TAPER: A Contextual Tensor- Based Approach for Personalized Expert Recommendation Hancheng Ge, James Caverlee and Haokai Lu Department of Computer Science and Engineering Texas A&M University, USA ACM RecSys 16:: September 18th, 2016
Recommender Systems Items (Movies, Songs, News, etc.) High-quality Content Producers
OUR GOAL: Recommend these experts to the right people High-Quality Content Producers (Experts) Personalized Expert Recommendation
Politics Technology Entertainment 1? 1? 1? 1 1? 1?? 1 1??? 1 1 1 1 1??? 1 1 Users?????? 1 1 1????????? 1?? 1?????????? 1???????????????? 1?????????????? 1???????? 1 1? 1??????????? 1 1 1 1??????? 1 1?????????????????????????????????? Topics Experts User-Expert-Topic preferences can be represented as a Tensor
Basic Tensor-based Personalized Expert Recommendation Matrix factorization CANNOT simultaneously consider all dimensions. Input: observed tensor, indicator tensor. Output: complete tensor X, latent matrices U (1),U (2),U (3). Minimize U (n),x Experts T 1 3X 2 kx [U (1),U (2),U (3) ]k 2 F + 2 subject to X = T,U (n) 0,n=1, 2, 3 n=1 ku (n) k 2 F, U 1 (3) U 2 (3) U R (3) U 1 (2) U 2 (2) U R (2) Users U 1 (1) U 2 (1) U R (1)
Challenges 1. Personal Experts 2. Sparsity 3. Complex Relationships David Jane Amy Ryan
Idea: Use Contextual Information Social Activities Temporal Q1: How to model? Q2: How to integrate? Location Topics Q3: Which is more important?
But First: Geo-tagged Twitter Lists A curated group of Twitter accounts. Allowing a user to label another user with an annotation (e.g., tech). List creators as users and members in the lists as experts Technology Foodie @HanchengGe @HanchengGe Allen @elonmusk Jane @jamieoliver Tech Food
Geo-Spatial Context 1 0.9 0.8 Ryan CDF 0.7 0.6 0.5 0.4 0.3 0.2 0.1 Atlanta Chicago Dallas Denver Houston Seattle San Francisco Washington DC 0 0 2000 4000 6000 8000 10000 12000 Distance CDF of the average distance bet. users and experts by locations Jane David The geo-spatial context does affect the preference for experts with varying degrees based on topics and locations.
Topic Context 0.8 Entertainment Food Ryan Cosine Similarity 0.6 0.4 0.2 0 0 1 2 3 4 5 6 7 8 9 Number of Shared Topics # Shared Topics VS. Similarity between Users Technology Technology Entertainment Entertainment Politics Jane David Politics The topic context does affect the preference for experts.
Social Context 1 0.8 Ryan CDF 0.6 0.4 Social Ties 0.2 users who follow the other users who do not follow 0 0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 Cosine Similarity between Users Jane David The social context does affect the preference for experts.
Modeling Contextual Preferences Bet. Homogeneous Entities Similarity Matrix S Bet. Heterogeneous Entities Users Adjacency Matrix A Adjacency Matrix AB Experts Topics Adjacency Matrix AC Similarity Matrix S Similarity Matrix S
Contextual Tensor-based Approach for Personalized Expert Recommendation (TAPER) Bet. Homo. Users Bet. Hetero. Bet. Hetero. Experts Bet. Hetero. Bet. Homo. minimize U (n),xx Topics Bet. Homo. Basic Tensor-based Recommendation n=1 Contextual Information bet. Homogeneous Entities 1 2 [[U (1),U (2),U (3) ]]k 2 X 3X F + ku (n) 2 X3X kx k k k 2 F + tr(z (n)t L n Z (n) ) 2 n=1 + 2 (ka U (1) U (2)T k 2 F + kb U (1) U (3)T k 2 F + kc U (2) U (3)T k 2 F ), Contextual Information bet. Heterogeneous Entities subject to X = T,U (n) = Z (n) 0,n=1, 2, 3 : control the weight of contextual information bet. homogeneous entities. : control the weight of contextual information bet. heterogeneous entities. Alternating Direction Method of Multipliers (ADMM) is applied. Recommendation are conducted based on the estimated tensor X.
State-of-the-art Methods Most Popular (MP) User-based Collaborative Filtering (UCF) Matrix Factorization (MF) Tensor Factorization (TF) Variants of Proposed TAPER Geo-based TAPER (G-TAPER) Topical-based TAPER (T-TAPER) Social-based TAPER (S-TAPER) Contextual Personalized Expert Recommendation (TAPER)
Experiments: Recommendation Effectiveness 0.35 0.3 TF > MF Precision@k 0.25 0.2 0.15 0.1 TF > MF TF > MF TAPER TF MF G TAPER T TAPER S TAPER UCF MP 0.05 0 Top 5 Top 10 Top 15 Tensor Factorization (TF) has a better performance than Matrix Factorization (MF)
Experiments: Recommendation Effectiveness 0.35 0.3 Precision@k 0.25 0.2 0.15 0.1 TAPER TF MF G TAPER T TAPER S TAPER UCF MP 0.05 0 Top 5 Top 10 Top 15 TAPER has the best performance comparing with other state-of-the-art methods
Experiments: Recommendation Effectiveness 0.35 0.3 Precision@k 0.25 0.2 0.15 0.1 TAPER TF MF G TAPER T TAPER S TAPER UCF MP 0.05 0 Top 5 Top 10 Top 15 Social ties of users and experts provide more significant contributions to the personalized expert recommendation
Experiments: Impact of Contextual Preferences 0.26 0.255 Bet. Heterogeneous Entities Bet. Homogeneous Entities 0.25 Precision@10 0.245 0.24 0.235 0.23 0.225 G TAPER T TAPER S TAPER TAPER Contextual preferences between homogeneous entities play a more important role than ones between heterogeneous entities
Experiments: Consistency Precision@10 0.3 0.25 0.2 0.15 0.1 TAPER TF MF 0.05 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Fraction of Training Data TAPER consistently outperforms both Matrix Factorization (MF) and Tensor Factorization (TF) in precision@10
Conclusions and Future Work The user of contextual information can lead to improve the accuracy of personalized expert recommendation based on a tensor-based approach. Social ties of users and experts provide more significant contributions to the personal expert recommendation than the geospatial and topical. Contextual preferences between homogeneous entities play a more important role than ones between heterogeneous entities. Future Work: Integrate additional contextual signals (e.g., temporal factors). Distributed TAPER for large-scale data.
Thank you! Q & A
Modeling Geo-Spatial Preferences Bet. Homogeneous Entities Bet. Heterogeneous Entities u i e j l(u i ) l(u j ) e i user i expert j location of user i num. of users selecting expert i in the location of user j across topics H G (u i,u j )=exp( Users Dist(u i,u j ) 2 2 2 ) A G B G ti adjacency matrix of the spatial popularity of an expert in the location of a user. adjacency matrix of the spatial popularity of a topic in the location of a user. distribution of distances bet. users and experts in a topic i. F AG = ka G U (1) U (2)T k 2 F F BG = kb G U (1) U (3)T k 2 F Experts Topics V G (e i,e j )=exp( Dist(l(e i ),l(e j )) 2 2 2 X u i 2U ( l(u i) e i l(u i ) e j ) 2 ) W G (t i,t j )=1 D KL ( ti k tj )
Modeling Topical Preferences Bet. Homogeneous Entities Bet. Heterogeneous Entities H T (u i,u j )= T T u i Tuj S T ui Tuj exp(x (o t u i o t u j )P t ) t2t T ui o t u i P t set of topics of a user is interested in num. of experts a user labeled in a topic t probability of being interested in a topic t Users T ei t e i set of topics of an expert has expertise num. of times an expert labeled by users in a topic t F BT = kb T U (1) U (3)T k 2 F Experts F CT = kc T U (2) U (3)T k 2 F Topics V T (e i,e j )= T T e i Tej S T ei Tej exp(x ( t t e i ej )P t ) t2t B T C T affinity matrix where an element indicates if a user is interested in certain topic affinity matrix where an element indicates the num. of times that an expert has been recognized by users in a topic
Modeling Social Preferences Bet. Homogeneous Entities Bet. Heterogeneous Entities H S (u i,u j )= F u i T Fuj F ui S Fuj Users F ei F ui set of users an expert follows set of users a user follows F AS = ka S U (1) U (2)T k 2 F A S adjacency matrix where the element indicates if a user follows an expert Experts Topics V S (e i,e j )= F e i T Fej F ei S Fej