毕业生知识库

With the emergence and rise of social media, people now usually engage in multiple online social networks simultaneously to serve for their diverse information needs. The cross-network user participation phenomenon is becoming more and more frequent. Meanwhile, to better improve user experience and increase user adoption, more and more internet companies also begin to release cross-network features and seek for strategic corporation, cross-network collaboration is gradually becoming the future trend of internet development. These all provide huge possibility for cross-network user behavior analysis and lay foundation for the cross-network collaborative application.

 

Compared with traditional single-network situation, cross-network analysis and application involves with a large amount of scattered information. These scattered information has some characteristics, such as heterogeneity, network gap and temporal dynamics, which make it a big challenge to conduct cross-network data integration and exploitation. How to extract valuable knowledge from these information for comprehensive user understanding and further design effective cross-network collaborative application is very important and promising. On one hand, the cross-network user analysis and application can help better understand the typical user interest from different perspectives and effectively address the cold start and data sparsity problems under single-network circumstance; On the other hand, it can also facilitate the understanding of unique network characteristics, so as to complement respective advantages between different networks and lead to brand new business model. Therefore, in this work, based on the cross-network user behavior analysis and aiming at designing novel collaborative application, we have conducted the following research work on three different aspects of common user behavior pattern analysis, cross-network data integration and cross-network user interest integration:

1. We propose to directly analyze the micro user behavior patterns in different social media networks from user level. Under cross-network circumstance, the user behaviors usually distribute on different networks, and the user behaviors of the same user in different networks have some sort of consistency and difference. Analyzing these characteristics can better understand the typical user interest and behavior patterns. Therefore, we conduct cross-network user analysis on two aspects, i.e., social behavior consistency and cross-network temporal dynamics. Social behavior consistency focuses on analyzing whether the same user's social links in different networks have some sort of consistency and whether the formation of a user's typical social links has certain relation with his/her social behaviors.  While cross-network temporal dynamics focuses on analyzing and comparing the reaction speed to the same event in different networks from three different levels, i.e., global level, user level and category level.

 

2. We propose a method to associate the heterogeneous knowledge in different networks by leveraging the collaborative behaviors of the common users in different networks. To represent the heterogeneous knowledge in different networks, we first conduct some typical topic modeling approaches in respective networks. Then we use the common users in different networks to connect the derived network spaces. With the same users' topical representations as a high-level supervision, we finally learn the distribution transfer function for cross-network data association. Based on the derived data association, we further design a cross-network YouTube video promotion application, by leveraging the high propagation efficiency of the Twitter network.

 

3. We propose a unified YouTube video recommendation framework via cross-network user collaboration, to address the typical new user, cold start and data sparsity problems in recommender systems. The proposed method first learns a transfer matrix, based on which users' Twitter social behaviors can be directly used to estimate their video preference on YouTube. To deal with the inconsistency of different user models on Twitter and YouTube, we further design an adaptive fusion strategy to balance their contributions on the final recommendation performance. Experimental results show that the proposed solution achieves superior performance on three kinds of typical users, in term of not only accuracy, but also diversity and novelty.

 

4. We propose a dynami?c association framework to associate different networks, by observing users’ temporal behaviors on them. A cold-start video recommendation solution is further designed based on the derived association. To capture the stable short-term user interest, we split the whole time space into many time sessions with equal length, and conduct the dynamic user modeling within each time session, then the derived user interest representations in each session are further used to compute the cross-network user interest association. Besides, assuming that the association patterns will not change dramatically in successive time sessions, we also incrementally update the derived association with time session by session.