Evaluation of session-based recommendation algorithms

Recommender systems help users find relevant items of interest, for example on e-commerce or media streaming sites. Most academic research is concerned with approaches that personalize the recommendations according to long-term user profiles. In many real-world applications, however, such long-term profiles often do not exist and recommendations therefore have to be made solely based on the observed behavior of a user during an ongoing session. Given the high practical relevance of the problem, an increased interest in this problem can be observed in recent years, leading to a number of proposals for session-based recommendation algorithms that typically aim to predict the user’s immediate next actions. In this work, we present the results of an in-depth performance comparison of a number of such algorithms, using a variety of datasets and evaluation measures. Our comparison includes the most recent approaches based on recurrent neural networks like gru4rec, factorized Markov model approaches such as fism or fossil, as well as simpler methods based, e.g., on nearest neighbor schemes. Our experiments reveal that algorithms of this latter class, despite their sometimes almost trivial nature, often perform equally well or significantly better than today’s more complex approaches based on deep neural networks. Our results therefore suggest that there is substantial room for improvement regarding the development of more sophisticated session-based recommendation algorithms.     

The effects of recommendations’ presentation on persuasion and satisfaction in a movie recommender system

The explosive growth of Internet applications and content, during the last decade, has revealed an increasing need for information filtering and recommendation. Most research in the area of recommendation systems has focused on designing and implementing efficient algorithms that provide accurate recommendations. However, the selection of appropriate recommendation content and the presentation of information are equally important in creating successful recommender applications. This paper addresses issues related to the presentation of recommendations in the movies domain. The current work reviews previous research approaches and popular recommender systems, and focuses on user persuasion and satisfaction. In our experiments, we compare different presentation methods in terms of recommendations’ organization in a list (i.e. top N-items list and structured overview) and recommendation modality (i.e. simple text, combination of text and image, and combination of text and video). The most efficient presentation methods, regarding user persuasion and satisfaction, proved to be the “structured overview” and the “text and video” interfaces, while a strong positive correlation was also found between user satisfaction and persuasion in all experimental conditions.     

Neural Explainable Recommender Model Based on Attributes and Reviews

Explainable recommendation, which can provide reasonable explanations for recommendations, is increasingly important in many fields. Although traditional embedding-based models can learn many implicit features, resulting in good performance, they cannot provide the reason for their recommendations. Existing explainable recommender methods can be mainly divided into two types. The first type models highlight reviews written by users to provide an explanation. For the second type, attribute information is taken into consideration. These approaches only consider one aspect and do not make the best use of the existing information. In this paper, we propose a novel neural explainable recommender model based on attributes and reviews (NERAR) for recommendation that combines the processing of attribute features and review features. We employ a tree-based model to extract and learn attribute features from auxiliary information, and then we use a time-aware gated recurrent unit (T-GRU) to model user review features and process item review features based on a convolutional neural network (CNN). Extensive experiments on Amazon datasets demonstrate that our model outperforms the state-of-the-art recommendation models in accuracy of recommendations. The presented examples also show that our model can offer more reasonable explanations. Crowd-sourcing based evaluations are conducted to verify our model’s superiority in explainability.

     

融合全局信息的注意力增强会话推荐方法

近年来,基于会话推荐系统（session-based recommender system,SRS）的应用和研究是推荐系统的一个热门方向。如何利用用户会话信息进一步提升用户满意度和推荐精确度,是基于会话推荐系统的主要任务。目前大多数SBR模型仅基于目标会话对用户偏好建模,忽略了来自其他会话的物品转换信息,导致无法全面了解用户偏好。为了解决其局限性,提出融合全局上下文信息注意力增强的图神经网络模型（global context information graph neural networks for session-based recommendation,GCI-GNN）。该模型利用所有会话上的物品转换关系,更准确地获取用户偏好。具体而言,GCI-GNN从目标会话和全局会话学习物品向量表示。使用位置感知注意网络,将反向位置信息纳入物品嵌入中。考虑会话长度信息学习用户表示进而达到更有效的推荐。在Diginetica和Yoochoose数据集上进行实验,实验结果表明,相对最优的基准模型,GCI-GNN模型在Diginetica数据集各项指标上的提高超过2个百分点,在Yoochoose数据...     

基于邻域感知图神经网络的会话推荐

在基于会话的推荐中,图神经网络及其改进模型将会话内复杂的交互关系建模为图结构并从中捕获项目特征,是现有推荐模型中性能较好的一类方法。然而大多数模型都忽略了不同会话之间可能存在的有效信息,仅对当前会话建模难以利用其他会话,也无法发挥邻域信息的辅助作用。因此提出基于邻域感知图神经网络的会话推荐（NA-GNN）。该模型构建会话层和全局邻域层的图结构捕获项目表示,结合注意力机制聚合两种项目表征,将会话序列之间的互信息最大化地结合到网络训练中。在真实的数据集Yoochoose和Diginetica上进行实验,与性能最优的基准模型相比,模型P@20在Yoochoose上提高了1.85%,在Diginetica上提升了7.19%;MRR@20分别提升了0.48%和8.36%,证明模型的有效性和合理性。     

A novel application recommendation method combining social relationship and trust relationship for future internet of things

Traditional collaborative filtering methods always utilize Cosine and Pearson methods to calculate the similarity of users. When the nearest neighbor doesn’t comment the predicted item, then the nearest neighbor has no influence on results, thus affecting the accuracy of collaborative filtering recommendation. And the traditional recommendation systems always have the problems of data sparsity, cold start and so on. In this paper, we consider social relationship and trust relationship, and put forward a novel application recommendation method that combines users’ social relationship and trust relationship. Specifically, we combine social relationship and user preference towards applications to calculate similarity score, we fuse the trust relationship based on familiarity and user reputation to calculate trust score. The final prediction score is calculated by fusing similar relationship and trust relationship properly. And the proposed method can effectively improve accuracy of recommendations.

     

Neural TV program recommendation with label and user dual attention

TV program recommendation is very important for users to find interesting TV programs and avoid confusing users with a lot of information. Currently, they are basically traditional collaborative filtering algorithms, which only recommend through the interactive data between users and programs ignoring the important value of some auxiliary information. In addition, the neural network method based on attention mechanism can well capture the relationship between program labels to obtain accurate program and user representations. In this paper, we propose a neural TV program recommendation with label and user dual attention (NPR-LUA), which can focus on auxiliary information in program and user modules. In the program encoder module, we learn the auxiliary information from program labels through neural networks and word attention to identify important program labels. In the user encoder module, we learn the user representation through the programs that the user watches and use personalized attention mechanism to distinguish the importance of programs for each user. Experiments on real data sets show that our method can effectively improve the effectiveness of TV program recommendations than other existing models.

     

Interactive recommender systems: A survey of the state of the art and future research challenges and opportunities

Recommender systems have been researched extensively over the past decades. Whereas several algorithms have been developed and deployed in various application domains, recent research efforts are increasingly oriented towards the user experience of recommender systems. This research goes beyond accuracy of recommendation algorithms and focuses on various human factors that affect acceptance of recommendations, such as user satisfaction, trust, transparency and sense of control. In this paper, we present an interactive visualization framework that combines recommendation with visualization techniques to support human-recommender interaction. Then, we analyze existing interactive recommender systems along the dimensions of our framework, including our work. Based on our survey results, we present future research challenges and opportunities.     

融合用户会话数据的上下文感知推荐算法

针对现有会话推荐算法未充分考虑用户的上下文信息的现状,为增强基于会话的推荐算法的个性化推荐效果,提出一种融合用户会话数据的上下文感知推荐算法。将上下文信息通过embedding映射成低维实数向量特征,通过Add、Stack、MLP三种组合方式将低维向量特征融入到基于会话的循环神经网络推荐模型,设计了基于BPR的损失函数动态刻画会话点击序列的用户偏好,以提升个性化推荐能力。在Adressa数据集上的实验表明,所提算法相比基线算法GRU4REC,在指标Recall@20上提高了3.2%,MRR@20上提高了27%。     

结合自监督学习的图神经网络会话推荐

基于会话的推荐方法由于短期用户交互数据有限,与传统推荐方法相比,其性能更容易受到数据稀疏性问题的影响。为增强会话数据以缓解数据稀疏对会话推荐性能的影响,提出一种结合自监督学习的图神经网络会话推荐（Ss-GNN）模型。构建会话图并建立基于图注意力网络的会话推荐任务来获取项目级表示和会话级表示;从会话级表示的角度出发,利用用户的一般兴趣和当前兴趣来构建辅助任务获取自监督信号;利用自监督学习实现推荐任务和辅助任务之间的互信息最大化,以增强会话数据,从而提升推荐性能。在Yoochoose和Tmall两个公开数据集上进行实验,与基线模型相比,提出的模型在Yoochoose上P@20和MRR@20至少提升了0.94%和0.79%,在Tmall上P@20和MRR@20至少提升了9.61%和4.67%,证明了Ss-GNN模型的有效性。     

Combining long-term and short-term user interest for personalized hashtag recommendation

Hashtags, terms prefixed by a hash-symbol #, are widely used and inserted anywhere within short messages (tweets) on micro-blogging systems as they present rich sentiment information on topics that people are interested in. In this paper, we focus on the problem of hashtag recommendation considering their personalized and temporal aspects. As far as we know, this is the first work addressing this issue specially to recommend personalized hashtags combining longterm and short-term user interest.We introduce three features to capture personal and temporal user interest: 1) hashtag textual information; 2) user behavior; and 3) time. We offer two recommendation models for comparison: a linearcombined model, and an enhanced session-based temporal graph (STG) model, Topic-STG, considering the features to learn user preferences and subsequently recommend personalized hashtags. Experiments on two real tweet datasets illustrate the effectiveness of the proposed models and algorithms.     

Explainable recommendation with fusion of aspect information

Explainable recommendation has attracted increasing attention from researchers. The existing methods, however, often suffer from two defects. One is the lack of quantitative fine-grained explanations why a user chooses an item, which likely makes recommendations unconvincing. The other one is that the fine-grained information such as aspects of item is not effectively utilized for making recommendations. In this paper, we investigate the problem of making quantitatively explainable recommendation at aspect level. It is a nontrivial task due to the challenges on quantitative evaluation of aspect and fusing aspect information into recommendation. To address these challenges, we propose an Aspect-based Matrix Factorization model (AMF), which is able to improve the accuracy of rating prediction by collaboratively decomposing the rating matrix with the auxiliary information extracted from aspects. To quantitatively evaluate aspects, we propose two metrics: User Aspect Preference (UAP) and Item Aspect Quality (IAQ), which quantify user preference to a specific aspect and the review sentiment of item on an aspect, respectively. By UAP and IAQ, we can quantitatively explain why a user chooses an item. To achieve information incorporation, we assemble UAPs and IAQs into two matrices UAP Matrix (UAPM) and IAQ Matrix (IAQM), respectively, and fuse UAPM and IAQM as constraints into the collaborative decomposition of item rating matrix. The extensive experiments conducted on real datasets verify the recommendation performance and explanatory ability of our approach.

     

基于时间效应的协同过滤算法

协同过滤算法已被成功应用在个性化推荐系统中,但传统的协同过滤算法很少考虑时间因素的影响,难以确保最近邻集的准确性和可靠性。虽然很多文献提出了各种改进推荐算法,但仍然没能在计算中有效地将时间因素和用户评分综合起来。因此,在原有的工作基础上提出基于时间效应的协同过滤算法,将时间因素纳入用户预测评分和用户相似性计算中,并综合这两个因素来动态分配每一项评分的权重,采用预测评分填充用户-项矩阵和二次计算用户相似性矩阵的方法,最终得到Top-N推荐集。实验表明,改进后的算法提高了推荐算法的精度和推荐质量。     

基于复杂结构信息的图神经网络序列推荐算法

图结构因其在序列推荐场景中的自然适应性而备受关注,而现有的基于图神经网络的会话序列推荐算法虽然能够利用图结构信息达到较好的推荐效果,但是没有考虑用户在会话序列中的重复点击行为和项目之间的复杂转换,且未很好地利用图中复杂的结构信息,导致推荐的效果受到一定程度的限制。提出有向与无向信息同注意力相融合的图神经网络序列推荐算法,并基于推荐算法给出项目隐含向量建模算法,结合会话序列图中的有向结构信息与无向结构信息,通过考虑用户的重复点击行为和引入注意力机制建立会话中点击项目的复杂转换模型。图节点在特征传播的过程中平衡邻居节点信息与自身信息的比例,以更准确地预测推荐过程中生成的会话向量。在Diginetica、Yoochoose 1/64、Yoochoose 1/4 3个数据集上的实验结果表明,与SR-GNN、TAGNN算法相比,该算法精度最高提升4.34%,能够更好地预测用户在会话中的下一次点击精度。     

Modeling Temporal Dynamics of Users’ Purchase Behaviors for Next Basket Prediction

Next basket prediction attempts to provide sequential recommendations to users based on a sequence of the user’s previous purchases. Ideally, a good prediction model should be able to explore the personalized preference of the users, as well as the sequential relations of the items. This goal of modeling becomes even more challenging when both factors are time-dependent. However, existing methods either take these two aspects as static, or only consider temporal dynamics for one of the two aspects. In this work, we propose the dynamic representation learning approach for time-dependent next basket recommendation, which jointly models the dynamic nature of user preferences and item relations. To do so, we explicitly model the transaction timestamps, as well as the dynamic representations of both users and items, so as to capture the personalized user preference on each individual item dynamically. Experiments on three real-world retail datasets show that our method significantly outperforms several state-of-the-art methods for next basket recommendation.

     

Path-enhanced explainable recommendation with knowledge graphs

Recommender systems, which are used to predict user requirements precisely, play a vital role in the modern internet industry. As an effective tool with rich semantics, knowledge graphs have recently attracted growing research attention in enhancing recommendation results. By mining multihop relations (i.e., paths) between user-item interactions within a knowledge graph, implicit user preferences and other side information can be clearly revealed. Nevertheless, existing knowledge graph-based recommendation methods have two fundamental limitations. First, the indiscriminate utilization of user-item path sets conveys unclear information and negatively influences explainability. Moreover, obtaining reliable recommendation results with these methods requires large amounts of prior knowledge, which indicates that they show poor performance in terms of accuracy and handling cold-start issues. To address these issues, we propose a novel model called the Path-enhanced Recurrent Network (PeRN). Specifically, PeRN integrates a recurrent neural network encoder with a metapath-based entropy encoder to increase explainability and accuracy and reduce cold-start costs. The recurrent network encoder has a strong ability to represent sequential path semantics in a knowledge graph, while the entropy encoder, as an efficient statistical analysis tool, leverages metapath information to differentiate paths in a single user-item interaction. A path extraction algorithm with a bidirectional scheme is also proposed to make PeRN more feasible. The experimental results on two real-world datasets demonstrate our significant improvements with reasonable explanations, promising accuracy and a minimal amount of prior knowledge compared with several state-of-the-art baselines.

     

Diversified recommendation method combining topic model and random walk

As one of the most widely used algorithms in recommendation field, collaborative filtering (CF) predicts the unknown rating of items based on similar neighbors. Although many CF-based recommendation methods have been proposed, there still be room for improvement. Firstly, the data sparsity problem still remains a big challenge for CF algorithms to find similar neighbors. Secondly, there are many redundant similar items in the recommendation list generated by traditional CF algorithms, which cannot meet the user wide interest. Therefore, we propose a diversified recommendation method combining topic model and random walk. A weighted random walk model is presented to find all direct and indirect similar neighbors on the sparse data, improving the accuracy of rating prediction. By taking both users’ behavior data and items’ lags into account, we give a diversity measurement method based on the topic distribution of items discovered by Linked-LDA model. Furthermore, a diversified ranking algorithm is developed to balance the accuracy and diversity of recommendation results. We compare our method with six other recommendation methods on a real-world dataset. Experimental results show that our method outperforms the other methods and achieves the best personalized recommendation effect.     

Hybrid recommendation approaches for multi-criteria collaborative filtering

Recommender systems are software tools and techniques for suggesting items in an automated fashion to users tailored their preferences. Collaborative Filtering (CF) techniques, which attempt to predict what information will meet a user’s needs from the neighborhoods of like-minded people, are becoming increasingly popular as ways to overcome the information overload. The multi-criteria based CF presents a possibility to provide accurate recommendations by considering the user preferences in multiple aspects and several methods have been proposed for improving the accuracy of these systems. However, the problem of multi-criteria recommendations with a single and overall rating is still considered an optimization problem. In addition, increasing the accuracy in predicting the appropriate items tailored to the users’ preferences is on of the main challenges in these systems. Hence, in this research new recommendation methods using Adaptive Neuro-Fuzzy Inference Systems (ANFIS) and Self-Organizing Map (SOM) clustering are proposed to improve predictive accuracy of criteria CF. In this research, SOM enables us to generate high quality clusters of dataset and ANFIS is used for discovering knowledge (fuzzy rules) from users’ ratings in multi-criteria dataset, generating appropriate membership functions (MFs), overall rating prediction and input selection. Using exhaustive search method for input selection, the effective inputs are determined to build the ANFIS models in all generated clusters. Furthermore, new fuzzy-based algorithms, Weighted Fuzzy MC-CF (WFuMC-CF), Fuzzy Euclidean MC-CF (FuEucMC-CF) and Fuzzy Average MC-CF (FuAvgMC-CF), are presented for prediction task in multi-criteria CF. FuEucMC-CF and FuAvgMC-CF algorithms uses the fuzzy-based Euclidian distance and fuzzy-based average similarity, respectively, the WFuMC-CF algorithm uses fuzzy-based user- and item-based prediction in a weighted approach. Experimental results on real-world dataset demonstrate that the proposed hybrid methods remarkably improve the accuracy of multi-criteria CF in relation to the previous methods based on multi-criteria ratings.     

CAESAR: context-aware explanation based on supervised attention for service recommendations

Explainable recommendations have drawn more attention from both academia and industry recently, because they can help users better understand recommendations (i.e., why some particular items are recommended), therefore improving the persuasiveness of the recommender system and users’ satisfaction. However, little work has been done to provide explanations from the angle of a user’s contextual situations (e.g., companion, season, and destination if the recommendation is a hotel). To fill this research gap, we propose a new context-aware recommendation algorithm based on supervised attention mechanism (CAESAR), which particularly matches latent features to explicit contextual features as mined from user-generated reviews for producing context-aware explanations. Experimental results on two large datasets in hotel and restaurant service domains demonstrate that our model improves recommendation performance against the state-of-the-art methods and furthermore is able to return feature-level explanations that can adapt to the target user’s current contexts.