基于知识感知采样的神经协同过滤

Neural Collaborative Filtering Based on Knowledge-Aware Sampling

目前,对于推荐系统的研究主要集中在如何使用不同的辅助数据进行混合推荐,来缓解数据稀疏和冷启动问题,以提高推荐的准确性,而对于自然噪声的研究相对较少。针对自然噪声中项目的曝光偏差问题,提出一种基于知识感知的负采样策略。该方法利用用户对项目的隐式反馈和项目的知识信息来构建知识图谱,然后根据知识图谱中正例与负例之间共同的知识实体来对负例进行采样。通过设计一个奖励函数来衡量采样负例的质量,然后通过最大化累积奖励函数期望来优化采样器。采样以后,将用户正例和最佳负例一起输入到推荐器中,用于训练推荐模型。随后,将该采样方法与神经协同过滤结合起来,得到一种基于知识感知采样的神经协同过滤。此外,为了能够灵活地控制采样比,用逐点损失代替原来的成对损失来优化推荐器。在真实数据集上进行了广泛实验,验证了所提方法的有效性。

At present,the research on recommendation systems is mainly focused on how to use different auxiliary data for mixed recommendation to alleviate data sparseness and cold start problems to improve the accuracy of recommendation,and there are relatively few researches on natural noise.Aiming at the exposure bias of items in natural noise,a negative sampling strategy based on knowledge perception is proposed.This method uses the user’s implicit feedback on the project and the knowledge information of the project to construct a knowledge graph,and then samples the negative examples according to the common knowledge entity between the positive and negative examples in the knowledge graph.A reward function is designed to measure the quality of sampled negative examples,and then optimizing the sampler by maximizing the cumulative reward function expectation.After sampling,the user’s positive examples and best negative examples are input into the recommender together for training the recommendation model.Subsequently,this sampling method is combined with neural collaborative filtering to obtain a neural collaborative filtering based on knowledge-aware sampling.In addition,in order to be able to flexibly control the sampling ratio,a point-by-point loss is used instead of the original paired loss to optimize the recommender.Extensive experiments have been conducted on real data sets to verify the effectiveness of the proposed method.