Difference between revisions of "Konstas et al. SIGIR 2009"

From Cohen Courses
Jump to navigationJump to search
Line 31: Line 31:
 
[[File:Last.fm.graph.png]]
 
[[File:Last.fm.graph.png]]
  
−
== Algorithm description ==
+
== Experiment results and observations==
 +
The experiment result is concluded as following table:
 +
[[File:Last.fm.result.png]]
  
−
#Construct a word relatedness graph
+
There are several observations from the result worth noticing:
−
#Define a random walk on the graph
 
−
#Compute the word's [[UsesMethod::hitting time]] for both the positive and negative sets of vertices
 
−
#If the hitting time for the positive set is greater than for the negative set, than the word is classified as negative. Otherwise, it is classified as positive. The ratio between the two hitting times could be used as an indication of how positive/negative the given word is.
 
  
−
Since computing the hitting time is time consuming especially when the graph is large, a Monte Carlo based estimating algorithm is proposed as such:
+
*The addition of social network and social tagging information increases the number of relevant retrieved (num_rel_ret) tracks in the [[Random walk with restart|RWR]] method.
−
[[File:Word_polarity_using_random_walks.png]]
+
*There is also a notable decrease in MAP and Precision at high ranks and in parallel to num_rel_ret increase and Precision at lower ranks. This tendency accounts for the fact that with the progressive addition of social knowledge the method retrieves more tracks thus effectively increases its recall but at lower ranks (e.g P@200), causing a harm to its precision.
 +
*the [[Random walk with restart|RWR]] method with the UTrUTg graph retrieves statistically significantly more relative tracks than the RWR method with the UTrUU graph (and also has higher P@200 and P@1000) compared to the simple [[Random walk with restart|RWR]] method.
 +
The above three observation highlights the benifits of the social graph.
 +
 
 +
*Using the same sub-matrix as the input, the [[Random walk with restart|RWR]] method always outperforms the [[Social_collaborative_filtering | CF]] method. This supports the argument that the [[Random walk with restart|RWR]] method is more effective than the [[Social_collaborative_filtering | CF]] method.
 +
 
 +
*Using the [[Social_collaborative_filtering | CF]] method, the addition of friendships and social tagging information deteriorates the performance increasingly. This supports the argument that the memory based [[Social_collaborative_filtering | CF]] method cannot provide with adequate non-trivial mechanisms to incorporate social knowledge.
  
 
== Experiment result ==
 
== Experiment result ==

Revision as of 04:41, 2 October 2012

Citation

Ioannis Konstas, Vassilios Stathopoulos, and Joemon M. Jose. On social networks and collaborative recommendation. In Proceedings of the 32nd international ACM SIGIR conference on Research and development in information retrieval, SIGIR ’09, pages 195–202, New York, NY, USA, 2009. ACM.

Online Version

online version

Summary

This paper uses Random walk with restart as a method of content recommendation in social network systems.

As it is in the context of social network, naturally and effectively, a graph based algorithm such as Random walk with restart can be considered to perform the recommendation task.

This paper argues that the extra knowledge provided by the users' social activity, such as the social annotation and friendships inherent in the social graph established among users, items and tags, can improve the performance of a recommendation system using methods such as Random walk with restart.

The argue is supported based on a series of comparison experiments between the Random walk with restart model and a user-based collaborative filtering method using the Pearson Correlation similarity. The results show that the graph model system benefits from the additional information embedded in social knowledge. In addition, the graph model outperforms the standard collaborative filtering method.

The dataset

The source of dataset comes from last.fm - a Web 2.0 website. As introduced in the dataset, there are three components serving as the vertices in the graph - users, music tracks and tags, deriving relation matrices as User-Track(UTr), User-Tag(UTg) and Track-Tag(TrTg).

The UTr sub-matrix consists of the playcount of every track by each user, i.e. the number of times it has been listened to.

The UU sub-matrix consists the average user playcount. This is crucial because the UTr sub-matrix may contain very large values and thus the binary values in UU will be suppressed.

In the case of the UTg sub-matrix, the associated tags for each user have been collected based on popularity. we applied an exponential decay function to the values of the tags of each user in the matrix, where the most popular tag gets the average user's playcount. The same process was performed for the TrTg sub-matrix accordingly.

The full social graph look like:

Last.fm.graph.png

Experiment results and observations

The experiment result is concluded as following table: Last.fm.result.png

There are several observations from the result worth noticing:

  • The addition of social network and social tagging information increases the number of relevant retrieved (num_rel_ret) tracks in the RWR method.
  • There is also a notable decrease in MAP and Precision at high ranks and in parallel to num_rel_ret increase and Precision at lower ranks. This tendency accounts for the fact that with the progressive addition of social knowledge the method retrieves more tracks thus effectively increases its recall but at lower ranks (e.g P@200), causing a harm to its precision.
  • the RWR method with the UTrUTg graph retrieves statistically significantly more relative tracks than the RWR method with the UTrUU graph (and also has higher P@200 and P@1000) compared to the simple RWR method.

The above three observation highlights the benifits of the social graph.

  • Using the same sub-matrix as the input, the RWR method always outperforms the CF method. This supports the argument that the RWR method is more effective than the CF method.
  • Using the CF method, the addition of friendships and social tagging information deteriorates the performance increasingly. This supports the argument that the memory based CF method cannot provide with adequate non-trivial mechanisms to incorporate social knowledge.

Experiment result

Comparing to other methods, this method is quite successful in both the settings of semi-supervised and unsupervised.

  • In the setting of using WordNet synonyms and hypernyms to construct the network and test set to the set of adjectives. It out performs the spin-model, bootstrap and short-path method.

Hassan & Radev ACL 2010 E1.png

  • It is also compared to the SO-PMI method in the setting of only 14 seeds. Though SO-PMI with a very large dataset performs slightly better than this method, this method is faster and does not need such large corpus.

Hassan & Radev ACL 2010 E2.png

Related papers

Study plan

Retrieved from "http://curtis.ml.cmu.edu/w/courses/index.php?title=Konstas_et_al._SIGIR_2009&oldid=13408"