and efficient solutions to learning such heterogeneous data.
Please use this identifier to cite or link to this item: