面向多属性推荐系统的对抗深度分解模型

Adversarial Deep Tensor Factorization for Multi-criteria Recommender Systems

对于单属性推荐系统,矩阵分解是广泛使用的方法之一,但对于包含多个不同属性的多属性推荐系统而言,矩阵分解方法效果存在局限性。目前已经开发了张量分解方法,以通过处理三维(3D)用户项目属性评分来学习多属性推荐系统中的预测模型。但是,它们在实际应用中会遇到数据稀疏和数据污染的问题。为了克服这些问题,该文提出了一种将深度表示学习和张量因子分解相结合的对抗性深度张量因子分解模型(adversarial deep tensor factorization, ADTF)的通用体系结构,其中嵌入了辅助信息以有效补偿张量稀疏性,并且采用了对抗性学习以增强模型的鲁棒性。通过结合对抗性堆叠降噪自动编码器(adversarial stacked denoising autoencoder, ASDAE)和CANDECOMP/PARAFAC(CP)张量因子分解来展示ADTF架构的实例化案例,其中用户和商品的额外信息都与稀疏的多属性评分紧密结合,而对抗性训练则是用于学习有效的潜在因子向量。在三个真实数据集上的实验结果表明,该ADTF方案优于多属性评分预测的基准方法。

Matrix factorization is one of the most successful methods for single-criterion recommender systems but not for multi-criteria recommender systems that contain multiple criterion-specific ratings.Tensor factorization has been developed to learn predictive models in multi-criteria recommender systems by dealing with the three-dimensional(3D)user-item-criterion ratings.However,they suffer from the data sparsity and contamination issues in real applications.In order to overcome these problems,we propose a general architecture of adversarial deep tensor factorization(ADTF)by integrating deep representation learning and tensor factorization,where the side information is embedded to provide an effective compensation for tensor sparsity,and the adversarial learning is adopted to enhance the model robustness.We exhibit a specific ADTF instantiation by combining adversarial stacked denoising autoencoder(ASDAE)and CANDECOMP/PARAFAC(CP)tensor factorization,where the side information of both users and items is tightly coupled with the sparse multi-criteria ratings,and the adversarial training is used on learning effective latent factors rather than on the extrinsic rating inputs.Experimental results on three real-world datasets demonstrate that the proposed ADTF outperforms state-of-the-art methods on multi-criteria rating predictions.