School of Electronics Engineering and Computer Science Peking University Beijing, P.R. China Ziqi Wang, Yuwei Tan, Ming Zhang
Content-based recommendation ◦ Recommends resources based on their content and not on user’s rating and opinion. Collaborative filtering ◦ It’s based on the assumption that similar users express similar interests on similar resources. Graph based recommendation ◦ User transitive associations between users and resources in the bipartite user-resource graph.
a = in every step there is a probability q = is a column vector of zeros with the element corresponding to the starting node set to 1 S = is the transition probability matrix and its element P (t) = denotes the probability that the random walk at step t How closely related are two nodes in a graph ?
Node 4 Node 1 Node 2 Node 3 Node 4 Node 5 Node 6 Node 7 Node 8 Node 9 Node 10 Node 11 Node Ranking vector More red, more relevant Nearby nodes, higher scores
QQ … SeaSunSkyWave {} {} CatForestGrassTiger {?, ?, ?,} ? A: RWR! [Pan KDD2004]
Test Image SeaSunSkyWaveCatForestTigerGrass Image Keyword Region
Test Image SeaSunSkyWaveCatForestTigerGrass Image Keyword Region {Grass, Forest, Cat, Tiger}
Treating tagging behavior directly as another form of rating. ◦ Assigning the minimum value of user rating to be the weight of each new edge ◦ Assigning the maximum value of user rating to be the weight of each new edge ◦ Assigning the average rating of the corresponding user to be the weight of the new edge Choose the best one in the experiment.
t i (k) = is the k th tag made by user u i. c i (k) = the frequency of tag t i (k) to describe the interest of user
Measuring the user’s similarity based on their tagging information. n i = is the number of tags user u i assigned. c i (k) = the frequency of tag t i (k)
The weight of the edge should be proportional to the similarity. k = is a parameter that we will test it in the experiment.
TS = stands for test set U = stands for users set RelevantNum = the number of relevant resources in the results RecommendLength = the number of resources that are recommended to a user
= Precision at rank K ◦ The proportion of resources ranked in the top K results. = Success at rank K ◦ The probability of finding a good resource among the top K results.
Method 1 = Assigning the minimum value Method 2 = Assigning the maximum value Method 3 = Assigning the average rating
Conclusions ◦ Two algorithms based on the framework of Random Walk with Restarts. ◦ This proves that our promotion algorithm performs better on sparse data sets. Future work ◦ Focus on recommendation on large scale data with better performance and lower time cost.