Recommendation method for graph construction and L1 regular matrix decomposition combination learning

Abstract

The invention discloses a recommendation method for graph construction and L1 regular matrix decomposition combination learning. The method comprises the following steps that: (1) according to user scoring data, establishing a scoring matrix; (2) on the basis of an L1 regular matrix decomposition model, taking graph edge weight as an unknown variable to be embedded into a target function so as to establish a new model; (3) on the basis of a score or a tag or a topic model, calculating an initial similarity between objects, and describing the similarity by a graph to obtain the initial graph edge weight; (4) randomly initializing a latent semantic feature vector to be solved in a decomposition model; (5) on the basis of the initial graph edge weight, using an OWL-QN (Orthant-Wise Limited-Memory Quasi-Newton) algorithm to adaptively update parameters in the decomposition model; and (6) obtaining an integral scoring matrix, and providing recommendation for a user according to the scoring matrix. By use of the method, through combination learning, the effective graph edge weight and the latent semantic feature vector can be obtained through the combination learning so as to obtain good prediction presentation.

Description

technical field [0001] The invention relates to a recommendation method for joint learning of graph construction and L1 regular matrix decomposition, and belongs to the technical field of recommendation learning. Background technique [0002] At present, existing recommendation technologies include content-based recommendation, collaborative filtering recommendation, and so on. The content-based recommendation algorithm constructs user preference information based on some historical information of users, calculates the similarity between recommended items and user preferences, and recommends items with high similarity to users. Collaborative filtering algorithms can be divided into neighborhood-based collaborative filtering algorithms and matrix factorization algorithms. Collaborative filtering recommendation algorithms based on neighbors include user-based collaborative filtering recommendation algorithms and item-based collaborative filtering algorithms. Matrix decomposi...

AI Technical Summary

Problems solved by technology

[0004] However, the traditional collaborative filtering method will be affected by data sparsity. In the collaborative filtering algorithm based on users and items, how to measure the similarity between users and items is the key link. Whether the similarity is effective directly affects the recommendation algorithm. Performance; in the recommendation algorithm based on manifold regularization matrix decomposition, due to data sparsity or incomplete label information, the neighbor graph (similarity) obtained based on sparse score or incomplete label calculation may not be effective, Moreover, such neighborhood graphs are local compositions, leading to difficulties in parameter selection (e.g., choosing the size of the neighborhood), and invalid graphs that are fixed in the learning task will produce a bad recommendation

Images