5. Background – Retweeting Behavior
5
• key mechanism for spreading information
• can help information spreading prediction, popularity prediction etc.
6. (Some) Related Work
6
• Retweeting behavior
• study of a number of features for retweetability of tweets [Suh et al., SocialCom’10]
• feature-aware factorization model [Feng et al, WSDM’13]
• considering information about user, tweet, and author
• who will retweet me? [Luo et al., SIGIR’13]
• using learning-to-rank framework
• non-parametric statistical models [Zhang et al. AAAI’15]
• combining structural, textual & temporal info.
7. (Some) Related Work
7
• Retweeting behavior
• study of a number of features for retweetability of tweets [Suh et al., SocialCom’10]
• feature-aware factorization model [Feng et al, WSDM’13]
• considering information about user, tweet, and author
• who will retweet me? [Luo et al., SIGIR’13]
• using learning-to-rank framework
• non-parametric statistical models [Zhang et al. AAAI’15]
• combining structural, textual & temporal info.
feature engineering is required
8. (Some) Related Work
8
• Convolutional Neural Network (CNN)
• image recognition
• video processing
• natural language processing
• Attention-based Neural Network
• machine translation
• speech recognition
• visual object classification
9. Proposed Approach – Variants of CNN approach
9
words of a tweet
• Vu: user embedding vector
• Vp: tweet embedding vector
11. Proposed Approach
11
• Modeling User Interests based on Tweet History [t1, t2 … tm]
• clustering m tweets of each user into n groups using K-means
• using the central tweet of each group as an interest of user
• user interest profile [t1, t2 … tn]
12. • Modeling User Interests based on Tweet History [t1, t2 … tm]
• clustering m tweets of each user into n groups
• using the central tweet of each group as an interest of user
• user interest profile [t1, t2 … tn]
• apply CNN for each tweet to obtain
tweet embeddings
Proposed Approach
12
21. Conclusions
21
• Proposed a novel attention-based deep neural network
• that can perform better than state-of-the-art methods for retweet prediction
• user, author embeddings, the similarity score and the user’s attention interests can
each significantly improve the performance
• the integration of these components provides the best performance