Personalized content targeting starts with understanding the user based on what content he has watched so far and the corresponding ratings.
Following are the titles watched/rated by
Using group collaborating filtering we can predict a rating for all the titles John has not watched yet. (this is the standard algorithm used by most recommendation engines today).
Based on predicting the rating we get following top 20 titles for John :
We use the above learning algorithm to drive the following shelf
Usually, users are more attached to tags associated with the title. Tags can be genre, mood, cast, directors, awards etc. Tags provide a more semantic meaning to the title. Therefore, predicting the tag rating compared to title rating is more effective in understanding what user likes.
By using our tag rate prediction algorithm we get the following top 20 tags for John:
We use the above learning algorithm to drive the following shelves
Users interest can be tracked at a finer level via extracting the combination of tags. Our tag cluster extraction algorithm finds the following top 20 clusters for John .
Meta-X uses the tag-clustering concepts to build following shelves.
Set of movies watched by user has semantic similarity to other movies. A user having watched a certain titles creates a bias towards watching similar titles in a transitive way. For example, user watched title A will lead to watching title B which in turn will lead to watching title C. This pattern is determined by building an inference graph between the titles. Based on building the inference graph, we find that will be interested watch the following titles
From the above inference analysis we build a highly personalized targeting of content to users.
We now build the inference graph as above on the tag space. Inference graph discovers similarity between tags. For example, if a user watching titles having tags “action” or “thriller” may like to watch titles with tags “mystery” or “adventure”. Analysis provides us the following top tags will be interested in.
We leverage inference based tag prediction to build following shelves.
A more filtered version of the above inference based prediction is used to generate “Because you liked” related shelves. Instead of inferring similar title from all currently rated titles, we track most important titles “X” which reflects useful concepts. We then build shelves from obtaining titles similar to “X” as shown below.
We also build the inference graph on the tag space to find out the TV-Shows only.
