PROJECT TITLE :

Enhancing Factorization Machines with Generalized Metric Learning

ABSTRACT:

The cold-start and data sparsity issues that plague recommender systems can be circumvented with the help of factorization machines (FMs), which are effective in incorporating side information. Traditional FMs use the inner product to model the second-order interactions between the various attributes, which are represented by feature vectors. This is done in order to simplify the modeling process. The issue is that the inner product does not adhere to the triangle inequality property of feature vectors, which is the source of the problem. As a consequence of this, it is unable to adequately capture fine-grained attribute interactions, which results in performance that is less than optimal. Recently, the Euclidean distance has been utilized in FMs to take the place of the inner product, which has resulted in improved overall performance. However, earlier FM methods, including the ones that were equipped with the euclidean distance, all focus on the modeling of interactions at the attribute level, while ignoring the crucial intrinsic feature correlations that exist within attributes. As a result, they are unable to model the intricate and varied interactions that are shown in the data coming from the real world. In this paper, we propose an FM framework that is equipped with generalized metric learning techniques to better capture these feature correlations. This will help us address the issue that we are currently facing. Specifically, we present a Mahalanobis distance method and a deep neural network (DNN) method, both of which are based on this framework. These methods are able to effectively model the linear and non-linear correlations between features, respectively. In addition to this, we develop an effective strategy for simplifying the model's functions. Experiments conducted on a number of benchmark datasets have shown that our proposed framework performs significantly better than a number of state-of-the-art baselines. In addition, we gather a brand new massive dataset on second-hand trading so that we can demonstrate the efficacy of our approach in overcoming cold-start and data sparsity issues that are common in recommender systems.