【作者】 王斌；

【导师】 高琪；

【作者基本信息】 北京理工大学 ， 控制科学与工程， 2018， 硕士

【摘要】 推荐算法是大数据时代解决信息过载问题强有力的技术手段之一,在理论研究和现实应用层面都具有深远意义和重大价值。目前,协同过滤推荐算法已广泛应用于电子商务等领域,为企业创造了巨大的经济价值。准确地推荐符合用户兴趣的项目是推荐算法的基本任务,但用户兴趣通常是多种多样的并且处于变化之中,因此,准确性与多样性困境成为了协同过滤面临的重要挑战之一。本文以准确性-多样性两难问题为研究目标,在协同过滤的框架之下,综合利用复杂网络、多目标优化等理论与方法,提出了一些有效的解决方案,并通过实验验证了算法效果。基于最近邻的推荐算法过于关注流行度高的项目而忽略长尾项目,本文提出了一种结合最近邻居和二阶邻居的推荐策略。该策略在最近邻居的邻域内寻找目标用户的二阶邻居,然后引入可调参数进行权重调整,最后利用最近邻居和二阶邻居共同进行推荐。多个数据集上的实验表明,当参数处于特定区间时,基于该策略的算法在保持足够高准确率的前提下,能较为明显地提高算法多样性和覆盖率。此外,算法在稀疏度低的数据集上的表现更突出。主流相似度方法关注用户间的项目交集,未能全面描述用户兴趣偏好上的相关性。本文提出一种改进的相似度方法,该方法将用户间相似度划分为历史相似度和趋势相似度,并通过有交集项目计算历史相似度,同时利用无交集项目计算趋势相似度。多个标准数据集上的实验表明,相较于其他改进算法,基于本文所提改进相似度方法的算法在准确率上所受到的影响最小;在准确率影响程度相同时,该算法对覆盖率和多样性的提高幅度更大。基于邻域的推荐中,最近邻之间存在严重的冗余项目信息,这使得少数项目权重被放大,同时推荐的覆盖率被限制。本文对任意目标用户建立了考虑准确性和多样性的双目标优化问题的模型,然后利用基于分解的多目标优化算法求解该问题,以寻找能尽可能代表用户多个维度兴趣的最优邻居。不同于协同过滤,该算法可以筛选多组最优邻居进行推荐。标准数据集上的实验表明,本文所提算法在保证推荐准确率相差不多甚至有所提高的前提下,明显地提高了推荐的多样性和覆盖率。准确性-多样性难题是推荐面临的重大挑战之一,本文对此进行多角度分析,同时提出了针对性的解决方案,最后在多个标准数据集上进行验证。实验表明,在保持足够高准确率甚至将准确率有所提高的前提下,所提算法显著地提高了推荐的多样性和覆盖率。同时,本文也为准确性-多样性难题提供了新的研究视角。更多还原

【Abstract】 As one of the most powerful technological means to address the information overload problem in the era of big data,the recommendation algorithms is of great significance and long-term value both in theoretical researches and practical applications.Collaborative filtering has been increasingly applied to lots of fields such as e-commerce,and created huge economic benefits for the enterprises.Providing items that conform with users’ interests is the basic task of the recommendation algorithm.Users’ interests,however,are usually varied and in a drift,owning to the changes in recommendation scenarios and users’ requirements.Therefore,the dilemma of accuracy and diversity has been a serious challenge with which the collaborative filtering is confronted at present.This paper focuses on this dilemma in the recommendation,and some effective solutions based on the theories of complex network and method of multi-objective optimization were proposed and verified,under the framework of collaborative filtering.Nearest neighbors-based recommendation algorithm tends to concentrate on the popular items and ignore the long tail ones.A recommendation strategy based on the nearest neighbors and second-order neighbors was proposed.The strategy selects the second-order neighbors of the target user in the neighborhood of his nearest neighbors,then the weights of the nearest neighbors are adjusted by introducing the adjustable parameters,finally both the nearest neighbors and second-order ones are used for recommendation.Experiments on several benchmark datasets demonstrated that,when the parameter ranges within a specific interval,the algorithm based on the proposed strategy could maintain high enough accuracy and significantly increase the diversity and coverage.Further,the algorithm could perform better on dataset with lower sparsity.The mainstream similarity methods focus on the intersection of users’ items,but fail to fully describe the relevance of the user’s interest preferences.In this paper,an improved similarity method was designed.This method divides similarity between users into historical similarity and tendency one,and calculate the historical similarity on the overlapping items and tendency similarity on non-overlapping items simultaneously.Experiments on several benchmark datasets showed that,compared with other two improved algorithms,the algorithm based on our improved similarity has the least loss on the accuracy.Meanwhile,this algorithm could raise the recommendation coverage and diversity by bigger percentages at the same loss on accuracy.In neighborhood-based recommendations,there is serious redundant item information among the nearest neighbors,which magnifies the weight of minority item and limits the recommendation coverage.In this paper,a bi-objective optimization model considering accuracy and diversity was established for arbitrary user.Then,a multi-objective optimization algorithm based on decomposition was used to solve the problem,in order to find the best neighbors that can represent the multi-dimension interest preference of the user as much as possible.Different from the collaborative filtering,the proposed algorithm could obtain multiple optimal neighbors who could all generate recommendation for the target user.Experiments on benchmark datasets showed that,the proposed algorithm obviously improves the diversity and coverage of recommendation results under the premise that the accuracy is similar or even improved.Accuracy and diversity dilemma is one of the major challenges that the recommender system encounters in today’s era.In this paper,we analyzed several causes of this dilemma and came up with specific solutions.Finally,experiments on several benchmark datasets showed that while maintaining high enough accuracy and even increasing accuracy to a certain extent,the proposed algorithm significantly improves the diversity and coverage of recommendation results.Meanwhile,this paper also provides a new research perspective on the accuracy-diversity problem.更多还原