Biased Random Walk based Social Regularization for Word Embeddings Ziqian Zeng*, Xin Liu*, and Yangqiu Song The Hong Kong University of Science and Technology * Equal Contribution
Outline Socialized Word Embeddings Motivation (Problems of SWE and Solutions) Algorithm Random Walk Methods Experiments
Socialized Word Embeddings (SWE) Everyone has his/her own personal characteristics of language use. Zeng et al., Socialized Word Embeddings, IJCAI, 2017
Socialized Word Embeddings (SWE) Socially connected individuals tend to use language in similar ways.1 Frequently Used Words Language feature pawsome, anipals, furever animal based punctuation kradam, glambert, glamily puns around pop star Adam Lambert tdd, mvc, linq acronyms [1] Yi Yang et al. Overcoming Language Variation in Sentiment Analysis with Social Attention, 2016. Table is from Bryden John et al, Word usage mirrors community structure in the online social network Twitter, 2013 with partial deletion.
Socialized Word Embeddings (SWE) Figure 1. Illustration of Socialized Word Embeddings1. Vector representation of a word for user : 1 Zeng et al., Socialized Word Embeddings, IJCAI, 2017
Socialized Word Embeddings (SWE) Personalized CBOW (1) Socialized Regularization (2) Socialized Word Embeddings Trade-off Parameter (3) Constraints
Problems of SWE Implicit modeling on transitivity Not affected node Node being considered Node being affected
Problems of SWE Implicit modeling on transitivity Not affected node Node being considered Node being affected
Problems of SWE Implicit modeling on transitivity Not affected node Node being considered Node being affected
Problems of SWE Implicit modeling on transitivity Not affected node Node being considered Node being affected
Problems of SWE The spread of language use is transitive. Not affected node Node being considered Node being affected
Solution Use random walk methods to generate paths Not affected node Node being considered Node being affected
Problems of SWE Lack of concern of users who have fewer friends Case 2 More Friends Not affected node Node being considered Node being affected
Solution Go to users who have fewer friends with larger probability. B A C
Solution Go to users who have fewer friends with larger probability. Node # friends B 1 C 4 B A C
Solution Go to users who have fewer friends with larger probability. B Node # friends B 1 C 4 A C
Social Regularization with Random Walk Algorithm Under SWE framework. Figure 2. Illustration of Socialized Word Embeddings. Zeng et al., Socialized Word Embeddings, IJCAI, 2017
Social Regularization with Random Walk Algorithm Under SWE framework. Generate paths using random walk methods.
Social Regularization with Random Walk Algorithm Under SWE framework. Generate paths using random walk methods. Each user in paths gets update.
Random Walk 1/3 Transition Probability A B C 1/3 2/3 1 1 A B 2/3 1 C
Random Walk 1/3 1 A B 2/3 1 C Path: A - Transition Probability A B C 1/3 2/3 1 1 A B 2/3 1 C Path: A -
Random Walk 1/3 1 A B 2/3 1 C Path: A - C Transition Probability A B C 1/3 2/3 1 1 A B 2/3 1 C Path: A - C
Random Walk 1/3 1 A B 2/3 1 C Path: A - C - B Transition Probability A 1/3 2/3 1 1 A B 2/3 1 C Path: A - C - B
Random Walk 1/3 1 A B 2/3 1 C Path: A - C - B - A Transition Probability A B C 1/3 2/3 1 1 A B 2/3 1 C Path: A - C - B - A
First-order Random Walk A random walker moves to next node based on the last node. Second Last Node Last Node Node Perozzi et al., Deepwalk: Online learning of social representations, KDD, 2014.
Second-order Random Walk A random walker moves to next node based on the last two nodes. Second Last Node Last Node Node B A Grover et al., node2vec: Scalable feature learning for networks, KDD, 2016.
Second-order Random Walk A random walker moves to next node based on the last two nodes. Second Last Node Last Node Node B A Grover et al., node2vec: Scalable feature learning for networks, KDD, 2016.
Second-order Random Walk A random walker moves to next node based on the last two nodes. Second Last Node Last Node Node B A Grover et al., node2vec: Scalable feature learning for networks, KDD, 2016.
Biased Second-order Random Walk A random walker moves to next node based on the last two nodes.
Biased Second-order Random Walk Number of friends of node Average number of friends Hyper-parameter
Experiments Dataset – Yelp Challenge Yelp Challenge Datasets contain billions of reviews, ratings, and large social networks. Table 1. Statistics of Yelp Round 9 and 10 datasets.
Sentiment Classification Two phenomena of language use help sentiment analysis task. Word such as “good” can indicate different sentiment ratings depending on the author. [1] Yang et al. Overcoming Language Variation in Sentiment Analysis with Social Attention, TACL, 2017.
Sentiment Classification average review + user Logistic Regression classifier rating Figure 2. Illustration of the procedure of sentiment classification
Sentiment Classification – Head & Tail Train on only head users or tail users. Head users published more reviews. Tail users published fewer. Half of total reviews from head users. Half from tail users. Tail Data Head Data
Sentiment Classification – Head & Tail Train on only head users or tail users. Head users published more reviews. Tail users published fewer. Half of total reviews from head users. Half from tail users. Statistics Table 2. Statistics of head users and tail users in the one-fifth of the training set.
Experiments A C U R Y Figure 3. Classification Accuracy on Head & Tail Users on Yelp Round 9
Experiments A C U R Y Figure 4. Classification Accuracy on Head & Tail Users on Yelp Round 10
Thank You Conclusion Explicitly model the transitivity Sample more users who have fewer friends. Thank You