| Author: | Lu, Wei |
| Title: | Probabilistic graphical modeling for latent feature learning |
| Advisors: | Chung, Fu-lai Korris (COMP) |
| Degree: | Ph.D. |
| Year: | 2017 |
| Subject: | Graphical modeling (Statistics) Probabilities. Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Computing |
| Pages: | xx, 114 pages : color illustrations |
| Language: | English |
| Abstract: | With increasing availability of digitized knowledge, it has been increasingly important to develop statistical models to manage large-scale and high-dimensional heterogeneous data, making hierarchical learning on these various kinds of data a challenging problem. Despite the extensive research on hierarchical topic mining and deep representations, there are still numerous issues that have not been sufciently addressed, such as dealing with sparsity issues, interpretability of inner structure, serendipitous recommendation and transfer of learned deep features to new domains. To overcome these challenges, there is pressing need to develop hierarchical learning methods for various kinds of dataset problems with diverse feature sets. The aims of this work are to develop novel probabilistic graphical models that can automatically learn good feature representation from sparse data using multiple sources and types of auxiliary data, and apply the models to machine learning tasks including semantic topic understanding, video recommender system and unsupervised/semi-supervised image classification. Targeting at the sparsity issue of text data applications, the first two approaches are introduced from topic modeling perspectives. Firstly, we investigate how auxiliary information can benefit content analysis for hierarchical topic mining when the text length are biased short. Through incorporating relational meta information, this algorithm takes advantage of the natural hierarchical structure and infers topics by jointly modeling word and taxonomic node assignments for documents. Secondly, addressing the sparseness phenomenon in a recommender system application scenario, instead of regard one of the two observations as auxiliary information, we consider the problem in a collaborative way. Motivated by a real world online video recommendation problem, we target at the long tail phenomena of user behavior and scarceness issues of item features, and propose a personalized compound recommendation framework for online video recommendation called Dirichlet mixture probit model for information scarcity (DPIS), a probit classi.er utilizing record topical clustering on the user part for recommendation. The third and fourth models also start from an unsupervised perspective while incorporating multi-layer features for recommendation and domain adaptation. The third model is based on a useful approach for complex multi-relational data learning and missing element completion from a tensor perspective, where a deep probabilistic tensor decomposition model for item recommendation and tag completion is proposed. We also apply the proposed algorithm to computational creativity, an emerging domain of application, emphasizing the use of big data to automatically design new knowledge, resulting to attain serendipitous recommendation. The fourth model is based on multi-layer sparse factorization. Deep architectures can now be well trained on massive labeled data. However, there exist many application scenarios, where labeled data are sparse or absent. Domain adaptation and multi-task transfer learning provide attractive options when related labeled data or tasks are abundant from di.erent domains. In this part, a new graphic modeling approach to multi-layer factorization based domain adaptation is explored to address the scenarios that sufficient labeled data are available from the source domain while only a small subset or no labeled data can be used for supervised learning. A deep convolutional factorization based transfer learning (DCFTL) is proposed to facilitate layer-wise transfer learning between domains. Completely based on graphical model representation, the proposed framework can seamlessly merge inference and learning, and has clear interpretability of conditional independence. The empirical performances on image classification tasks in both supervised and semi-supervised adaptation settings illustrate the effectiveness and generalization of knowledge transfer framework. |
| Rights: | All rights reserved |
| Access: | open access |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| b29616785.pdf | For All Users | 3.83 MB | Adobe PDF | View/Open |
Copyright Undertaking
As a bona fide Library user, I declare that:
- I will abide by the rules and legal ordinances governing copyright regarding the use of the Database.
- I will use the Database for the purpose of my research or private study only and not for circulation or further reproduction or any other purpose.
- I agree to indemnify and hold the University harmless from and against any loss, damage, cost, liability or expenses arising from copyright infringement or unauthorized usage.
By downloading any item(s) listed above, you acknowledge that you have read and understood the copyright undertaking as stated above, and agree to be bound by all of its terms.
Please use this identifier to cite or link to this item:
https://theses.lib.polyu.edu.hk/handle/200/8927