| Author: | Wang, Zhitao |
| Title: | Information diffusion prediction in social media |
| Advisors: | Li, Wenjie Maggie (COMP) |
| Degree: | Ph.D. |
| Year: | 2020 |
| Subject: | Online social networks Social media Sociotechnical systems Information behavior Hong Kong Polytechnic University -- Dissertations |
| Department: | Department of Computing |
| Pages: | xvii, 165 pages : color illustrations |
| Language: | English |
| Abstract: | The emergence of online social media has fundamentally changed the way of information diffusion in human society. These changes trigger a large amount of information diffusion processes in our daily life. Thanks to various social media platforms, the diffusion data become retrievable and traceable, providing unprecedented opportunities as well as great challenges for information diffusion studies. Modeling and predicting information diffusion in social media are meaningful for a variety of real-life applications, such as recommendation services, social marketing campaigns and stability maintenance. In this thesis, we investigate the information diffusion prediction problem from a micro perspective, which aims to model how individuals in social media affect each other in the information propagation process and predict potential individuals who will participate in the future. Specifically, we identify three key research problems to be solved in information diffusion prediction task, i.e, 1. How to develop an effective prediction model when only historical diffusion processes are observed? 2. How to utilize additional social relationship network to improve the prediction performance and generalization ability of the diffusion model? 3. How to capture interplay effect between information diffusion and social network, and predict diffusion processes and network links jointly? Inspired by great success of representation learning and neural network techniques on various fields, we propose several models based on the two powerful frameworks to solve the above problems. According to the category of the proposed models, this thesis is naturally divided into two parts. In the first part (work 1 and 2), we present two representation learning based models for problem 2 and problem 3, respectively. The poor generalization of discrete graph-based models in previous work motivates us to project diffusion users into a continuous latent space as user representations, which capture unique diffusion characteristics of users. In the representation space, any possible diffusion influence can be flexibly measured by the distance between user representations. In work 1, we focus on problem 2, which asks for integrating social network structure information into diffusion prediction model. In this work, a novel network-regularized role-based representation learning model is proposed. The model learns the user representations based on the objective of maximizing the likelihood of observed diffusion cascades and employs another objective of reconstructing structural proximities as a regularization. The network regularization provides additional constraints on representation learning, correcting the biased information or supplementing the missed information of diffusion relationships. In work 2, we move the attention forward on problem 3, which aims to capture correlations between diffusion cascades and network structure and predict diffusion processes and network links jointly. To achieve this goal, we propose a joint user representation learning model that embeds users as shared representations in a common latent space to characterize their behaviors correspond to both information diffusion and link creation. The proposed model defines two consistent objectives with maximization likelihood estimation on two behaviors and incorporates them in a unified learning framework. The shared representations latently capture the interplay effects and improve the generalization ability on both prediction tasks. The second part (work 3 and 4) delves into neural network based solutions for problem 2 and problem 1, respectively. Due to the sequential form of diffusion processes, most recent studies formulated diffusion prediction as a sequence prediction task and employed recurrent neural network (RNN) for the problem. However, few existing RNN-based models consider the observed network structure when modeling diffusion cascades (sequences), which means that they cannot be applied to problem 2. Therefore, in work 3, we propose a novel sequential neural model with structure attention to inject network structure information. The RNN framework is employed to model the sequential information. A structure attention mechanism is designed to capture the important structural information of diffusion users in the given social network. A gating mechanism is further developed to effectively integrate the sequential and structural information. With the injection of structure information, the prediction performance and the generalization ability are further improved. Although retrievable diffusion processes are recorded in the sequential form, we find that non-sequential properties exist, which do not strictly follow the assumptions of previous RNN-based work. In work 4, we propose a hierarchical diffusion attention network with a non-RNN framework for problem 1, which asks to predict diffusion without knowing the underlying social network. The model adopts two-level attention mechanisms, i.e., a dependency attention at user level for capturing historical user-to-user dependencies, and a time-aware influence attention at the cascade (sequence) level for inferring possible future user's dependencies on historical users. The evaluations demonstrate our non-sequential attention network is more effective than previous RNN-based sequential models. In summary, we present a systemic study of information diffusion prediction in social media. The effectiveness of the proposed models is demonstrated on public benchmark diffusion datasets or synthetic datasets. The proposed models will benefit a wide range of potential applications in real life, such as advertisement recommendation, product influence optimization and personal opinion prediction. Moreover, the possible extensions of current work will provide a deeper and more comprehensive understandings on diffusion related behaviors in social media. |
| Rights: | All rights reserved |
| Access: | open access |
Files in This Item:
| File | Description | Size | Format | |
|---|---|---|---|---|
| 991022385355703411.pdf | For All Users | 2.55 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/200/10489