Leveraging Kernel-Incorporated Matrix Factorization for App Recommendation
Chenyang Liu, Jian Cao, Shanshan Feng · ACM Transactions on Knowledge Discovery from Data · 2019
The ever-increasing number of smartphone applications (apps) available on different app markets poses a challenge for personalized app recommendation. Conventional collaborative filtering-based recommendation methods suffer from sparse and binary user-app implicit feedback, which results in poor performance in discriminating user-app preferences. In this article, we first propose two kernel incorporated probabilistic matrix factorization models, which introduce app-categorical information to constrain the user and app latent features to be similar to their neighbors in the latent space. The two models are solved by Stochastic Gradient Descent with a user-oriented negative sampling scheme. To further improve the recommendation performance, we construct pseudo user-app ratings based on user-app usage information, and propose a novel kernelized non-negative matrix factorization by incorporating non-negative constraints on latent factors to predict user-app preferences. This model also leverages user--user and app--app similarities with regard to app-categorical information to mine the latent geometric structure in the pseudo-rating space. Adopting the Karush--Kuhn--Tucker conditions, a Multiplicative Updating Rules based optimization is proposed for model learning, and the convergence is proved by introducing an auxiliary function. The experimental results on a real user-app installation usage dataset show the comparable performance of our models with the state-of-the-art baselines in terms of two ranking-oriented evaluation metrics.