Search

Context-aware recommender systems (CARS), which capture user preferences by incorporating user interactions and side information, have attracted widespread attention in today's information age. With the emergence of graph convolution networks (GCN) in recent years, graph methods have been shown effective in establishing global connections between users, items, and contexts. In this paper, we first introduce a unified graph recommendation framework to incorporate both GCN and traditional models, showing the similarities and differences between them in terms of parameter configurations and message passing styles. Then, based on the empirical analysis of existing models, we propose a novel Fused Graph Context-aware Recommender system (FGCR) model to address the identified limitations. FGCR fuses the parameter configuration of both item-based models and GCN, modeling robust node relationships in user-item-context interaction graph through a sparse item correlation matrix and dense node embeddings. In addition, FGCR applies a novel masked graph convolution strategy to refine the information aggregation process with various types of nodes. Experiments conducted on five real-world datasets show that FGCR significantly improves performance compared to the seven baseline models. Ablation study and user group analysis validate the effectiveness of each component in FGCR, especially in modeling active users and sparse contextual information. [ABSTRACT FROM AUTHOR]

Graph Convolutional Networks (GCNs) gain success for recommendation, but still face great challenges of data sparseness and negative sampling in implicit feedback-based recommendation. In particular, they ignore the unique graph structure for information propagation and thus fail to fully explore the potentials of GCN. To tackle the problem, we propose a GCN-based self-training approach, called STL, which exploits the learning results in the training procedure, and the potential relations in the embedding space of GCN. First, to handle the data sparsity, we modify the interaction graph structure by adding edges linking a selected part of users and their potential positive items. Second, we adaptively expand the set of positive samples that are used in the pairwise loss function, which not only supplements the dataset but also avoids sampling noises. Further, similarity of structural neighbors on graph is used to mine hard negative sample for improving the sample quality. Experiments on three representative GCN-based recommenders and four widely used public datasets show that STL alleviates the problem of data sparsity, thereby improving recommendation performance compared to normal training. [ABSTRACT FROM AUTHOR]

Leveraging knowledge graphs (KGs) to enhance recommender systems has gained considerable attention, with researchers obtaining user preferences by aggregating entity pairs with explicit relations in KGs via graph convolutional networks (GCNs). Existing approaches currently overlook many entity pairs without relations, which, however, may have potentially useful information. To address this issue, we propose a novel relation-aware attentional GCN (RAAGCN) with the following improvements over vanilla GCNs: (1) it aggregates all entity pairs with and without explicit relations and (2) it distinguishes the importance of different relational context information. Based on the proposed RAAGCN, we further propose a user preference and item attractiveness capturing model (UPIACM) for KG-based recommendation. In the UPIACM, the user preference is decomposed into interest and rating preferences. The interest preference is the user's interest taste toward the items with specific features, while the rating preference reflects the intention of rating high or low. Additionally, our model accounts for item attractiveness, which reflects an item's popularity among users. Additionally, we incorporate a gated filtering mechanism to further improve our model's performance. Through extensive experiments, we show that the proposed UPIACM outperforms state-of-the-art baseline methods. • Designing relation-aware attentional GCN to aggregate more fine-gained information. • Proposing KG-based UPIACM to capture user preference and item attractiveness. • Integrating gated filtering mechanism to filter out the possible impurity information. • Extensive experiments with popular state-of-the-art baselines on real-world datasets. [ABSTRACT FROM AUTHOR]

Many bipartite networks describe systems where an edge represents a relation between a user and an item. Measuring the similarity between either users or items is the basis of memory-based collaborative filtering, a widely used method to build a recommender system with the purpose of proposing items to users. When the edges of the network are unweighted, the popular common neighbors-based approaches, allowing only positive similarity values, neglect the possibility and the effect of two users (or two items) being very dissimilar. Moreover, they underperform with respect to model-based (machine learning) approaches, although providing a higher interpretability. Inspired by the functioning of Decision Trees, we propose a method to compute similarity that allows also negative values, the Sapling Similarity. The key idea is to look at how the information that a user is connected to an item influences our prior estimation of the probability that another user is connected to the same item: if it is reduced, then the similarity between the two users will be negative, otherwise it will be positive. We show that, when used to build memory-based collaborative filtering, Sapling Similarity provides better recommendations than existing similarity metrics. Then we compare the Sapling Similarity Collaborative Filtering (SSCF, an hybrid of the item-based and the user-based) with state-of-the-art models using standard datasets. Even if SSCF depends on only one straightforward hyperparameter, it has comparable or higher recommending accuracy, and outperforms all other models on the Amazon-Book dataset, while retaining the high explainability of memory-based approaches. [ABSTRACT FROM AUTHOR]

The cold-start problem has been of great concern in the recommendation domain. To address this problem, meta-learning frameworks have been widely adopted due to their fast adaptation to new tasks. However, existing meta-learning-based methods always use higher-order graph structures to obtain global user preferences but neglect to consider local preferences at different rating levels in a fine-grained manner. Since differences in user ratings on items truly reflect differences in users' local preferences, we propose a relation-propagation meta-learning on explicit preference graph for cold-start recommendation (RPMLG-Rec) to improve the generalization performance of existing meta-learning-based methods. Specifically, we focus on capturing the relationships between local preferences. First, our RPMLG-Rec approach concatenates different local preferences to form the nodes of local user preference and further constructs an explicit preference graph. Second, the relationships between local preferences, including intraclass commonality and interclass uniqueness, are used to guide the propagation of relationships in the explicit preference graph with graph convolutional networks and produce more distinguishable local preference nodes. Third, the precise global user preference is obtained with an attention mechanism. Finally, prior knowledge is learned based on a set of training tasks and quickly adapted to make recommendations for new tasks, following the optimization-based meta-learning training strategy. To the best of our knowledge, this is the first time that the relationships between local user preference nodes have been explicitly considered in cold-start recommendation. In addition, we conducted extensive experiments on two real-world datasets, and the experimental results demonstrate the effectiveness of our proposed framework. [ABSTRACT FROM AUTHOR]

Recommender systems (RSs) have been employed for many real-world applications including search engines, social networks, and information retrieval systems as powerful intelligent techniques for creating personalized recommendations. RSs mainly utilize different data resources such as user-item ratings, tag data, and social relationships in their recommendation process to overcome the data sparsity issue. One of the main challenges is how to integrate the results obtained by these data resources. Moreover, most of the existing RSs focus on the recommendation accuracy, and do not consider the reliability of the recommendations. To address these challenges, in this study, a reliable recommendation method is developed, which employs deep neural networks and reinforcement learning. The proposed method firstly uses a stacked denoising autoencoder model to obtain the latent features from the rating, tag, and trust data. The similarity value between each pair of users is then calculated in three different aspects according to the extracted latent features. An ensemble approach is developed by integrating the three similarities using reinforcement learning to determine the final similarity value. The ensemble similarity values are used in the rating prediction process, and the reliability of the predicted ratings is evaluated using an effective measure. Finally, a recommendation strategy is developed based on the integration of the predicted ratings and their reliability values. We evaluated the effectiveness of the proposed method by conducting numerous experiments on three datasets where the results show that the proposed method is far superior to other recommendation models in making more accurate recommendations. [ABSTRACT FROM AUTHOR]

Recommender systems are popularly used to deal with an information overload issue. Existing systems mainly focus on user–item interactions and semantic information derived from metadata of users and items to improve recommendation accuracy. Item images provide useful information to infer users' individual preferences, especially for those domains where visual factors are influential such as fashion items. However, this type of information has been ignored by most previous work. To bridge this gap and meet the requirements of performance from the aspects of Accuracy, Scalability, and Explainability evaluation metrics, this paper proposes a scalable and explainable visually-aware recommender system framework called SEV-RS. This framework contains a visually-augmented heterogeneous information network, a scalable meta-path feature extraction method for multi-hop relations, and a shallow explainable meta-path based Collaborative Filtering recommendation approach. We compared SEV-RS with the state-of-the-art models such as the deep learning model using Graph Attention Network on two real-world datasets and one synthetic dataset. The results show that SEV-RS produced more accurate and more explainable recommendations. Also, SEV-RS has substantially less computational time than the compared deep learning models. [ABSTRACT FROM AUTHOR]

Recently, Generative Adversarial Networks (GANs) have received much attention in recommender systems because they can capture complex data distributions. They rely on two sub-networks, generator and discriminator models, to generate 'fake' yet reliable data, so better recommendation accuracy can be obtained. However, most variants of existing GANs achieve accuracy improvements by greatly increasing the model complexity, especially the amount of parameters, leading to difficulty in the deployment of those methods. Therefore, we aim to resolve this issue by learning compact yet accurate GANs that can make up valuable data with less generator parameters. To this end, we integrate knowledge distillation (KD, in the form of a teacher–student architecture) into GANs to reduce the model complexity while improving the accuracy of GANs. To the best of our knowledge, this work is the first to learn compact GANs for recommender systems by applying a KD framework named StuGAN. Specifically, we use a student discriminator to refine the teacher's knowledge, and then both the generator and discriminator are enhanced by leveraging the refined knowledge via adversarial learning, which constrains the generator to produce fake data approximating both the ground truth and teacher's predicted preferences, and enables the discriminator to distinguish between the preferences of the generator and more confusing preferences of the teacher from the ground truth. Finally, we conduct extensive experiments on two real-world datasets (i.e., Ciao and LastFM), and the results show that our approach can reduce the volume of model parameters by as much as half while maintaining comparable recommendation accuracy. [ABSTRACT FROM AUTHOR]

Data sparsity and cold start are two critical issues which need to be addressed in recommender systems (RSs). Currently, most methods address these issues by applying user history files or some side information to improve the user model and complete the rating matrix. However, such methods cannot perform well when labeled data is scarce or unavailable. In this paper, we propose a dual learning-based recommendation approach (DLRA). DLRA can trigger initial recommendation and improve the quality of recommendations by using the duality characteristics of RSs, even when the available labeled information is scarce. Specifically, DLRA regards the recommendation task as two independent subtasks — primal task and dual task, and these two tasks show strong duality in DLRA. The primal task is item-centered which aims to find users who can rate high for items, while the dual task is user-centered that aims to recommend the most favorite items to users. These two tasks have strong dualities in terms of the recommendation space, selection probability and recommendation basis. Based on these dualities, we design three dual learning strategies to couple the whole recommendation process and realize the self-tuning and self-improvement of each task model, and finally optimize the whole recommendation model. Based on the dataset of Movielens and BookCrossing, we simulate data sparsity and cold start recommendation scenarios, the experimental results show that DLRA achieves substantial improvement when the labeled data is scare, and it outperforms other hybrid recommendation approaches and deep learning strategies with a smaller predictive error as well as better recommendation accuracy. [ABSTRACT FROM AUTHOR]

Recently, contextual multiarmed bandits (CMAB)-based recommendation has shown promise for applications in dynamic domains such as news or short video recommendation, where items are changing over time. There are two key challenges in constructing a CMAB-based system. First, a user's historical rating data are usually sparse, which restricts the precise representation of dynamic contextual information. Second, it is difficult to design a personalized recommendation policy because of the high diversity of users' selection behaviors over time. Therefore, existing CMAB-based recommendation methods mainly focus on the improvement of recommendation accuracy. This leads to difficulty in capturing the dynamic personalized preferences of users. To overcome these limitations, we design a novel knowledge-enhanced contextual-bandit approach that combines the dynamic recommendation strategy of the bandit algorithm and the knowledge representation of the knowledge graph. We argue that a user's selection history for items reflects his/her favorite attributes, which will be repeated in the future. Accordingly we propose the knowledge-enhanced CMAB model, which leverages the contextual information of both users and items obtained by knowledge graph-based embedding. We validate the performance of the proposed model using two public datasets. Experimental results show that our method outperforms state-of-the-art methods in terms of both recommendation accuracy and diversity. • We combine a knowledge graph and the bandit method for constructing personalized, dynamic recommendations. • We propose an intention-selection mechanism for item selection in recommender systems. • We propose the one choice multi change strategy for a contextual multiarmed bandit. • Knowledge graphs effectively address the lack of knowledge in recommender systems. [ABSTRACT FROM AUTHOR]

Traditional recommendation algorithms can be used to develop techniques that can help people choose desirable items of interest. However, in many real-world applications, it is important to quantify each recommendation's (un)certainty, in addition to a set of recommendations. The conformal recommender system uses the experience of a user to output a set of recommendations, each associated with a precise confidence value. A significance level ɛ provides a bound ɛ on the probability of making a wrong recommendation. The conformal framework uses a key concept called nonconformity measure that measures the strangeness of an item concerning other items. One of the significant design challenges of any conformal recommendation framework is the integration of nonconformity measures with the recommendation algorithm. This paper introduces an inductive variant of a conformal recommender system. We propose and analyze different nonconformity measures in the inductive setting. In addition, we provide theoretical proofs on the error bound and time complexity. Extensive empirical analysis on seven benchmark datasets reveals that the inductive variant substantially improves the performance in computation time while preserving the accuracy. • We introduce the concept of inductive conformal recommendation. • We propose and analyze different nonconformity measures in the inductive setting. • We provide theoretical proofs on the error-bound and the time complexity. • Extensive comparative studies validate the efficiency of our algorithm. [ABSTRACT FROM AUTHOR]

To provide precise recommendations, traditional recommender systems (RS) collect personal data, user preference and feedback, which are sensitive to each user if such information is maliciously used for extra analysis. In recent years, differential privacy (DP) has been widely applied in RS to provide privacy protection for sensitive information. Prior studies explored the combination of DP and RS, while neglecting the disparate effect on model accuracy of imbalanced subgroups as large user groups control the trained model, and DP can worsen the disparate effect of degrading the performance of recommender systems significantly. Besides, the number of uploaded contributions can differ among users for training a recommender system, so it is necessary to set the user-level privacy guarantee. In this paper, we make four contributions. First, we propose an efficient way of constructing datasets for training a recommender system based on prior theories. Second, we compute the user-level priors based on user metadata to optimize the VAE model. Besides, we add noise into the calculation process to protect user metadata. Third, we analyze and propose a tighter theoretical bound on gradient updates for DP Stochastic Gradient Descent (DPSGD). Finally, we exploit these theoretical results and propose a novel DP-VAE based recommender system. Extensive experimental results on multiple datasets show that our system can achieve high recommendation precision while maintaining a reasonable privacy guarantee. [ABSTRACT FROM AUTHOR]

Recommender systems aim to suggest items of interest from historical interactions between users and items. Advanced methods such as path-based recommendations attempt to capture high-order user–item interactions for better recommendation performance. However, these methods focus more on the local view of user–item interactions, ignoring the global view, which limits performance. Two more drawbacks further restrict the performance of existing methods. First, most do not consider the relations (or relation directions) between successive nodes when constructing a multi-hop path (i.e., high-order local interactions) from a user to a target item. Second, the influential factors of path length and number of relation types are ignored when computing path importance to aggregate paths for better local representation. To resolve these issues, we propose a recommendation model that can well capture high-order local and global (HOLG) user–item interactions. A path-embedding module learns a local representation of user–item interactions via a long short-term memory network, taking (directed) relations of successive nodes as input. Multiple local representations are aggregated with an attention network, using both path length and number of relation types as important factors. A graph-embedding module learns a global representation of user–item interactions by constructing a subgraph from sampled user–item paths. Experiments on the LinkedIn, MovieLens-1M, and Yelp datasets validate our approach, which performs best in comparison with eight baselines. [ABSTRACT FROM AUTHOR]

Leveraging knowledge graphs for post-hoc recommendation explanations has been investigated in recent years. Existing approaches rely mainly on the overlap properties (encoded by knowledge graphs) that characterize both user liked items and the recommended ones. These approaches, however, do not fully leverage the property hierarchy of knowledge graphs which may lead to flawed explanations. In this paper we introduce an approach that takes the whole property hierarchy into account. This is done with a limited computation time overhead thanks to efficient algorithmic optimizations relying on sub-ontology extraction. The hierarchical relationships among properties are also considered to avoid redundant properties for explanation. We carried out a user study of 155 participants in the movie recommendation domain and used both offline and online metrics to assess the proposed approach. Significant improvements, in terms of informativeness (by 39%), persuasiveness (by 22%), engagement (by 29%) and user trust (by 26%), are suggested by the obtained results, as compared to the state-of-the-art property-based explanation model. Our findings indicate the superiority of accounting for the whole property hierarchy when dealing with post-hoc recommendation explanations. [ABSTRACT FROM AUTHOR]

Nowadays, Knowledge Distillation (KD) has been widely studied for recommender system. KD is a model-independent strategy that generates a small but powerful student model by transferring knowledge from a pre-trained large teacher model. Recent work has shown that the knowledge from the teacher's representation space significantly improves the student model. The state-of-the-art method, named Distillation Experts (DE), adopts cluster-wise distillation that transfers the knowledge of each representation cluster separately to distill the various preference knowledge in a balanced manner. However, it is challenging to apply DE to a new environment since its performance is highly dependent on several key assumptions and hyperparameters that need to be tuned for each dataset and each base model. In this work, we propose a novel method, dubbed Personalized Hint Regression (PHR), distilling the preference knowledge in a balanced way without relying on any assumption on the representation space nor any method-specific hyperparameters. To circumvent the clustering, PHR employs personalization network that enables a personalized distillation to the student space for each user/item representation, which can be viewed as a generalization of DE. Extensive experiments conducted on real-world datasets show that PHR achieves comparable or even better performance to DE tuned by a grid search for all of its hyperparameters. [ABSTRACT FROM AUTHOR]

Bayesian personalized ranking (BPR) has been proposed as an effective method to model pairwise learning, and it is widely used in many personalized recommender systems. However, the effectiveness of BPR can be seriously affected by an imbalanced data distribution because it tends to rank popular items ahead of personalized items. As a result, the personalized needs of users cannot be well met. In this paper, we propose a novel personalized recommendation method called similarity pairwise ranking (SPR) to rank users' favorite items first. SPR eliminates the differences in the scores between popular and personalized items based on their similarity by using a new penalty. In such a way, the SPR-enhanced recommendation will render meaningful and personalized results that better meet the individual needs of users, and it overcomes the negative impact of imbalanced datasets. We design a model to illustrate the improvement of SPR: similarity pairwise ranking matrix factorization (SPRMF). Experimental results obtained using six datasets indicate the superiority in recommendation quality of SPRMF over the recent state-of-the-art methods. • A pairwise method based on item similarity is proposed. • Item similarity is used to overcome the impact of data imbalance on the pairwise ranking method. • The method narrows the prediction score between similar item pairs in Bayesian personalized ranking method. • A uniform sample method based on ratings is used to obtain similar item pairs. • The method is validated in six datasets with promising results. [ABSTRACT FROM AUTHOR]

Inferring the substitutable and complementary products for a given product is an essential and fundamental concern for the recommender system. To achieve this, existing approaches take advantage of the knowledge graphs to learn more evidences for inference, whereas they often suffer from invalid reasoning for lack of elegant decision making strategies. Therefore, we propose a novel Knowledge-Aware Path Reasoning (KAPR) model which leverages the dynamic policy network to make explicit reasoning over knowledge graphs, for inferring the substitutable and complementary relationships. Our contributions can be highlighted as three aspects. Firstly, we model this inference scenario as a Markov Decision Process in order to accomplish a knowledge-aware path reasoning over knowledge graphs. Secondly, we integrate both structured and unstructured knowledge to provide adequate evidences for making accurate decision-making. Thirdly, we evaluate our model on a series of real-world datasets, achieving competitive performance compared with state-of-the-art approaches. Our code is released on https://gitee.com/yangzijing_flower/kapr/tree/master. [ABSTRACT FROM AUTHOR]

Due to the development of graph neural networks, graph-based representation learning methods have made great progress in recommender systems. However, data sparsity is still a challenging problem that most graph-based recommendation methods are confronted with. Recent works try to address this problem by utilizing side information. In this paper, we exploit the relevant and easily accessible textual information by advanced natural language processing (NLP) models and propose a light RGCN-based (RGCN, relational graph convolutional network) collaborative filtering method on heterogeneous graphs. Specifically, to incorporate rich textual knowledge, we utilize a pre-trained NLP model to initialize the embeddings of text nodes. Afterward, by performing a simplified RGCN-based node information propagation on the constructed heterogeneous graph, the embeddings of users and items can be adjusted with textual knowledge, which effectively alleviates the negative effects of data sparsity. Moreover, the matching function used by most graph-based representation learning methods is the inner product, which is not appropriate for the obtained embeddings that contain complex semantics. We design a predictive network that combines graph-based representation learning with neural matching function learning, and demonstrate that this architecture can bring a significant performance improvement. Extensive experiments are conducted on three publicly available datasets, and the results verify the superior performance of our method over several baselines. [ABSTRACT FROM AUTHOR]