There are many recommender systems out there. As a general rule, they look at some properties of the item and match that to each individual user. This paper explains how you can use network analysis to derive 2 properties of the item being recommended. The authors focused on message items from Orkut as their data set.
The two properties being derived from the social network are context and completeness.
“Context of a message relates to its comprehensibility (Maglaughlin & Sonnenwald 2002) or the simplification it
provides (Bryant & Zillman 2002), based on how well the message content explains the relationship of the message to
its recipient. Thus, comprehensibility and simplification can be considered as outcomes of the amount of context in the
message; messages that are more contextual will be more comprehensible and improve understanding for recipients.”
“Completeness of a message denotes the depth and breadth of topics covered by the message. A concrete definition of
depth and breadth is proposed by (Zhu & Gauch 2000), as the depth and breadth of the topic ontology graph covered
by the message. The scope of the message (Maglaughlin & Sonnenwald 2002), or the opinion diversity provided by
the message (Bryant & Zillman 2002), can be considered as outcomes of the amount of completeness in the message.”
How do graph or network analysis come in?
The intuitive perspective here is that user has STRONG ties to the author of a message, then the context should be higher. This makes sense because birds of the same feather flock together. Similarly, if the user has WEAK ties to the author of the message, then the completeness should be higher.
So it’s possible to estimate the completeness and context using graph (network) metrics. The authors have to first identify STRONG and WEAK ties between users. This is accomplished using a clustering algorithm to group users together. Those that are grouped together have strong ties; those that don’t, have weak ties.
Then for each user, the authors proceed to calculate 3 other network metrics to further improve the model : a local clustering coefficient for each cluster to estimate the degree of shared context, an integration coefficient for each user to estimate the amount of context provided by each user’s messages; and a local credibility of the user.
Now context = clustering coefficient x integration coefficient. Completeness is a little more complicated and has to include the user’s relationship to other clusters (remember weak ties means a user bridges many different clusters). Using a variety of derivations of these metrics, 6 metrics were calculated.
- NA = Evidence variable for the new amount of context provided by a message
- NO = Evidence variable for the current amount of context already provided to the user
- MO = Evidence variable for the current amount of completeness already provided to the user
- MA = Evidence variable for the new amount of completeness provided by a message
- NF = Evidence variable for freshness of the contextual information provided by a message
- MF = Evidence variable for freshness of the completeness provided by a message
A learning phase is executed against data from Orkut, a popular social networking site. Different functions of the difference evidence measures were used as classifiers.
Results
The diagram shows the plots of TPR (true positive rates) versus FPR (false positive rates) for the various functions of the evidence variables.
As can be seen, many of the models score well on having high true positive rates while keeping false positives low.
So what?
There is much more refinement to this approach that can be taken. But in general it points to the possibility of using a social network based approach to recommend digital content items to users. Since there are no “pagerank” type methods to rank order digital content, using this SNA approach is the next best thing by combining user views and interactions as indications of the relevance of digital content items.
This recommendation approach should be applied to any website that has content wrapped by social network features.
References
Aaditeshwar Seth, Jie Zhang. A Social Network Based Approach to Personalized Recommendation of Participatory Media Content. Int. Conf. on Weblogs and Social Media (ICWSM) (2008)
Tags: recommendations, Social networks
