基于交互的协同过滤算法研究

发布时间：2018-03-30 04:20

本文选题：推荐系统　切入点：协同过滤　出处：《西南石油大学》2017年硕士论文

【摘要】：互联网的出现和普及推动了电子商务行业的崛起,然而,网络数据的急剧膨胀使人们难以获得有用的信息和服务。推荐系统致力于解决信息过载问题,帮助用户找到所感兴趣的资源,因此已成为当前研究的热点。协同过滤是推荐系统的常用核心算法,其基本思想是根据用户对产品的评分数据,找到相似用户或产品(称为邻居),最终借鉴邻居的偏好进行推荐。然而,已有协同过滤算法在获得邻居时多使用传统的相似度指标,且很少考虑到用户与推荐系统之间的交互,影响了预测精度和推荐效果。因此,研究在用户-推荐系统交互场景中的,利用新颖的相似度指标来进行推荐具有较强的现实意义。本文提出一种基于用户-推荐系统交互场景的协同过滤算法,提供用户感兴趣的个性化推荐。第一,提出了近邻相似度计算指标Triangle,并结合传统的Jaccard和Cosine相似度,定义了两种JCT指标JCT_A和JCT_M。Jaccard可以衡量两个样本集合的相似性,它被定义为两个集合的交集和他们的并集元素个数的比率,其值越高意味着邻居共同评价过的项目越多,他们的评分相似性也就越可靠;Cosine用两个评分向量的夹角余弦值来衡量邻居对不同项目的评分偏好,其值越大,评分向量之间的夹角越小,评分的偏好性越一致;Triangle相似度对评分的绝对数值敏感,其值越大,评分之间的绝对差距越小。JCT_A将这三种相似度相加,而JCT_M则将它们相乘。第二,设计并实现了批量反馈型的用户-推荐系统交互场景。用户随机登录推荐系统后,浏览推荐列表中所有自己感兴趣的项目,并将自己的选择和评分反馈给系统。系统则根据用户的反馈和历史评分信息,为用户提供更准确和多样的推荐。该场景可以高效地为用户一次推荐多个感兴趣的项目,同时又兼顾了信息提供者将自己的资源尽可能多地推荐给用户的需求,使系统具有一定的挖掘长尾物品的能力。第三,通过两组在 MovieLens 100K,MovieLens 1M,Each Movie 和 Dou Ban 四个电影评分数据集上的实验验证算法有效性。第一组实验以平均绝对误差(MAE)和均方根误差(RMSE)为指标,比较PIP、NHSM、JCT__A、JCT_M等七种近邻计算相似度的评分预测精度。第二组实验以召回率(recall)、准确率(precision)和覆盖率(coverage)为指标,首先比较了单一反馈型和批量反馈型的交互场景的TopN推荐效果,然后比较了 Cosine、Pearson、JCTA和JCTM四种近邻计算相似度在批量反馈型的交互场景的TopN推荐效果。实验表明,在评分预测方面,在MovieLens 100K,MovieLens 1M和Dou Ban数据集上JCT_M的MAE和RMSE值均低于其他相似度指标;在Each Movie数据集上,JCT_A取得MAE和RMSE最小值。在TopN推荐方面,在同一种相似度下,批量反馈型的交互场景比单一反馈型的交互场景能得到更高的recall、precision和coverage;JCT_A在Top N推荐上性能优于其他相似度。
[Abstract]:The emergence and popularity of the Internet has promoted the rise of e-commerce industry. However, the rapid expansion of network data makes it difficult for people to obtain useful information and services.Recommendation system is dedicated to solve the problem of information overload and help users find interesting resources, so it has become a hot research topic.Collaborative filtering is the core algorithm of recommendation system. The basic idea of collaborative filtering is to find similar users or products (called neighbors or products) according to the users' scoring data.However, the existing collaborative filtering algorithms often use the traditional similarity index in obtaining neighbors, and seldom consider the interaction between users and recommendation systems, which affects the prediction accuracy and recommendation effect.Therefore, it is of great practical significance to study the application of novel similarity index in the user-recommendation system interaction scenario.In this paper, a collaborative filtering algorithm based on user-recommendation system interaction scenario is proposed to provide personalized recommendation of user interest.First, the nearest neighbor similarity index Triangle. combining with the traditional Jaccard and Cosine similarity, two JCT indexes, JCT_A and JCT_M.Jaccard, are defined to measure the similarity between the two sets of samples.It is defined as the ratio of the intersection of two sets to the number of elements of their union, and the higher its value is, the more items neighbors evaluate together,The more reliable their score similarity is, the more reliable Cosine is in measuring neighbors' preferences for different items using the cosine value of the angle between the two score vectors, and the larger the value, the smaller the angle between the score vectors.The more consistent the preference of the score is, the more sensitive the similarity is to the absolute value of the score, and the greater the value, the smaller the absolute difference between the scores. JCTA adds these three similarity degrees and JCT_M multiplies them.Secondly, a batch feedback user-recommendation system interaction scenario is designed and implemented.After the user logs into the recommendation system at random, browse all the items in the recommendation list that are of interest to them, and feedback their selection and rating to the system.The system provides users with more accurate and diverse recommendations based on user feedback and historical scoring information.This scenario can efficiently recommend multiple items of interest to the user at a time, and at the same time, it also takes into account the requirement of the information provider to recommend his own resources to the user as much as possible, so that the system has a certain ability to mine long-tailed items.Thirdly, the validity of the algorithm is verified by two groups of experiments on the four sets of MovieLens 100K movieLens 1Me ach Movie and Dou Ban.In the first group of experiments, the mean absolute error (mae) and the root mean square error (RMSE) were used as the indexes to compare the prediction accuracy of PIPNHSM / JCTS / JCTSP / AJ / JCTSP / AJ / JCTM / AJCTM.In the second group of experiments, the TopN recommendation effect of single feedback interactive scenario and batch feedback interactive scene was compared with the recall rate, accuracy rate and coverage coverage.Then we compare the TopN recommendation effect between Cosine Pearsonian JCTA and JCTM in batch feedback interactive scenarios.The experimental results show that the MAE and RMSE values of JCT_M on MovieLens 100K Movie Lens 1m and Dou Ban datasets are lower than those of other similarity indexes, and the minimum MAE and RMSE values are obtained on Each Movie datasets.In the aspect of TopN recommendation, under the same similarity, the performance of batch feedback interaction scene is higher than that of single feedback interaction scenario, and the performance of covering JCTA is better than that of Top N recommendation.
【学位授予单位】：西南石油大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】