查看文章

Recommender systems are ubiquitous in today’s customer facing services. Traditionally, researchers have modelled recommender system as a collaborative filtering problem. Data prepared for this type of problem can be represented by a tuple (c, r, i), indicating customer c has interacted through relation r with item i. These tuples can also be stored in matrix form, with each array element representing the user’s prior interactions with the items (eg watched, rated, bought the item) for each user-item pair. The downside of traditional collaborative filtering method is that it only makes use of information in the utility matrix but not metadata available for items. In other words, the problem formulation only looks for similar users who also bought the item, or similar items bought by the same user, ignoring other similarity in items such as movies starred by the same actor or books authored by the same person. Inclusion of knowledge graph addresses this issue, but introduces a rather sophisticated set of new relations and millions more entities. Combining user-item interactions with knowledge graph results in a hybrid graph that requires more advanced handling and training methods. Current state-of-the-art models which build on graph convolutional networks retain existing neighborhood structure as-is. While this approach respects the given network structure, it leaves out potential correlations among similar items and structures in the network. As pointed out by Wu (1), distant neighbors bring little information, thus decreasing the utility of additional convolution layers, yet no existing work can be found on alternative local neighborhood definition. More …