基于多目标演化优化的个性化推荐算法研究

Improving accuracy and diversity is an important research direction in personalized recommendation. The difficulties of solving this problem are to establish robust models and design efficient optimization methods for personalized recommendation. On the one hand, focusing on improving recommendation accuracy and diversity, we will define effective objective functions and model personalized recommendations as multi-objective problems in this proposal. Effective and robust multi-objective evolutionary optimization algorithms for multi-objective personalized recommendation will be proposed. On the other hand, user’s locations will be introduced into multi-objective personalized recommendation. We will establish location-based recommendation models based on multi-objective optimization by combining locality preference and user’s history behaviors preference. Then efficient multi-objective evolutionary optimization approaches will be designed to solve these models. We aim to improve accuracy and diversity of recommendation and alleviate data sparsity and cold-start problems. Based on these multi-objective personalized recommendation models, we will solve multi-objective optimization problems of recommending popular and unpopular locations in real travel recommendation. This proposal has high theoretic meaning and realistic value for multi-objective optimization problems in personalized recommendation.

提高推荐准确性和多样性是个性化推荐中一个重要的研究内容。其难点是对个性化推荐的建模与高效优化。本项目针对推荐准确性和多样性两难问题，定义合适的目标函数，将个性化推荐建模为多目标优化问题，设计高效稳健的多目标个性化推荐演化优化算法；在多目标个性化推荐中引入用户的位置信息，结合用户的位置偏好和历史行为偏好，建立位置感知个性化推荐的多目标优化模型，设计相应的多目标演化优化算法。通过本项目的研究，在提高推荐准确性和多样性的同时，缓解个性化推荐的数据稀疏和冷启动问题；在多目标个性化推荐理论基础上，研究真实旅游场景中热门景点和冷门景点推荐的多目标优化问题。本项目的研究对个性化推荐系统中的多目标优化问题有重要的理论意义和应用价值。

本课题在执行过程中严格按照研究计划进行，主要研究成果包括：.本课题针对推荐系统中推荐准确率和多样性两难优化问题，设计相互冲突的目标函数分别衡量推荐算法的准确性和多样性，同时分析推荐系统中用户位置等因素对推荐结果的影响，将个性化推荐问题建模为多目标优化问题。.提出了高效的多目标进化算法求解该多目标推荐优化问题。针对进化多目标推荐优化中全局搜索与局部搜索的平衡、算法参数的自适应调整，设计了适用于多目标优化的参数自适应学习模型，提出了局部搜索能力强、具备参数自适应学习能力的多目标进化算法，解决了进化多目标推荐优化中全局与局部搜索的自适应平衡难题。所设计的进化多目标优化算法提高了推荐算法准确性，并且可以很好地缓解数据稀疏和冷启动问题。.上述研究成果在《Information Sciences》、《Applied Soft Computing》等期刊共计发表高水平论文5篇，申请/授权国家发明专利7项，培养博士2人、硕士研究生5人，研究成果获2020年陕西省自然科学一等奖，圆满完成了预期研究目标。.