Download presentation
Presentation is loading. Please wait.
1
华东师范大学软件学院 王科强 (第一作者), 王晓玲
Bayesian Probabilistic Multi-Topic Matrix Factorization for Rating Prediction 华东师范大学软件学院 王科强 (第一作者), 王晓玲 中国人民大学信息学院 赵鑫 Wayne Xin Zhao
2
推荐系统 User-item matrix m1 m2 u1 1 ? u2 5 u3 3 u4 2 Rating prediction
3
Rating Prediction User-based KNN Item-based KNN Matrix factorization
Feature-based MF (Context-based MF) SVD++ SVDFeature LibFM
4
推荐系统 User-item matrix m1 m2 u1 yes ? u2 u3 no u4 Item recommendation
5
Item recommendation User-based KNN Item-based KNN Matrix factorization
Pairwise-ranking recommendation
6
Our Focus Rating Prediction Matrix Factorization
7
最新的研究进展之一 稀疏性是困扰矩阵分解的一个主要问题之一 大部分评分矩阵只有少数ratings(相对比例)
8
最新的研究进展之一 Local Matrix Factorization (LMF) 缺乏一个较为形式化的刻画; 可解释性较差;
9
吸取前人工作的精华 Local Matrix Factorization
Each submatrix can be considered as a local cluster of users and items; A user or an item has multiple latent representations in different submatrices
10
Our Approach Cluster Topic
Group items by topics Use topic models Cluster-specific latent factor Topic-specific latent factor Each user or item has K topic-specific latent factors
11
Modeling the Rated Items
For each rating record <u,m>, we treat a user u as a document and an item m as a word token, then build the topic models
12
Modeling the Ratings Now we model the ratings
13
The generative process
14
Learning Algorithm
15
Connections with previous studies
16
Experiments Datasets Metric Baselines
17
Results
18
Quality of “clusters”
19
Summary Topics as “local clusters” Topic-specific latent factors
A bayesian approach
20
The End 谢谢!
Similar presentations