Methods, systems, and computer readable media for predicting ratings using graph filters

A method for computing user rating predictions, the method comprising:

• at a recommender implemented on at least one processor and memory storing executable instructions for the processor:
• implementing a recommender system using graph signal processing, wherein implementing the recommender system using graph signal processing includes:
• during a training phase, determining a plurality of filter coefficients for a graph filter from a training user-item dataset;
• after the training phase, receiving a request to select one or more recommended items for a user from a plurality of unrated items that are unrated by the user to the recommender;
• predicting ratings for the first user for each of the unrated items by applying the graph filter utilizing the determined filter coefficients to user-item rating data, wherein the user-item rating data comprises a plurality of ratings by the first user of the unrated items other than a first item, a plurality of ratings by users other than the first user of the first item, wherein applying the graph filter includes estimating a user relationship which quantifies similarity between tastes any two users for all items in the user-item rating data, estimating an item relationship which quantifies similarity between the any two items based on ratings from all users in the user-item rating data, and applying a transformation via at least one of the user relationship and the item relationship upon the user-item rating data, wherein the transformation weighs, based on the user relationship and the item relationship, the ratings that the users who, according to the user relationship, are similar and dissimilar to the first user gave to the items that, according to the item relationship, are similar and dissimilar to the first unrated item;
• and
• selecting the one or more recommended items from the unrated items using the user rating predictions for the unrated items and outputting the recommended items in response to the request.

A non-transitory computer readable medium having stored thereon executable instructions that when executed by at least one processor of a computer cause the computer to perform steps comprising:

• at a recommender implemented on at least one processor and memory storing executable instructions for the processor:
• implementing a recommender system using graph signal processing, wherein implementing the recommender system using graph signal processing includes:
• during a training phase, determining a plurality of filter coefficients for a graph filter from a training user-item dataset;
• after the training phase, receiving a request to select one or more recommended items for a user from a plurality of unrated items that are unrated by the user to the recommender;
• predicting ratings for the first user for each of the unrated items by applying the graph filter utilizing the determined coefficients to user-item rating data, wherein the user-item rating data comprises a plurality of ratings by the first user of the unrated items other than a first item, a plurality of ratings by users other than the first user of the first item, wherein applying the graph filter includes estimating a user relationship which quantifies similarity between tastes any two users for all items in the user-item rating data, estimating an item relationship which quantifies similarity between the any two items based on ratings from all users in the user-item rating data, and applying a transformation via at least one of the user relationship and the item relationship upon the user-item rating data, wherein the transformation weighs, based on the user relationship and the item relationship, the ratings that the users who, according to the user relationship, are similar and dissimilar to the first user gave to the items that, according to the item relationship, are similar and dissimilar to the first unrated item;
• and
• selecting the one or more recommended items from the unrated items using the user rating predictions for the unrated items and outputting the recommended items in response to the request.