融合主题信息和卷积神经网络的混合推荐算法

针对传统的协同过滤算法中数据稀疏和推荐结果不准确的问题，提出了一种基于隐狄利克雷分布（LDA）与卷积神经网络（CNN）的概率矩阵分解推荐模型（LCPMF），该模型综合考虑项目评论文档的主题信息与深层语义信息。首先，分别使用LDA主题模型和文本CNN对项目评论文档建模；然后，获取项目评论文档的显著潜在低维主题信息及全局深层语义信息，从而捕获项目文档的多层次特征表示；最后，将得到的用户和多层次的项目特征融合到概率矩阵分解（PMF）模型中，产生预测评分进行推荐。在真实数据集Movielens 1M、Movielens 10M与Amazon上，将LCPMF与经典的PMF、协同深度学习（CDL）、卷积矩阵因子分解模型（ConvMF）模型进行对比。实验结果表明，相较PMF、CDL、ConvMF模型，所提推荐模型LCPMF的均方根误差（RMSE）和平均绝对误差（MAE）在Movielens 1M数据集上分别降低了6.03%和5.38%、5.12%和4.03%、1.46%和2.00%，在Movielens 10M数据集上分别降低了5.35%和5.67%、2.50%和3.64%、1.75%和1.74%，在Amazon数据集上分别降低17.71%和23.63%、14.92%和17.47%、3.51%和4.87%，验证了所提模型在推荐系统中的可行性与有效性。

Hybrid recommendation algorithm by fusion of topic information and convolution neural network

Aiming at the problems of data sparsity and inaccuracy of recommendation results in the traditional collaborative filtering algorithms, a Probability Matrix Factorization recommendation model based on Latent Dirichlet Allocations (LDA) and Convolutional Neural Network (CNN) named LCPMF was proposed, which considers the topic information and deep semantic information of project review document comprehensively. Firstly, the LDA topic model and the text CNN were used to model the project review document respectively. Then, the significant potential low-dimensional topic information and the global deep semantic information of project review document were obtained in order to capture the multi-level feature representation of the project document. Finally, the obtained features of users and multi-level projects were integrated into the Probability Matrix Factorization (PMF) model to generate the prediction score for recommendation. LCPMF was compared with the classical PMF, Collaborative Deep Learning (CDL) and Convolutional Matrix Factorization (ConvMF) models on the real datasets Movielens 1M, Movielens 10M and Amazon. The experimental results show that, compared to PMF, CDL and ConvMF models, on the Movielens 1M dataset, the Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) of the proposed recommender model LCPMF are reduced by 6. 03% and 5.38%, 5.12% and 4.03%, 1.46% and 2.00% respectively; on the Movielens 10M dataset, the RMSE and MAE of LCPMF are reduced by 5.35% and 5.67%, 2.50% and 3.64%, 1.75% and 1.74% respectively; while on the Amazon dataset, the RMSE and MAE of LCPMF are reduced by 17.71% and 23.63%, 14.92% and 17.47%, 3.51% and 4.87% respectively. The feasibility and effectiveness of the proposed model in the recommendation system are verified.