Hypergraphical Models for Object Tagging

Description

The extraordinary volatility of user-generated objects (UGO) has presented the current search engines and social platforms with enormous—and growing fast—difficulties due to inaccuracy and inconsistency of tags, even with the benefit of the Wiki-style user input. Tag modeling and categorization have therefore become crucially important for sustainable online and mobile lifestyle for the user and commerce for the industry. In reality, developing new technologies for tag modeling has become imperative in this 4G era. In this project, we propose to investigate the nature of tags and topics in UGO and develop new techniques for tagging UGO data.With fast growth of Web technologies, we have witnessed an exponential explosion of UGO (e.g., blogs/messages, photos/images, music, video, articles, etc.) and the mushrooming of a plethora of online communities (e.g., del.icio.us, flickr, YouTube, Facebook, arXiv, CiteULike, Mendeley, etc.) that enable users to share UGO. In addition, such massive UGO have gained considerable momentum as ubiquitous computing, sensor technologies and 4G technologies have become more mature and widely accessible. Unfortunately, techniques that can better serve the user by taking advantages of the various properties/attributes in UGO for effective filtering, searching and organization have not kept pace with data growth. This is a challenging problem because of the multi-tagging nature of individual objects. For example, an article may focus on multiple topics and an image may have multiple regions of interests. In this project, we will work on a substantial project for object tagging to help users organize, filter, search and use UGO efficiently.Tags include user-provided labels and latent semantic labels inferred from object contents. Traditional collaborative filtering (CF) techniques based on user-provided tags are often found to be inefficient, inadequate and ineffective because users to don’t have the same level of diligence and consistency in tagging making search difficult and wasting huge amount of memory due to unnecessary data duplications. UGO is a growing and dynamic data set and in a sense we need a “life-long learning” mechanism to adapt to the changing data. It is important to investigate the user behavior data of online communities, which will help us gain new insights into the object-tagging problem to be used in our system. Developing new models and algorithms that combine community structures into object tagging is therefore a large component of this research. In this project, we will investigate new hypergraphical models for tagging UGO data. Using hypergraphical representations, we can bridge two statistical models, hierarchical Bayesian models (HBMs) and Markov random fields (MRFs), for both the observed/latent tags and complicated tag-object relations within a unified framework so as to fulfill the following criteria: 1)accuracy(assigning correct tags to each object); 2)scalability(fast tagging for big and changing UGO data); and 3)interpretability.Based on our extensive research experience in the relevant areas, we are confident that, by combining HBM and MRF based on hypergraphs, we will be able to develop techniques to achieve a superior performance for multi-tagging UGO data. Through this study, we expect to achieve the following goals: 1. Establish theoretical foundation and key algorithms of the hypergraphical models that combine HBMs and MRFs within a unified framework. 2. Investigate the multi-tagging properties of UGO data, especially the “recognition by component” theory to represent each object as “a bag of features”. 3. Develop multi-tag techniques based on the hypergraph-based HBM to model the property of UGO data. 4. Use the hypergraph-based MRF to model and investigate user behaviors in some major online communities. 5. Develop fast and memory-efficient inference and learning algorithms to handle big and changing UGO data using the proposed hypergraphical models.?