Enhancing Factorization Machines With Generalized Metric Learning

Factorization Machines (FMs) are effective in incorporating side information to overcome the cold-start and data sparsity problems recommender systems. Traditional FMs adopt inner product model second-order interactions between different attributes, which represented via feature vectors. The problem is that violates triangle inequality property of As a result, it cannot well capture fine-grained attribute interactions, resulting sub-optimal performance. Recently, euclidean distance exploited replace has delivered better However, previous FM methods including ones equipped with all focus on attribute-level interaction modeling, ignoring critical intrinsic correlations inside attributes. Thereby, they fail complex rich exhibited real-world data. To tackle this problem, paper, we propose framework generalized metric learning techniques these correlations. In particular, based framework, present Mahalanobis deep neural network (DNN) methods, can effectively linear non-linear features, respectively. Besides, design an efficient approach for simplifying functions. Experiments several benchmark datasets demonstrate our proposed outperforms state-of-the-art baselines by large margin. Moreover, collect new large-scale dataset second-hand trading justify effectiveness method over