1. MODEL ADAPTATION FOR PERSONALIZED OPINION ANALYSIS MOHAMMAD AL BONI KEIRA ZHOU

2. OUTLINE Introduction Related Work Model Adaptation Framework Experiment and Discussion Conclusion and Future Work 1

3. SENTIMENT ANALYSIS Exploring user’s sentiment: $8 for A bag of chips Master student Phd student Professor Ratings: 1 This is ridiculously expensive. A bag of chips for $8? I’d rather eat potatoes. Ratings: 3 The chip is expensive. But the taste is pretty good. Ratings: 5 This bag of chips worths the money! Some people may think it expensive but the taste is really good! 2

4. Exploring user’s sentiment: $8 for A bag of chips Master student Phd student Professor Ratings: 1 This is ridiculously expensive. A bag of chips for $8? I’d rather eat potatoes. Ratings: 3 The chip is expensive. But the taste is pretty good. Ratings: 5 This bag of chips worths the money! Some people may think it expensive but the taste is really good! SENTIMENT ANALYSIS 3

5. THE BIG QUESTION How to do personalization for sentiment analysis? 4

6. RELATED WORK Sentiment Analysis: Classify text documents into predefined sentiment classes, e.g., positive v.s. negative, (Dave et al., 2003; Kim and Hovy, 2004) Identify topical aspects and corresponding opinions (Wang et al., 2010; Jo and Oh, 2011) Opinion summarization (Hu and Liu, 2004; Ku et al., 2006) Transfer Learning: Aims to help improve the learning of target predictive problem by using the knowledge from different but related problems (Pan and Yang, 2010) In opinion mining community, transfer learning is mostly exploited for domain adaptation Blitzer et al. (2006) proposed structural correspondence learning to identify the correspondences among features between different domains via the concept of pivot features. 5

7. ANSWER TO THE BIG QUESTION How to do personalization for sentiment analysis? Inspired by a personalized ranking model adaptation method developed in (Wang et al., 2013) Use transfer learning to adapting a generic sentiment classification model for each user 6

8. Adapting the global model for each individual user PERSONALIZED MODEL ADAPTATION Master student Phd student Professor Ratings: 1 This is ridiculously expensive. A bag of chips for $8? I’d rather eat potatoes. Ratings: 3 The chip is expensive. But the taste is pretty good. Ratings: 5 This bag of chips worth the money! Some people may think it expensive but the taste is really good! 7

9. Adjusting the generic model’s parameters with respect to each individual user’s review data PERSONALIZED MODEL ADAPTATION Shifting 8

10. Adjusting the generic model’s parameters with respect to each individual user’s review data PERSONALIZED MODEL ADAPTATION Scaling 9

11. Adjusting the generic model’s parameters with respect to each individual user’s review data PERSONALIZED MODEL ADAPTATION Rotation 10 grouping

12. Adapting global model for each individual user Loss function from any linear classifier - logistic regression in our case Complexity of adaptation: prefer no adaptation PERSONALIZED MODEL ADAPTATION 11

13. EXPERIMENTS AND DISCUSSION Dataset and Preprocessing Amazon review data from Stanford SNAP website Remove users who have more than 1000 reviews Ratings greater than 4 stars are labeled as Positive; others are Negative Unigram and Bigram Bag-of-words feature representation Chi-square and information gain for feature selection 12

14. EXPERIMENTS AND DISCUSSION (CONT’D) Baseline Model First baseline model Instance-based adaptation K-nearest neighbors of test set from training set Second baseline model Individual logistic regression models for every user with regularization term Force the personalized model to be close to the global model 13

15. EXPERIMENTS 14

16. EXPERIMENTS 2. Different training size for global: # of users (global): 100, 200, 300, 400 & 500. # of users (Personalized): 100. 15

17. EXPERIMENTS 3. Different transformation operations: Operation: Shifting, Shifting + Scaling, Shifting + Scaling + Rotation. Global model: 100 users, 5050 features. Personalized model: Users GroupTotalPosNeg Light [2,10]: 204368740868697 (78.6%)18711 (21.4%) Medium [11,50]: 12458273435211234 (77.25%)62201 (22.75%) Heavy [50,100]: 3624921882 (75.52%)610 (24.48%) 16

18. 3. Different transformation operations: Heavy Users: EXPERIMENTS Users GroupTotalPosNeg Light [2,10]: 204368740868697 (78.6%)18711 (21.4%) Medium [11,50]: 12458273435211234 (77.25%)62201 (22.75%) Heavy [50,100]: 3624921882 (75.52%)610 (24.48%) 17

19. 3. Different transformation operations: Heavy Users: 18 EXPERIMENTS Users GroupTotalPosNeg Light [2,10]: 204368740868697 (78.6%)18711 (21.4%) Medium [11,50]: 12458273435211234 (77.25%)62201 (22.75%) Heavy [50,100]: 3624921882 (75.52%)610 (24.48%)

20. 3. Different transformation operations: Medium Users: 19 EXPERIMENTS Users GroupTotalPosNeg Light [2,10]: 204368740868697 (78.6%)18711 (21.4%) Medium [11,50]: 12458273435211234 (77.25%)62201 (22.75%) Heavy [50,100]: 3624921882 (75.52%)610 (24.48%)

21. 3. Different transformation operations: Light Users: 20 EXPERIMENTS Users GroupTotalPosNeg Light [2,10]: 204368740868697 (78.6%)18711 (21.4%) Medium [11,50]: 12458273435211234 (77.25%)62201 (22.75%) Heavy [50,100]: 3624921882 (75.52%)610 (24.48%)

22. 4. Performance gain over global model: MethodUser ClassPos F1Neg F1 ReTrain Heavy 0.389-0.063 Medium 0.3890.054 Light 0.0890.076 Reg-LR Heavy 0.533-0.044 Medium 0.509-0.015 Light 0.2120.017 Shifting+Scalin g Heavy 0.593*-0.083 Medium 0.693*-0.022 Light 0.500*-0.023 Shifting+Scalin g +Rotation Heavy 0.543-0.102 Medium 0.574-0.038 Light 0.241-0.018 * p-value <0.05 with paired t-test. 21 EXPERIMENTS Users GroupTotalPosNeg Light [2,10]: 204368740868697 (78.6%)18711 (21.4%) Medium [11,50]: 12458273435211234 (77.25%)62201 (22.75%) Heavy [50,100]: 3624921882 (75.52%)610 (24.48%)

23. EXPERIMENTS 5. Highest polarity variance: Absolute difference between feature coefficients. Variance across all users. Top 10 features with highest polarity variance: great, love, best, bad, like, read, time, excellent, high, and work. great, love, best, like: Same in global, different between users. read, time, work: Neutral in global, varies between users. 22

24. CONCLUSION AND FUTURE WORK Developed a transfer learning based solution for personalized opinion mining Linear transformations, i.e. scaling, shifting and Rotation Empirical evaluations verify that personalized adaptation improves sentiment classification Future work Further explore this linear transformation based adaptation from different perspectives, e.g., sharing adaptation operations across users or review categories. Submitted to Associations of Computational Linguistics (ACL)! 23

25. REFERENCES Kushal Dave, Steve Lawrence, and David M Pennock. 2003. Mining the peanut gallery: Opinion extraction and semantic classification of product reviews. In Proceedings of the 12th international conference on World Wide Web, pages 519–528. ACM. Soo-Min Kim and Eduard Hovy. 2004. Determining the sentiment of opinions. In Proceedings of the 20th international conference on Computational Linguistics, page 1367. Association for Computational Linguistics. Hongning Wang, Yue Lu, and Chengxiang Zhai. 2010. Latent aspect rating analysis on review text data: a rating regression approach. In Proceedings of the 16th ACM SIGKDD international conference on Knowledge discovery and data mining, pages 783–792. ACM. Yohan Jo and Alice H Oh. 2011. Aspect and sentiment unification model for online review analysis. In Proceedings of the fourth ACM international conference on Web search and data mining, pages 815–824. ACM. Minqing Hu and Bing Liu. 2004. Mining and summarizing customer reviews. In Proceedings of the tenth ACM SIGKDD international conference on Knowledge discovery and data mining, pages 168–177. ACM. Lun-Wei Ku, Yu-Ting Liang, and Hsin-Hsi Chen. 2006. Opinion extraction, summarization and tracking in news and blog corpora. In AAAI spring symposium: Computational approaches to analyzing weblogs, volume 100107. John Blitzer, Ryan McDonald, and Fernando Pereira. 2006. Domain adaptation with structural correspondence learning. In Proceedings of the 2006 conference on empirical methods in natural language processing, pages 120–128. Association for Computational Linguistics. Hongning Wang, Xiaodong He, Ming-Wei Chang, Yang Song, Ryen W White, and Wei Chu. 2013. Personalized ranking model adaptation for web search. In Proceedings of the 36th international ACM SIGIR conference on Research and development in information retrieval, pages 323–332. ACM. Sinno Jialin Pan and Qiang Yang. 2010. A survey on transfer learning. Knowledge and Data Engineering, IEEE Transactions on, 22(10):1345–1359. 24

26. THANKS! 25