Recommendation in heterogeneous information network via dual similarity regularization

Abstract

Recommender system has caught much attention from multiple disciplines, and many techniques are proposed to build it. Recently, social recommendation becomes a hot research direction. The social recommendation methods tend to leverage social relations among users obtained from social network to alleviate data sparsity and cold-start problems in recommender systems. It employs simple similarity information of users as social regularization on users. Unfortunately, the widely used social regularization suffers from several aspects: (1) The similarity information of users only stems from social relations of related users; (2) it only has constraint on users without considering the impact of items for recommendation; (3) it may not work well for dissimilar users. To overcome the shortcomings of social regularization, we design a novel dual similarity regularization to impose the constraint on users and items with high and low similarities simultaneously. With the dual similarity regularization, we further propose an optimization function to integrate the similarity information of users and items under different semantic meta-paths, and a gradient descend solution is derived to optimize the objective function. Experiments with differ-This paper is an extension version of the PAKDD2016 Long Presentation paper "Dual Similarity Regularization for Recommendation" [40].

Recurrent knowledge graph embedding for effective recommendation

“…They design regularization terms for both users and items with the help of meta path based similarity. Similarly, Wang et al [37] and Zheng et al [45] also devise matrix factorization approaches by regularizing user-user relations with the computed meta path based similarity.…”

“…They connect various types of information related to items (e.g., genre, director, actor of a movie) in a unified global space, which helps to develop insights on recommendation problems that are difficult to uncover with user-item interaction data only. Stateof-the-art methods [15,28,37,42,45] mainly extend the latent factor model (LFM) [29] with entity similarity derived from paths (e.g., meta paths [33]) in a KG, based on the intuition that paths connecting two entities represent entity relations of different semantics. Such an intuition facilitates the inference of user preferences based Figure 1: A KG in the movie domain, which contains users, movies, actors, directors and genres as entities; rating, categorizing, acting, and directing as the entity relations.…”

Learning Topological Representation for Networks via Hierarchical Sampling

“…We evaluate our method on various real-world datasets, including Movielens 1 , MIT [23], DBLP [2], Douban [24], and Yelp 2 . These datasets are commonly used in NRL field.…”

Representation Learning for Heterogeneous Information Networks via Embedding Events

“…We evaluate our method on four real-world datasets, including DBLP [25], Douban [32], IMDB 1 and Yelp 2 . e brief information of each dataset is shown as follows.…”