Personalized Recommendations using Knowledge Graphs

Summary
In this talk, Rose Catherine discussed their recent work on personalized recommender systems by using Knowledge Graphs (KG), which has been leveraged as external information to supplement content-based recommendation. Content-based methods depend on general contents, such as users’ demographics, a movie’s actors, genre, directors and release countries, or a restaurant’s location and cuisine, etc. However, not all of them use the interconnections between the contents or external knowledge sources, which are referred to knowledge graphs. A movie recommendation example is depicted in Figure 1. Entity contains users, movies, actors and directors. The links in the figure show the content association with each of the movies as well as the movies that each of the users watched.

Figure 1. Example of movie recommendation.

In their work, they investigate three methods for KG based recommendation, by using a general-purpose probabilistic logic system called Programming with Personalized PageRank (ProPPR) which was firstly proposed by W. Y. Wang, K. Mazaitis, and W. W. Cohen, 2013 [1]. The first and simplest method EntitySim learns users’ preferences over the contents and leverages the link structure of the KG to make predictions. The second method TypeSim extends EntitySim by additionally considering type popularities and type similarities. Beyond that, they also present a complicated method called GraphLF which is a latent factorization model over KG.

To evaluate the performances of their methods, they compare results with the state-of-the-art graph based recommendation strategy HeteRec_p [2]. By testing on two datasets constructed from Yelp and IMDb, it is shown that KG based algorithm can obtain large improvements, despite that no consistent results can prove that which method is the best one among EntitySim, TypeSim and GraphLF. Even though, KG shows the most prominent advantages especially dataset is sparse (see Figure 2), implying its valuable applications in real-word setting.

Figure 2. Performance comparison between different methods by varying data density of Yelp.

[1] W. Y. Wang, K. Mazaitis, and W. W. Cohen. Programming with personalized pagerank: A locally groundable first-order probabilistic logic. In Proc. CIKM ’13, pages 2129–2138, 2013.

[2] X. Yu, X. Ren, Y. Sun, Q. Gu, B. Sturt, U. Khandelwal, B. Norick, and J. Han. Personalized entity recommendation: A heterogeneous information network approach. In Proc. WSDM, 2014.

Details about talk

Title: Personalized Recommendations using Knowledge Graphs

Speaker: Rose Catherine

Talk URL: http://halley.exp.sis.pitt.edu/comet/presentColloquium.do?col_id=10665