Factorization Machine Based On Ranking Learning And Neural Network Enhancement

The recommender system algorithm based on earning to rank is an effective method for personalized recommendation of implicit feedback datasets.It transforms the recommendation problem into a ranking problem,i.e.,given a user and a set of items,find the optimal order of items that best match the user’s interests.Set Rank is a novel Bayesian collaborative ranking algorithm based on Setwise learning to rank,which effectively utilizes the partial order relationship between observed items and the set of unobserved items for ranking optimization.It proposes using permutation probabilities to encourage an observed item to be ranked ahead of a set composed of multiple unobserved items,with weaker independence assumptions and broader ranking constraints.Set Rank has achieved state-of-the-art results in Top-n recommendation.However,existing methods have some limitations,such as ignoring content information and assuming independent user behavior while neglecting the influence between users.This paper conducts an in-depth study of the Set Rank model,proposes and implements a Setwise ranking learning and neural network-enhanced factorization machine.The main research contents of the paper are as follows:(1)To solve the problem that Set Rank only considers collaborative information,ignores the effective utilization of content information carried by users and items themselves,and lacks effective modeling of content information,this paper combines entity content information with implicit feedback,selects factorization machines as predictors to model content information,and proposes a Setwise ranking factorization machine framework(SRFMs).(2)To address the problems of fixed feature representation and lack of high-order interaction in standard FM,and refine feature representation for different data scenarios while simultaneously learning low-order and high-order feature interactions,this paper combines input-aware and neural networks into the SRFMs framework to construct a Setwise ranking deep input-aware factorization machine model(SRDIFM).(3)To alleviate the problem that Set Rank neglects the possibility of influence or dependence between user behaviors,and to overcome the limitation of independence assumptions,this paper introduces multiple set rankings and user behavior similarity to mine potential preferred items,divides unobserved item sets,and proposes a Multi-Setwise ranking factorization machine framework(MSRFMs)and MSRDIFM models.The modeling ability of the user’s intrinsic dependence relationship in user activity in implicit feedback datasets is further enhanced while maintaining the advantages of the original models.This paper conducts experimental evaluations on two real datasets and compares with some classical and state-of-the-art methods.The experimental results show that the proposed models achieve the best performance and have significant advantages in various evaluation metrics,verifying the effectiveness and stability of the proposed models.