基于深度自动编码器的托攻击集成检测方法

在采用协同过滤技术的推荐系统中,恶意用户通过注入大量虚假概貌使系统的推荐结果产生偏离,达到其攻击目的。为了检测托攻击,根据用户的评分值或基于攻击时间的集中性假设,从不同视角提取攻击概貌的特征。但是,这些基于人工特征的检测方法严重依赖于特征工程的质量,而且人工提取的检测特征多限于特定类型的攻击,提取特征也需要较高的知识成本。针对这些问题,从用户评分项目的时间偏好信息入手,提出一种利用深度稀疏自动编码器自动提取检测特征的托攻击集成检测方法。利用小波变换将项目在不同时间间隔内的流行度设定为多个等级,对用户的评分数据预处理得到用户-项目时间流行度等级矩阵。然后,采用深度稀疏自动编码器对用户-项目时间流行度等级矩阵自动进行特征提取,得到用户评分模式的低层特征表达,消除了传统的人工特征工程。以SVM作为基分类器,在深度稀疏自动编码器的每层提取特征并进行攻击检测,生成最终的集成检测结果。在Netflix数据集上的实验表明,提出的检测方法对均值攻击、AoP攻击、偏移攻击、高级项目攻击、高级用户攻击具有较好的检测效果。     

一种基于流行度分类特征的托攻击检测算法

基于协同过滤的推荐系统容易受到托攻击的危害, 如何检测托攻击成为推荐系统可靠性的关键. 针对现有托攻击检测手段使用基于评分的分类特征易受混淆技术干扰的局限, 本文从用户选择评分项目方式入手, 分析由此造成的用户概貌中已评分项目的流行度分布情况的不同, 提出用于区分正常用户与虚假用户基于流行度的分类特征, 进而得到基于流行度的托攻击检测算法. 实验表明该算法在托攻击检测中具有更强的检测性能与抗干扰性.     

Noisy but non-malicious user detection in social recommender systems

Social recommender systems largely rely on user-contributed data to infer users’ preference. While this feature has enabled many interesting applications in social networking services, it also introduces unreliability to recommenders as users are allowed to insert data freely. Although detecting malicious attacks from social spammers has been studied for years, little work was done for detecting Noisy but Non-Malicious Users (NNMUs), which refers to those genuine users who may provide some untruthful data due to their imperfect behaviors. Unlike colluded malicious attacks that can be detected by finding similarly-behaved user profiles, NNMUs are more difficult to identify since their profiles are neither similar nor correlated from one another. In this article, we study how to detect NNMUs in social recommender systems. Based on the assumption that the ratings provided by a same user on closely correlated items should have similar scores, we propose an effective method for NNMU detection by capturing and accumulating user’s “self-contradictions”, i.e., the cases that a user provides very different rating scores on closely correlated items. We show that self-contradiction capturing can be formulated as a constrained quadratic optimization problem w.r.t. a set of slack variables, which can be further used to quantify the underlying noise in each test user profile. We adopt three real-world data sets to empirically test the proposed method. The experimental results show that our method (i) is effective in real-world NNMU detection scenarios, (ii) can significantly outperform other noisy-user detection methods, and (iii) can improve recommendation performance for other users after removing detected NNMUs from the recommender system.     

基于目标项目识别的用户概貌攻击检测算法

为了准确检测协同过滤推荐系统中的用户概貌注入攻击,从攻击的目的性考虑,通过对攻击概貌中目标项目与填充项目之间的特征差异进行分析,提出一种项目评分背离度的计算方法;利用项目评分背离度来确定受攻击的目标项目,并在此基础上提出一种基于目标项目识别的用户概貌注入攻击检测算法.实验结果表明,该算法进一步提高了攻击检测的精度,确保了系统推荐的质量.     

基于特征指标推荐系统托攻击半监督检测*

推荐系统托攻击检测算法监督学习过度依赖训练集,无监督算法依赖于攻击概貌之间相似性。本文提出一种半监督托攻击检测模型,对标记用户分类计算簇中心,给出中心用户相似度特征属性。对不同攻击选择合适的特征指标,把输入用户划分到不同的簇集中,通过簇集中输入用户全部评分项为最大值的均值与标记用户对该项均值差,确定攻击项。依据特征指标对不同簇集进行两次分类,进而确定攻击对象。实验证明,该检测算法对不同的托攻击有较高的检测准确率。     

基于属性优化矩阵补全的抗托攻击推荐算法

托攻击是当前推荐系统面临的严峻挑战之一。由于推荐系统的开放性,恶意用户可轻易对其注入精心设计的评分从而影响推荐结果,降低用户体验。基于属性优化结构化噪声矩阵补全技术,提出一种鲁棒的抗托攻击个性化推荐（SATPR）算法,将攻击评分视为评分矩阵中的结构化行噪声并采用L2,1范数进行噪声建模,同时引入用户与物品的属性特征以提高托攻击检测精度。实验表明,SATPR算法在托攻击下可取得比传统推荐算法更精确的个性化评分预测效果。     

引入兴趣稳定性的时间敏感协同过滤算法

针对物品流行偏置现象,将物品流行度引入到用户兴趣中建模,提出了基于物品流行度的用户兴趣特征相似度模型。针对传统模型没有考虑到用户兴趣稳定性和难以实时捕获用户兴趣问题,在计算用户兴趣相似度过程中引入时间敏感和兴趣稳定性,提出了引入时间敏感的用户兴趣稳定性的相似度模型。最后,融合权重因子将两种相似度模型线性加权,提出了一种新颖的IPSTS算法模型。实验表明,该算法相比传统协同过滤算法在平均绝对误差（MAE）和均方根误差（RMSE）上均有明显降低的同时也能挖掘出长尾物品,缓解了物品流行偏置现象。     

Integrating similarity and dissimilarity notions in recommenders

Collaborative recommenders rely on the assumption that similar users may exhibit similar tastes while content-based ones favour items that found to be similar with the items a user likes. Weak related entities, which are often considered to be useful, are neglected by those similarity-driven recommenders. To take advantage of this neglected information, we introduce a novel dissimilarity-based recommender that bases its estimations on degrees of dissimilarities among items’ attributes. However, instead of using the proposed recommender as a stand-alone method, we combine it with similarity-based ones to maintain the selective nature of the latter while detecting, through our recommender, information that may have been overlooked. Such combinations are established by IANOS, a proposed framework through which we increase the accuracy of two popular similarity-based recommenders (Naive Bayes and Slope-One) after their combination with our algorithm. Improved accuracy results in experimentation on two datasets (Yahoo! Movies and Movielens) enhance our reasoning. However, the proposed recommender comes with an additional computational complexity when combined with other techniques. By using Hadoop technology, we developed a distributed version of IANOS through which execution time was reduced. Evaluation on IANOS procedures in terms of time performance endorses the use of distributed implementations.     

Rating prediction for recommendation: Constructing user profiles and item characteristics using backpropagation

Recommender systems play an important role in quickly identifying and recommending most acceptable products to the users. The latent user factors and item characteristics determine the degree of user satisfaction on an item. While many of the methods in the literature have assumed that these factors are linear, there are some other methods that treat these factors as nonlinear; but they do it in a more implicit way. In this paper, we have investigated the effect of true nature (i.e., nonlinearity) of the user factors and item characteristics, and their complex layered relationship on rating prediction. We propose a new deep feedforward network that learns both the factors and their complex relationship concurrently. The aim of our study was to automate the construction of user profiles and item characteristics without using any demographic information and then use these constructed features to predict the degree of acceptability of an item to a user. We constructed the user and item factors by using separate learner weights at the lower layers, and modeled their complex relationship in the upper layers. The construction of the user profiles and the item characteristics, solely based on rating triples (i.e., user id, item id, rating), overcomes the requirement of explicit demographic information be given to the system. We have tested our model on three real world datasets: Jester, Movielens, and Yahoo music. Our model produces better rating predictions than some of the state-of-the-art methods which use demographic information. The root mean squared error incurred by our model on these datasets are 4.0873, 0.8110, and 0.9408 respectively. The errors are smaller than current best existing models’ errors in these datasets. The results show that our system can be integrated to any web store where development of hand engineered features for recommending products is less feasible due to huge traffics and also that there is a lack of demographic information about the users and the items.     

基于项目分类和用户群体兴趣的协同过滤算法

由于用户评分数据在极端稀疏的情况下会导致传统协同过滤算法的推荐质量下降,针对该问题,提出一种基于项目分类和用户群体兴趣的协同过滤算法。该算法根据项目类别信息对项目进行分类,相同分类的项目具有较高的相似性;利用评分数据计算各个项目分类上的用户相似性矩阵,并计算用户群体在各个分类上的兴趣,通过二者构造加权的用户相似性矩阵;利用用户加权相似性矩阵寻找用户的最近邻以获得最佳的推荐效果。实验结果表明,该算法能有效提高推荐质量。     

Collaborative filtering and deep learning based recommendation system for cold start items

Recommender system is a specific type of intelligent systems, which exploits historical user ratings on items and/or auxiliary information to make recommendations on items to the users. It plays a critical role in a wide range of online shopping, e-commercial services and social networking applications. Collaborative filtering (CF) is the most popular approaches used for recommender systems, but it suffers from complete cold start (CCS) problem where no rating record are available and incomplete cold start (ICS) problem where only a small number of rating records are available for some new items or users in the system. In this paper, we propose two recommendation models to solve the CCS and ICS problems for new items, which are based on a framework of tightly coupled CF approach and deep learning neural network. A specific deep neural network SADE is used to extract the content features of the items. The state of the art CF model, timeSVD++, which models and utilizes temporal dynamics of user preferences and item features, is modified to take the content features into prediction of ratings for cold start items. Extensive experiments on a large Netflix rating dataset of movies are performed, which show that our proposed recommendation models largely outperform the baseline models for rating prediction of cold start items. The two proposed recommendation models are also evaluated and compared on ICS items, and a flexible scheme of model retraining and switching is proposed to deal with the transition of items from cold start to non-cold start status. The experiment results on Netflix movie recommendation show the tight coupling of CF approach and deep learning neural network is feasible and very effective for cold start item recommendation. The design is general and can be applied to many other recommender systems for online shopping and social networking applications. The solution of cold start item problem can largely improve user experience and trust of recommender systems, and effectively promote cold start items.     

Discerning individual interests and shared interests for social user profiling

Traditionally, research about social user profiling assumes that users share some similar interests with their followees. However, it lacks the studies on what topic and to what extent their interests are similar. Our study in online sharing sites reveals that besides shared interests between followers and followees, users do maintain some individual interests which differ from their followees. Thus, for better social user profiling we need to discern individual interests (capturing the uniqueness of users) and shared interests (capturing the commonality of neighboring users) of the users in the connected world. To achieve this, we extend the matrix factorization model by incorporating both individual and shared interests, and also learn the multi-faceted similarities unsupervisedly. The proposed method can be applied to many applications, such as rating prediction, item level social influence maximization and so on. Experimental results on real-world datasets show that our work can be applied to improve the performance of social rating. Also, it can reveal some interesting findings, such as who likes the “controversial” items most, and who is the most influential in attracting their followers to rate an item.     

Modeling and broadening temporal user interest in personalized news recommendation

User profiling is an important step for solving the problem of personalized news recommendation. Traditional user profiling techniques often construct profiles of users based on static historical data accessed by users. However, due to the frequent updating of news repository, it is possible that a user’s fine-grained reading preference would evolve over time while his/her long-term interest remains stable. Therefore, it is imperative to reason on such preference evaluation for user profiling in news recommenders. Besides, in content-based news recommenders, a user’s preference tends to be stable due to the mechanism of selecting similar content-wise news articles with respect to the user’s profile. To activate users’ reading motivations, a successful recommender needs to introduce “somewhat novel” articles to users.In this paper, we initially provide an experimental study on the evolution of user interests in real-world news recommender systems, and then propose a novel recommendation approach, in which the long-term and short-term reading preferences of users are seamlessly integrated when recommending news items. Given a hierarchy of newly-published news articles, news groups that a user might prefer are differentiated using the long-term profile, and then in each selected news group, a list of news items are chosen as the recommended candidates based on the short-term user profile. We further propose to select news items from the user–item affinity graph using absorbing random walk model to increase the diversity of the recommended news list. Extensive empirical experiments on a collection of news data obtained from various popular news websites demonstrate the effectiveness of our method.     

基于仿生模式识别的未知推荐攻击检测

针对已有检测方法不能有效地检测未知推荐攻击的问题,提出了一种基于仿生模式识别(bionic pattern recognition)的检测方法。首先,依据项目流行度划分项目到不同的窗口,把用户对窗口内项目的评分视为随机事件发生。在此基础上,利用信息熵(information entropy)提取评分分布特征作为检测推荐攻击的通用特征。然后,在特征空间中,利用仿生模式识别技术覆盖真实概貌样本,将覆盖范围外的测试数据判为推荐攻击。在MovieLens数据集上进行实验,结果表明,该方法在检测未知推荐攻击时具有较高的命中率和较低的误报率。     

Distributional semantic pre-filtering in context-aware recommender systems

Context-aware recommender systems improve context-free recommenders by exploiting the knowledge of the contextual situation under which a user experienced and rated an item. They use data sets of contextually-tagged ratings to predict how the target user would evaluate (rate) an item in a given contextual situation, with the ultimate goal to recommend the items with the best estimated ratings. This paper describes and evaluates a pre-filtering approach to context-aware recommendation, called distributional-semantics pre-filtering (DSPF), which exploits in a novel way the distributional semantics of contextual conditions to build more precise context-aware rating prediction models. In DSPF, given a target contextual situation (of a target user), a matrix-factorization predictive model is built by using the ratings tagged with the contextual situations most similar to the target one. Then, this model is used to compute rating predictions and identify recommendations for that specific target contextual situation. In the proposed approach, the definition of the similarity of contextual situations is based on the distributional semantics of their composing conditions: situations are similar if they influence the user’s ratings in a similar way. This notion of similarity has the advantage of being directly derived from the rating data; hence it does not require a context taxonomy. We analyze the effectiveness of DSPF varying the specific method used to compute the situation-to-situation similarity. We also show how DSPF can be further improved by using clustering techniques. Finally, we evaluate DSPF on several contextually-tagged data sets and demonstrate that it outperforms state-of-the-art context-aware approaches.     

基于用户评分和评论信息的协同推荐框架*

用户的反馈通常包含一个数值评分和一段文本形式的评论.文中利用用户评论学习商品特征在不同主题上的分布及用户对商品不同特征的偏好程度,把商品特征和用户偏好的契合度引入传统的协同过滤算法中,提出基于用户评分和评论信息的协同推荐框架.使用该框架可较方便地将用户评论信息引入到现有的协同过滤算法中.通过引入用户评论信息,可一定程度缓解传统协同过滤算法面临的数据稀疏性问题.在22个亚马逊的真实数据集上的实验证明文中方法的有效性.     

Personalized recommendation based on review topics

The traditional collaborative filtering algorithm is a successful recommendation technology. The core idea of this algorithm is to calculate user or item similarity based on user ratings and then to predict ratings and recommend items based on similar users’ or similar items’ ratings. However, real applications face a problem of data sparsity because most users provide only a few ratings, such that the traditional collaborative filtering algorithm cannot produce satisfactory results. This paper proposes a new topic model-based similarity and two recommendation algorithms: user-based collaborative filtering with topic model algorithm (UCFTM, in this paper) and item-based collaborative filtering with topic model algorithm (ICFTM, in this paper). Each review is processed using the topic model to generate review topic allocations representing a user’s preference for a product’s different features. The UCFTM algorithm aggregates all topic allocations of reviews by the same user and calculates the user most valued features representing product features that the user most values. User similarity is calculated based on user most valued features, whereas ratings are predicted from similar users’ ratings. The ICFTM algorithm aggregates all topic allocations of reviews for the same product, and item most valued features representing the most valued features of the product are calculated. Item similarity is calculated based on item most valued features, whereas ratings are predicted from similar items’ ratings. Experiments on six data sets from Amazon indicate that when most users give only one review and one rating, our algorithms exhibit better prediction accuracy than other traditional collaborative filtering and state-of-the-art topic model-based recommendation algorithms.     

基于项目之间相似性的兴趣点推荐方法*

针对评分数据稀疏的情况下传统相似性计算的不足,提出了一种基于项目之间相似性的协同过滤算法。该算法结合用户对项目的评分和项目之间的兴趣度进行项目之间的相似性计算,在一定程度上减小了评分数据稀疏的负面影响。实验结果表明,该算法在评分数据稀疏的情况下,能使推荐系统的推荐质量明显提高。     

自适应用户的Item-based协同过滤推荐算法

传统Item-based协同过滤算法计算两个条目间相似性时, 将每个评分视为同等重要, 忽略了共评用户（对两个条目共同评分的用户）与目标用户间的相似性对条目间相似性的影响。针对此问题, 提出了一种自适应用户的Item-based协同过滤算法。该算法将共评用户与目标用户的相似性作为共评用户评分重要性的权重, 以实现针对不同的目标用户, 为目标条目选择不同的、适合目标用户的最近邻居集, 从而提高推荐准确性。实验结果表明, 提出的算法可以显著提高推荐系统的推荐质量。