In the field of recommender systems, diversity as the measure of recommendation quality has gained much attention recently. Unfortunately, many researchers have shown that it has a trade-off relation with accuracy. Meanwhile, tensor factorization has been used as a useful technique that considers multi-correlations between user-item-other factors directly. However, it generally suffers from the model sparsity caused by high dimensionality and requirement of high computational costs. To improve diversity and response time while preserving accuracy in multi-criteria recommender systems (MCRS), we propose a decrease and conquer-based parallel tensor factorization (DnCPTF). In the DnCPTF, sentiment analysis alleviates the sparsity problem, and a two-phase clustering groups similar user reviews into sub-models. Furthermore, a controllable subdivision guarantees high diversity and short response time. The sub-models are then factorized in parallel to predict ratings, and top-N items are recommended via ratings consolidated from the sub-models. On a real-world dataset gathered from TripAdvisor, experimental results demonstrated that the DnCPTF significantly improve recommendation diversity (55× of a conventional tensor factorization (CTF)) and response time (182× of the CTF) with preserving high precision and recall. Furthermore, it outperformed recent techniques in precision, diversity and required response time within 1 s on average.

在推荐器系统领域中，多样性作为推荐质量的量度最近受到了广泛的关注。不幸的是，许多研究人员表明，它与精确度之间存在取舍关系。同时，张量分解已被用作直接考虑用户项-其他因素之间的多重相关性的有用技术。但是，它通常会受到以下原因造成的模型稀疏性的影响：高维度和高计算成本的要求。为了在保持多准则推荐系统（MCRS）准确性的同时提高多样性和响应时间，我们提出了一种基于减少和征服的并行张量因子分解（DnCPTF）。在DnCPTF中，情感分析缓解了稀疏性问题，并且两阶段聚类将相似的用户评论分组为子模型。此外，可控的细分可确保高度的多样性和较短的响应时间。然后将子模型并行分解以预测收视率，并通过从子模型合并的收视率来推荐前N个项目。在从TripAdvisor收集的真实数据集上，实验结果表明，DnCPTF显着改善了推荐多样性（传统张量因子分解（CTF）的55倍）和响应时间（CTF的182倍），同时保持了较高的准确性和查全率。此外，它在精度，多样性和所需的平均响应时间均在1 s之内超过了最新技术。