PRemiSE: Personalized News Recommendation via Implicit Social Experts
Feb 24, 2016
PRemiSE:Personalized News Recommendation via Implicit Social Experts
1. Introduction2. Expert model3. PRemiSE4. Experimental5. Future work
Overview
Google News
Existing news recommender systems
Content-based Recommenders bag-of-word model : document word topic models : document topic word Collaborative Filtering KNN MF PMF Hybrid Recommenders combining social network
1. data sparsity2. cold-start problem
PRemiSE: incorporating content information, collaborative filtering and information diffusion in virtual social network into probabilistic matrix factorization.
Two problems in previous studies
1. Capable of handling the cold-start problem2. Semantically interpretable3. Producing better predictions
Our contribution
Expert Model
An illustrative example of implicit social network
Building Implicit Social Network
Step0: Compute time span & number of visits for each itemStep1: Plot the time span , number of visits , find the abnormal items , remove itStep2: Build the graph, based on user-item accessing historyif U1 access the same item V after U2, and access_time(U1) – access_time(U2) < time_window , we say in the graph , there is an directed edges from U2 to U1.Step3 : Normalized weights
Time_window:1. find enough neighbors for each user2. precisely find the right experts
Empirical study on a real data set
1. How probably the given user will follow the expert’s adoption on the same item?
2. How probably any individual will follow the expert’s adoption on the same item?
Local Expert and Global Expert
find global expert?
1. Matrix factorization2. Probabilistic matrix factorization3. PRemiSE4. Learning in PRemiSE5. Inference in PRemiSE
PRemiSE
Matrix factorization
user u
𝑝𝑢
=
𝑞𝑖
item i
Now how to get
Gradient descent
Probabilistic matrix factorization
Linear probability model
PRemiSE
Learning in PRemiSE
See detailed in the paper
Optimization Algorithm
Existing Item by Existing User
Existing Item by New User
New Item by Existing User
New Item by New User
Inference in PRemiSE
Real-World Dataset 1. crawled from several popular news service websites 2. two types of elements : news stories and named entities.
Rating 1. Rating in Story: binary 2. Rating in Entity: numerical
Experimental Evaluation
Step1: eliminate outlier items employing by ELKIStep2: The size of time-window set to be 8 days. we delete edges that are caused by a delayed co-consumption (9 days or even longer)step 3, we normalize the edges weight, and empirically set the edge weight threshold as 0.001
Construction of Networks
Parameters of Global Expert Model
Comparative Study
Cold-start problem
Semantics of factors
We integrate this “expert” model with the content information and collaborative filtering, and propose a hybrid recommendation framework, called PRemiSE.
1. effectively handle the cold-start problem2. better Semantics Explanation3. better performance in recommendation accuracy
FUTURE WORK : social media & information diffusion model & export model
Conclusion AND Future work
Thank you for your time!