[Talk Summary 3] Personalized Recommendations using Knowledge Graphs

Rose Catherine from Language Technologies Institute, Carnegie Mellon University presented the talk Personalized Recommendations using Knowledge Graphs by 09/23/2016.

At the beginning, Rose introduced general ideas about personalized recommendations and gave some examples such as Amazon recommending customers products they would like to buy and introducing movies to users based on their favorite ones.

And then, Rose focused on improving the performance of recommender systems using knowledge graphs (KGs). She introduced other works that use the combination of content-based and collaborative filtering techniques and connected concept-based KGs on recommendations. To improve the current methods, Rose's research group investigates three techniques for making recommendations. Those KG-based recommendations use a probabilistic logic system called ProPPR which stands for Programming with Personalized PageRank.
  • The first approach is EntitySim that uses only the links of the graph to find the seedset of the user. Which means finding a set of entities in which each user is interested. From that, a random walker starts from a node in the set to other entities in the graph.
  •  The second model extended from EntitySim is called TypeSim that considers the types of entities to boost its capabilities. it models the general popularity of each types by learning the overall predictability offered by the entity's type. For example, users is likely to prefer the actors of a movie more than the country of a movie. Additionally, the model learns the overall
    predictability offered by the entity itself.
  • The third method is GraphLF that integrates latent factorization and graph-based recommendation. And, this method does not use type to increase its applicability to a wide range of data domains because types are not always available.
 Model complexity:
  • EntitySim: O(n)
  • TypeSim: O(n + e + t2)    
  • GraphLF: O(n + m + e)
Finally, Rose showed the experiments that compare her methods with the state-of-the-art method HeteRec_p and NB recommender on the two datasets, Yelp2013 and IM100k, in different settings. The results generally proved that her methods were able to achieve a large improvement in performance. However, NB recommender's performance which is poor at low densities improves with higher densities and eventually is better than all the KG-based methods.

Thanks to the conclusion, to build a good recommendation system, we should consider what kind of datasets we will use and their densitiy to choose the effective methods.

407A/B BAUM, The Offices at Baum

5607 Baum Blvd
University of Pittsburgh


Popular posts from this blog

[Talk Summary] Machine Learning and Privacy: Friends or Foes?

SoRec: Social Recommendation Using Probabilistic Matrix Factorization