Vol.58 No (May 2017) 1 2,a) , CF CF Matrix Factorization MF MF Collaborative Topic Regression CTR CTR CTR MF Matrix Fa

Size: px

Start display at page:

Download "Vol.58 No (May 2017) 1 2,a) , CF CF Matrix Factorization MF MF Collaborative Topic Regression CTR CTR CTR MF Matrix Fa"

はすなうみのなか
5 years ago
Views:

1 1 2,a) , CF CF Matrix Factorization MF MF Collaborative Topic Regression CTR CTR CTR MF Matrix Factorization LDA Examination on Recommendation Performance of Information Source on Collaborative Filtering Using Latent Topics Akihiro Nishimura 1 Yoshinori Hijikata 2,a) Nobuchika Sakata 1 Received: August 14, 2016, Accepted: February 9, 2017 Abstract: Collaborative filtering is one of the popular methods for selecting contents or items suited to user s interest or preference from a content (item) set. Among the existing methods of CF, Matrix Factorization (MF) is superior for the dataset with much missing data. However, even MF cannot perform well when the number of users ratings are extremely small. To solve this problem, an approach that uses not only users rating values but also text information of items are becoming popular recently. We focus on one major model of this approach, Collaborative Topic Regression (CTR). In CTR, text information of items is used as information source to extract topics. However, the case using text information of users has not been evaluated. In this study, we compare the case using text information of items and the case using text information of users in CTR to know the performance difference. We compare the both cases and MF (which is a baseline) according to accuracy metrics and usefulness metrics. We found that using text information of users realizes the recommendation with high accuracy and using text information of items realizes the recommendation with high usefulness from the experiment. Keywords: recommender system, collaborative filtering, Matrix Factorization, topic model, LDA 1 Graduate School of Engineering Science, Osaka University, Toyonaka, Osaka , Japan 2 School of Business Administration, Kwansei Gakuin University, Nishinomiya, Hyogo , Japan 1. 1 a) contact@soc-research.org c 2017 Information Processing Society of Japan 1102

2 content-based filtering CBF collaborative filtering CF [1] CF *1 CF Matrix Factorization MF [2], [3] MF [4] MF [5], [6] MF [7], [8], [9], [10] MF [7] Collaborative Topic Regression CTR [7] CTR [11] *1 Amazon ( LinkedIn ( Hulu ( 5 CTR ictr uctr MF ( 1 ) CTR (2) (3) 2 3 CTR CF 2 [12] 1 [13], [14] 1 c 2017 Information Processing Society of Japan 1103

3 [2], [15] [16] [4], [17] R U T V R U, V [2] Matrix Factorization MF Blei Latent Dirichlet Allocation LDA [18] CF CF [19] [19] [5], [20], [21] CF [7], [8], [9], [10] collaborative topic regression CTR [7] Wang LDA Blei CF CTR 3 3. CTR 1 Collaborative Topic Regression Fig. 1 Graphical model of Collaborative Topic Regression. 3.1 Collaborative Topic Regression itemoriented CTR [7] 1-(a) CTR ictr α β LDA [18] v λ v MF MF Probabilistic Matrix Factorization PMF [22], [23] CTR LDA θ PMF v θ v i I j J k K w j r ij r ij {0, 1} r ij =0 i j c 2017 Information Processing Society of Japan 1104

4 w j y Y w j n z jn w jn θ j k θ jk =1 θ j z jn β k y β ky =1 k u i K v j K λ u λ v u i v j u i v j α θ j MF PMF [22], [23] (1) u i N(0,λ 1 u I K ) v j N(0,λ 1 v I K ) (1) r ij N(u T i v j,c 1 ij ) I K K N(μ, σ) c ij r ij c ij =0 c ij =1 PMF MAP MF (2) min U,V (r ij u T i v j ) 2 + λ u u i 2 + λ v v j 2 (2) ij 2 3 CTR c ij { a (r ij =1 ) c ij = b (r ij =0 ) (3) a b r i,j a >b>0 CTR 1. i N(0,λ 1 u I K ) u i 2. j a. Dir(α) θ j b. N(0,λ 1 v I K ) K ɛ j v j = ɛ j + θ j c. w jn i. Mult(θ) z jn ii. Mult(β zjn ) w jn 3. (i, j) N(u T i v j,c 1 ij ) r ij CTR u v θ β [7] L (4) L coordinate ascent [24] L = c ij (r ij u T i v j ) 2 i j ( ) + log θ jk β k,wjn j n k λ u u i 2 λ v v j θ j 2 (4) i j u v (5) U V u i v j C i C j Cdd i = c id C j dd = c dj r i r j r i =(r i1,r i2,...,r ij ) r j =(r 1j,r 2j,...,r Ij ) u i L u i v j u i =(VC i V T + λ u I K ) 1 VC i r i (5) v j =(UC j U T + λ v I K ) 1 (UC j r j + λ v θ j ) θ L (4) 2 (4) θ j L(θ j ) φ jnk = q z jn = k Jensen (6) ( ) L(θ j )= λ v v j θ j 2 θ jk β k,wjn φ jnk + log φ jnk n λ v v j θ j 2 (6) + φ jnk (log(θ jk β k,wjn ) log(φ jnk )) n k = L(θ j,φ j ) L(θ j,φ j ) L(θ j ) k c 2017 Information Processing Society of Japan 1105

5 φ jnk φ jnk θ jk β kwjn θ j θ j simplex projection gradient [24] β (7) β kw φ jnk δ(j, n, w) (7) j n { 1(w jn = w ) δ(i, n, w) = 0(w jn w ) u v θ β û ˆv (8) ˆr i r 0 (i, j) ˆr ˆr ij j r ij ˆr ij = u T i v j (8) 3.2 Collaborative Topic Regression useroriented CTR 1(a) u v λ u λ v I J uctr uctr 1(b) uctr ictr uctr K θ u v uctr MAP [25] N Precision Recall [7] [8], [9], [23] Prediction Coverage PC Catalogue Coverage CC [26] 2 [27] Catalogue Coverage CC aggregate diversity [28] inter-user diversity [29] temporal diversity [30] [25] IUD i L i I Recall Recall = T i IL i (9) T i T i i Recall CC CC = i=1,...,i IL i (10) B c 2017 Information Processing Society of Japan 1106

6 B CC IUD IUD = 1 U C 2 U u1 U d u1,u2 (11) u2 d u1,u2 =1 IL u1 IL u2 S S = IL u1 = IL u2 IUD 2 u1 u2 IUD [0, 1] ictr uctr MF [7] CiteULike *2 CiteULike CiteULike 5.1 CiteULike he the this TreeTagger *3 SlothLib *4 5.2 CiteUlike 1 *2 Web Web *3 schmid/tools/ TreeTagger/ *4 Version1/SlothLib/NLP/Filter/StopWord/word/ English.txt 1 Table 1 An example of text information on user. Profile 3 - PhD, ATR specialization in art therapy. - I am a member of the Software Engineering - Productivity tools department at *** - India. - I am generally interested in studies of learning and the design of instruction, especially in the areas of reading comprehension, science, and mathematics. In particular, I am interested in how objects external to the person (physical artifacts, symbolic and graphical representations) affect learning and reasoning. Interest 3 - Cheminformatics, chemometrics, statistics, chemistry, metabolomics, systems biology, semantic web. - Complex networks, Spatial distribution of networks, Telecommunications, Graph theory, Network design - I work on Geno-transcriptomic study of solid tumours, diffusely infiltrating and thus with a complex architecture. My interest are therefore inter and intra tumour heterogeneity and tumour evolution. My biological background is based on Genomics and few Transcriptomics and Proteomics and Protein Networks. I have a technical biological background in : SNP and Expression Microarrays. FISH, PCR, Western Blot. And in Informatics : R, Python, SQL, PHP, Perl. 0 ictr article title abstract uctr profile interest profile interest 3 1 profile interest % nlt= c 2017 Information Processing Society of Japan 1107

7 2 Table 2 Statistical data of our dataset. (item) (user) (%) (item) (user) (%) nlt=2 51,435 6,671 39,202 1,121 4,779,689 15, nlt=4 26,575 6,416 25,027 1,091 2,474,700 14, nlt=6 10,925 5,969 14,354 1,034 1,012,098 14, nlt=8 5,796 5,597 9,657 1, ,410 13, nlt=10 3,534 5,312 7, ,581 12, K K =5, 10, 30, 50, 100, 200 K =50 λ v λ u [7] ictr λ v =1, 10, 100 λ u =0.1, 1, 10 uctr λ v λ u ictr (λ v = 100,λ u =0.1) uctr (λ v =0.1,λ u = 100) [7] a =1 b =0.01 α =1 θ β LDA θ β LDA Collapsed Gibbs Sampling [31] 500 MF [7] CTR λ v =0.1 λ u =0.1 a =1 b =0.01 α =1 θ Perplexity Perplexity ictr Perplexity (12) { j n Perplexity(D) =exp log( k θ } jkβ k,wjn ) j n 1 (12) 3.1 uctr j Perplexity 2 Perplexity nlt=2 Fig. 2 Perplexity (the most sparse when nlt=2). nlt Perplexity 2 2 ictr uctr uctr nlt Perplexity uctr ictr uctr ictr uctr Perplexity Recall 3 ictr uctr MF 5 nlt Recall c 2017 Information Processing Society of Japan 1108

8 Fig. 3 3 Recall nlt=2 Recall in each model and in each sparse ratio (the most sparse when nlt=2). 4 Fig. 4 Catalogue Coverage nlt=2 Catalogue Coverage in each model and in each sparse ratio (the most sparse when nlt=2). CTR MF Recall 100 MF nlt=2 nlt=10 Recall 0.55 ictr Recall 0.2 uctr 0.35 MF ictr uctr MF MF CTR ictr uctr uctr ictr Recall ictr uctr uctr MF Catalogue Coverage CC 4 Recall nlt CC CC ictr 2 MF uctr nlt=2 CC uctr CC CC c 2017 Information Processing Society of Japan 1109

9 5 Fig. 5 Inter-User Diversity nlt=2 Inter-User Diversity in each model and in each sparse ratio (the most sparse when nlt=2). MF Inter-User Diversity IUD 5 ictr IUD 3 uctr ictr MF 25 IUD MF ictr MF nlt=2 MF nlt=4 MF nlt=2 MF nlt=4 25 IUD MF 1 MF IUD ictr uctr 5.4 MF uctr MF uctr ictr MF MF ictr ictr IUD 6. Collaborative Topic Regression CTR Matrix Factorization MF ictr uctr MF CTR MF c 2017 Information Processing Society of Japan 1110

10 ictr uctr ictr uctr MF ictr uctr K λ u λ v CiteULike 15K12150 [1] Vol.48, No.9, pp (2007). [2] Koren, Y., Bell, R. and Volinsky, C.: Matrix Factorization Techniques for Recommender Systems, Computer, Vol.42, No.8, pp (2009). [3] Koren, Y.: Factorization Meets the Neighborhood: A Multifaceted Collaborative Filtering Model, Proc. KDD, pp (2008). [4] Paterek, A.: Improving Regularized Singular Value Decomposition for Collaborative Filtering, Proc. ACM SIGKDD Cup and mworkshop (2007). [5] Schein, A.I., Popescul, A., Ungar, L.H., Pennock, D.M.: Methods and Metrics for Cold-Start Recommendations, Proc. ACM SIGIR, pp (2002). [6] Chang, S., Harper, M.F. and Terveen, L.: Using Groups of Items for Preference Elicitation in Recommender Systems, Proc. CSCW, pp (2015). [7] Wang, C. and Blei, D.M.: Collaborative Topic Modeling for Recommending Scientific Articles, Proc. ACM SIGKDD, pp (2011). [8] Purushotham, S., Liu, Y. and Kuo, C.-C.J.: Collaborative Topic Regression with Social Matrix Factorization for Recommendation Systems, Proc. ICML, pp (2012). [9] Ding, X., Jin, X., Li, Y. and Li, L.: Celebrity Recommendation with Collaborative Social Topic Regression, Proc. IJCAI, pp (2013). [10] Kim, Y. and Shim, K.: TWILITE: A Recommendation System for Twitter using a Probabilistic Model based on Latent Dirichlet Allocation, Journal of Information Systems, Vol.42, pp (2014). [11] Twitter Vol.56, No.3, pp (2015). [12] Su, X. and Khoshgoftaar, T.M.: A Survey of Collaborative Filtering Techniques, Journal of Advances in Artificial Intelligence, Vol.2009, No.4 (2009). [13] Resnick, P., Iacovou, N., Suchak, M., Bergstorm, P. and Riedl, J.: GroupLens: An Open Architecture for Collaborative Filtering of Netnews, Proc. ACM CSCW, pp (1994). [14] Sarwar, B., Karypis, G., Konstan, J. and Riedl, J.: Item-based Collaborative Filtering Recommendation Algorithms, Proc. ACM WWW, pp (2001). [15] Canny, J.: Collaborative Filtering with Privacy via Factor Analysis, Proc. ACM SIGIR, pp (2002). [16] Hofmann, T. and Puzicha, J.: Latent Class Models for Collaborative Filtering, Proc. IJCAI, pp (1999). [17] Cremonesi, P., Koren, Y. and Turrin, R.: Performance of Recommender Algorithms on Top-N Recommendation Tasks, Proc. ACM RecSys, pp (2010). [18] Blei, D.M., Ng, A.Y. and Jordan, M.I.: Latent Dirichlet Allocation, The Journal of Machine Learning Research, Vol.3, pp (2003). [19] Burke, R.: Hybrid Recommender Systems: Survey and Experiments, User Modeling and User-Adapted Interaction, Vol.12, No.4, pp (2002). [20] Popescul, A., Ungar, L.H., Pennock, D.M. and Lawrence, S.: Probabilistic Models for Unified Collaborative and Content-based Recommendation in Sparse-data Environments, Proc. UAI, pp (2001). [21] Kim, B.M. and Li., Q.: Probabilistic Model Estimation for Collaborative Filtering based on Items Attributes, Proc. IEEE/WIC/ACM WI, pp (2004). [22] Salakhutdinov, R. and Mnih, A.: Probabilistic Matrix Factorization, Proc. NIPS 21 (2008). [23] Hu, Y., Koren, Y. and Volinsky, C.: Collaborative Filtering for Implicit Feedback Datasets, Proc. IEEE ICDM, pp (2008). [24] Bertsekas, D.P.: Nonlinear Programming, 2nd Edition, Athena Scientific, Belmont, Massachusetts (1999). [25] Vol.29, No.6, pp (2014). [26] Ge, M., Delgado-Battenfeld, C. and Jannach, D.: Beyond Accuracy: Evaluating Recommender Systems by Coverage and Serendipity, Proc. ACM RecSys, pp (2010). [27] Herlocker, J.L., Konstan, J.A., Terveen, L.G. and Riedl, J.T.: Evaluating Collaborative Filtering Recommender Systems, ACM Trans. Information Systems (TOIS), Vol.22, No.1, pp.5 53 (2004). [28] Adomavicius, G. and Kwon, Y.: Improving Aggregate Recommendation Diversity Using Ranking-Based Techniques, IEEE Trans. Knowledge and Data Engineering, Vol.24, No.5, pp (2012). [29] Zhou, T., Kuscsik, Z., Liu, J.-G., Medo, M., Wakeling, J.R. and Zhang, Y.-C.: Solving the Apparent Diversityaccuracy Dilemma of Recommender Systems, Proc. National Academy of Sciences, Vol.107, No.10, pp (2010). [30] Lathia, N., Hailes, S., Capra, L. and Amatriain, X.: Temporal Diversity in Recommender Systems, Proc. ACM SIGIR, pp (2010). [31] Griffiths, T.L. and Steyvers, M.: Finding Scientific Topics, Proc. National Academy of Sciences, Vol.101, pp (2004). c 2017 Information Processing Society of Japan 1111

Best Paper Award DEWS2006 2011 WebDB 2011 2012

2013 HITLAB NZ University of Canterbury 2007 4

11 GroupLens Research ACM IUI Best Paper Award DEWS WebDB WebDB WebDB HITLAB NZ University of Canterbury NTS AR c 2017 Information Processing Society of Japan 1112

IPSJ-TOM

IPSJ-TOM Vol. 2 No. 2 47 57 (Mar. 2009) 1, 2 1 3 1 Web Performance Evaluation of Recommendation Algorithms Based on Rating-recommendation Interaction Akihiro Yamashita, 1, 2 Hidenori Kawamura, 1 Hiroyuki Iizuka