【摘要】 针对个性化推荐系统中用户偏好的进化学习与高维稀疏数据处理的问题。受隐马尔科夫模型(HMM)结构特征启发,提出了一种考虑上下文感知的两阶段用户偏好集推理策略的个性化推荐算法(HHRA算法)。通过对系统历史评分信息的处理,将用户偏好的提取过程抽象为一个HMM模型,来进行第一阶段的用户偏好集学习与推理。然后在此基础上,引入用户的实时上下文信息,构建了一种融入用户实时偏好的张量模型,并基于一种改进的高阶奇异值分解算法来处理高维稀疏的数据集,对模型进行优化求解,生成最优推荐集合。实验设计在3个具有不同特征的真实数据集上将HHRA算法与传统经典推荐算法进行对比分析,结果显示HHRA算法具有较好的适应性和推荐质量。更多还原

【Abstract】 Evolutionary learning of user preference and the processing of high-dimensional sparse data have emerged as an important topic issue in a personalized recommendation system.Inspired by the structural features of hidden Markov model( HMM),this paper proposed a personalized recommendation algorithm( HHRA) based on a strategy for two-stage user preference sets inference considering context awareness.In the first stage,the learning and inference of user preference sets were conducted through processing historical information of system scoring.Meanwhile,the extraction process of user preference was abstracted as a HMM model.Then,in the second stage,user’s real-time context information was introduced to construct a tensor model containing real-time user preference.To deal with high-dimensional sparse datasets,an improved high-order singular value decomposition(HOSVD) algorithm was adopted to provide optimization solutions for the proposed model and generate sets with optimal recommendation.Comparison analysis was performed between HHRA algorithm and traditional recommendation algorithms by using three real-life datasets with different characteristics derived from experiment designs.Results show that HHRA algorithm has better adaptability and recommendation quality performance.更多还原