Presentation is loading. Please wait.

Presentation is loading. Please wait.

Supervised Random Walks: Predicting and Recommending Links in Social Networks Lars Backstrom (Facebook) & Jure Leskovec (Stanford) Proc. of WSDM 2011 Present.

Similar presentations


Presentation on theme: "Supervised Random Walks: Predicting and Recommending Links in Social Networks Lars Backstrom (Facebook) & Jure Leskovec (Stanford) Proc. of WSDM 2011 Present."— Presentation transcript:

1 Supervised Random Walks: Predicting and Recommending Links in Social Networks Lars Backstrom (Facebook) & Jure Leskovec (Stanford) Proc. of WSDM 2011 Present by Mo Mingzhen

2 Problem Friendship is important on social networks How to predict the future interaction How to recommend potential friends to new user? Link Prediction Problem

3 Motivation Predicting future interaction brings direct business consequences: possible collaborations Beyond social networks: predicting coauthor/collaborations In link prediction problem, how to combine the node and edge attributes remains an open challenge

4 Method Based on the Supervised Random Walks – Combines the network structure with the characteristics of nodes and edges Develop an algorithm to estimate the edge strength – bias a PageRank-like random walk to visits given nodes more often

5 Problem Formulation Given G(V, E) A start point s, learning candidate C = {c i } Destination nodes D = {d 1,…,d k }, no-link nodes L = {l 1,…,l n }, C = D ∪ L For edge (u, v) we compute the strength a uv = f w (ψ uv )

6 Optimization p is the vector of PageRank scores A “soft” version

7 Algorithm

8 Experiments on Synthetic Data A scale-free graph G with 10,000 nodes Evaluated by classification accuracy Strength func. *AUC – Area under the ROC curve. 1.0 means perfect classification and 0.5 means random guessing.

9 Experiments on Real Data Four co-authorship networks and the Facebook network of Iceland Strength func.

10 Interaction Procedure The method basically converges in only about 25 iterations

11 Results LR: logistic regression, Prec@20: precision at top 20

12 Methods Comparison some unsupervised baselines & two supervised learning methods

13 Conclusion The Supervised Random Walks has great improvement over Random Walks. It outperforms supervised machine learning techniques It combines rich node and edge features with the structure of the network Apply to: recommendations, anomaly detection, missing link, and expertise search and ranking


Download ppt "Supervised Random Walks: Predicting and Recommending Links in Social Networks Lars Backstrom (Facebook) & Jure Leskovec (Stanford) Proc. of WSDM 2011 Present."

Similar presentations


Ads by Google