top-N Recommendation Task Research Articles

CF-MGAN: Collaborative filtering with metadata-aware generative adversarial networks for top-N recommendation

Generative Adversarial Networks (GANs) have several advantages over latent variable and autoregressive models, such as directly computing the input data distribution, fast sampling, and best samples generated. Furthermore, the metadata information of users and items plays a crucial role in effectively learning the user-item latent representations. Existing neural network approaches are deterministic, incapable of accounting for user-item latent representations, and the rating prediction process is calculative rather than generative. However, only some papers have utilized GANs for information retrieval tasks. However, they have several limitations, such as the inability to abstract data effectively, the limitation of the model's ability to learn accurate latent representations, and the absence of metadata information. To address these issues, we propose a GAN-based recommendation framework named Collaborative Filtering with Metadata-awareGenerative Adversarial Networks (CF-MGAN) to learn complex rating distribution for collaborative filtering (CF) tasks. We use a negative item generator for the implicit user purchase vector to prevent GAN from generating biased ratings. We exploited metadata information to overcome data sparsity issues. We use Variational Autoencoder to abstract user-item metadata features, a regularizer for GAN-generated rating. Extensive experiments on datasets reveal the effectiveness of the proposed model, showing an improvement in the top-N recommendation task.

Distributional Fairness-aware Recommendation

Fairness has been gradually recognized as a significant problem in the recommendation domain. Previous models usually achieve fairness by reducing the average performance gap between different user groups. However, the average performance may not sufficiently represent all the characteristics of the performances in a user group. Thus, equivalent average performance may not mean the recommender model is fair, for example, the variance of the performances can be different. To alleviate this problem, in this article, we define a novel type of fairness, where we require that the performance distributions across different user groups should be similar. We prove that with the same performance distribution, the numerical characteristics of the group performance, including the expectation, variance, and any higher-order moment, are also the same. To achieve distributional fairness, we propose a generative and adversarial training framework. Specifically, we regard the recommender model as the generator to compute the performance for each user in different groups, and then we deploy a discriminator to judge which group the performance is drawn from. By iteratively optimizing the generator and the discriminator, we can theoretically prove that the optimal generator (the recommender model) can indeed lead to the equivalent performance distributions. To smooth the adversarial training process, we propose a novel dual curriculum learning strategy for optimal scheduling of training samples. Additionally, we tailor our framework to better suit top-N recommendation tasks by incorporating softened ranking metrics as measures of performance discrepancies. We conduct extensive experiments based on real-world datasets to demonstrate the effectiveness of our model.

GraphHINGE: Learning Interaction Models of Structured Neighborhood on Heterogeneous Information Network

Heterogeneous information network (HIN) has been widely used to characterize entities of various types and their complex relations. Recent attempts either rely on explicit path reachability to leverage path-based semantic relatedness or graph neighborhood to learn heterogeneous network representations before predictions. These weakly coupled manners overlook the rich interactions among neighbor nodes, which introduces an early summarization issue. In this article, we propose GraphHINGE ( H eterogeneous IN teract and aggre G at E ), which captures and aggregates the interactive patterns between each pair of nodes through their structured neighborhoods. Specifically, we first introduce Neighborhood-based Interaction (NI) module to model the interactive patterns under the same metapaths, and then extend it to Cross Neighborhood-based Interaction (CNI) module to deal with different metapaths. Next, in order to address the complexity issue on large-scale networks, we formulate the interaction modules via a convolutional framework and learn the parameters efficiently with fast Fourier transform. Furthermore, we design a novel neighborhood-based selection (NS) mechanism, a sampling strategy, to filter high-order neighborhood information based on their low-order performance. The extensive experiments on six different types of heterogeneous graphs demonstrate the performance gains by comparing with state-of-the-arts in both click-through rate prediction and top-N recommendation tasks.

NFC: a deep and hybrid item-based model for item cold-start recommendation

New items, also called cold-start items, are introduced every day in the catalogs of numerous online systems. Due to the absence of previous preferences, recommending these items is difficult but important task for a recommender system. For this reason, the item cold-start recommendation problem still represents an interesting research topic for the community. In this work, we propose Neural Feature Combiner (NFC), a novel deep learning, item-based approach for cold-start item recommendation. The model learns to map the content features of the items into a low-dimensional hybrid embedding space. The features that compose the embeddings are then combined in order to reproduce collaborative item similarity values. We compare NFC with three variants of the same model that learn from user feedback, showing the advantages of learning from similarities in terms of accuracy and convergence time. With an extensive set of experiments on four datasets, we show that NFC outperforms several cutting edge approaches in the top-n recommendation task of cold-start items. Results in extremely cold (i.e., with a very low amount of interactions for training) and cold-warm hybrid scenarios prove that NFC effectively exploits collaborative information, leading to state-of-the-art accuracy. We finally conduct a qualitative analysis of the embeddings generated by different models, and we provide an analysis of the importance that different models assign to the input features, empirically demonstrating the robustness of the hybrid representations produced by our new model.

A synchronized heterogeneous autoencoder with feature-level and label-level knowledge distillation for the recommendation

Motivated by the success of deep learning techniques, there are numerous deep learning models developed for recommender systems. User-oriented autoencoder (UAE) and item-oriented autoencoder (IAE) are two typical approaches developed for the recommendation by learning features from items or users respectively. Existing works demonstrate that IAE model outperforms UAE model for rating prediction tasks, while UAE model outperforms IAE model for top-N recommendation task. This fact motivates us to develop a new synchronized heterogeneous autoencoder (SHAE) for top-N recommendation by considering both features learned by IAE and UAE models. Especially, we develop two novel heterogeneous knowledge distillation methods in feature-level and label-level to build relations between IAE and UAE models. Compared with matrix factorization approaches, our methods are more efficient with acceptable computational complexity for the recommendation. We conduct comprehensive experiments on public datasets to compare with several state-of-the-art approaches. Experimental results demonstrate that the proposed method significantly outperforms other comparisons for top-N recommendation tasks.

GLIMG: Global and local item graphs for top-N recommender systems

Graph-based recommendation models work well for top-N recommender systems due to their capability to capture the potential relationships between entities. However, most of the existing methods only construct a single global item graph shared by all the users and regrettably ignore the diverse tastes between different user groups. Inspired by the success of local models for recommendation tasks, this paper provides the first attempt to investigate multiple local item graphs along with a global item graph for graph-based recommendation models. We argue that recommendation on global and local graphs outperforms that on a single global graph or multiple local graphs. Specifically, we propose a novel graph-based recommendation model named GLIMG (Global and Local IteMGraphs), which simultaneously captures both the global and local user tastes. By integrating the global and local graphs into an adapted semi-supervised learning model, users’ preferences on items are propagated globally and locally. Extensive experimental results on real-world datasets show that our proposed method consistently outperforms the state-of-the-art counterparts on the top-N recommendation task.

A novel deep quantile matrix completion model for top-N recommendation

Matrix completion models have been receiving keen attention due to their wide applications in science and engineering. However, the majority of these models assumes a symmetric noise distribution in their completion processes and uses conditional mean to characterize data distribution in a data set, the assumption of which incurs noticeable bias toward outliers. Recognizing the fact that noise distribution tends to be asymmetric in the real-world, this paper proposes a novel Deep Quantile Matrix Completion model, abbreviated as DQMC, which aims to accurately capture noise distribution in a data set by modeling conditional quantile of the data set instead of its conditional mean as traditionally handled by many state-of-the-art methods. Implemented via a deep computing paradigm, the newly proposed model maps a data set from its input space to the latent spaces through a two-branched deep autoencoder network. Such a mapping can effectively capture complex information latent in the data set. The proposed model is empowered by two key designed elements, including: (1) its two-branched deep autoencoder network that provides a flexible computing pathway to attain completion results with a high quality; (2) the introduction of a quantile loss function in combination with the proposed deep network, leading to a new unsupervised learning algorithm for tackling the matrix completion tasks with a superior capability. Comparative experimental results consistently demonstrate the superiority of the proposed DQMC model in conducting the top-N recommendation tasks involving both explicit and implicit rating data sets with respect to a series of state-of-the-art recommendation algorithms.

A mixed heterogeneous factorization model for non-overlapping cross-domain recommendation

Cross-domain collaborative filtering has attracted a great deal of attention for its capability of dealing with data sparsity by transferring valuable knowledge from auxiliary domains to assist in recommendation in a target domain. State-of-the-art research has mainly focused on the scenario in which auxiliary domains share users or items with a target domain. However, such auxiliary data is rare or the acquisition of them is limited due to privacy concerns in real-world applications. We investigate a more realistic scenario in which auxiliary domains have neither users nor items overlapped with a target domain, thereby facilitating the collection of auxiliary data. In order to extract and transfer the sharing knowledge between auxiliary domains and target domain, we assume the user-item feedback in each domain consist of two parts: domain-transferrable information containing the shared knowledge and domain-reserved information reflecting the domain-specific characteristics. Correspondingly, we propose a mixed heterogeneous factorization model to capture the sharing knowledge and the domain-specific characteristics based on adapted tensor factorization and biased matrix factorization respectively and then combine them together in an accumulative way. Meanwhile, three types of domain heterogeneity including preference heterogeneity, characteristics heterogeneity and rating bias are taken into account in this model. Experimenting on four publicly available datasets across different domains, we show that our model is superior to state-of-the-art methods in rating prediction and top-N recommendation tasks.

Mixture Matrix Approximation for Collaborative Filtering

Matrix approximation (MA) methods are integral parts of today's recommender systems. In standard MA methods, only one feature vector is learned for each user/item, which may not be accurate enough to characterize the diverse interests of users/items. For instance, users could have different opinions on a given item, so that they may need different feature vectors for the item to represent their unique interests. To this end, this article proposes a mixture matrix approximation (MMA) method, in which we assume that the user-item ratings follow mixture distributions and the user/item feature vectors vary among different stars to better characterize the diverse interests of users/items. Furthermore, we show that the proposed method can tackle both rating prediction and the top-N recommendation problems. Empirical studies on MovieLens, Netflix and Amazon datasets demonstrate that the proposed method can outperform state-of-the-art MA-based collaborative filtering methods in both rating prediction and top-N recommendation tasks.

Scalable and Explainable 1-Bit Matrix Completion via Graph Signal Learning

One-bit matrix completion is an important class of positive-unlabeled (PU) learning problems where the observations consist of only positive examples, e.g., in top-N recommender systems. For the first time, we show that 1-bit matrix completion can be formulated as the problem of recovering clean graph signals from noise-corrupted signals in hypergraphs. This makes it possible to enjoy recent advances in graph signal learning. Then, we propose the spectral graph matrix completion (SGMC) method, which can recover the underlying matrix in distributed systems by filtering the noisy data in the graph frequency domain. Meanwhile, it can provide micro- and macro-level explanations by following vertex-frequency analysis. To tackle the computational and memory issue of performing graph signal operations on large graphs, we construct a scalable Nystrom algorithm which can efficiently compute orthonormal eigenvectors. Furthermore, we also develop polynomial and sparse frequency filters to remedy the accuracy loss caused by the approximations. We demonstrate the effectiveness of our algorithms on top-N recommendation tasks, and the results on three large-scale real-world datasets show that SGMC can outperform state-of-the-art top-N recommendation algorithms in accuracy while only requiring a small fraction of training time compared to the baselines.

Speed up Differential Evolution for ranking of items in recommendation systems

Recommendation systems can suggest users list of items they have not yet seen but might be interested in. To improve the quality of the generated recommendations, different techniques are often used which try to personalize recommendations. Usually user preferences are stored in the form of a vector in which individual values describe to what extent a given feature is desired by the user. To find this vector, metaheuristic algorithms can be used, however their main drawback is their computational complexity. Therefore, in this paper, a modification of the Differential Evolution algorithm is proposed to enable faster computation of the ranking score for each item in the system, which is used to create a recommendation list. Experiments have been performed on the current MovieLens 25m database and they show that our modification can significantly speed up the process of finding a preference vector, without losing their quality for the top-N recommendation task. We will also address the vulnerability of recommendation systems to profile injection attacks, as a result of which an attacker can influence the generated recommendations.

Attribute-aware deep attentive recommendation

Since the rich semantics of attribute information has become a great supplement to the ratings data in designing recommender systems, fusing attributes information into ratings has shown promising performance for many recommendation tasks. However, the use of attribute information is not easy, because different attributes are often: (1) multi-source, that is, attributes may come from many different fields, (2) unstructured, (3) unbalanced, (4) heterogeneous. In this paper, we explore effective fusion of user-item ratings and item attributes to improve recommendations, we propose an attribute-aware deep attentive recommendation model, which embeds attribute information into the latent semantic space of items through the attention mechanism, forming more accurate item representations. Extensive experiments show that our method is superior to the existing methods on both rating prediction and Top-N Recommendation tasks.

A collaborative filtering recommendation framework based on Wasserstein GAN

Compared with the original GAN, Wasserstein GAN minimizes the Wasserstein Distance between the generative distribution and the real distribution, can well capture the potential distribution of data and has achieved excellent results in image generation. However, the exploration of Wasserstein GAN on recommendation systems has received relatively less scrutiny. In this paper, we propose a collaborative filtering recommendation framework based on Wasserstein GAN called CFWGAN to improve recommendation accuracy. By learning the real user distribution, we can mine the potential nonlinear interactions between users and items, and capture users’ preferences for all items. Besides, we combine two positive and negative item sampling methods and add the reconstruction loss to the generator’s loss. This can well handle the problem of discrete data in recommendation (relative to the continuity of image data). By continuously approximating the generative distribution to the real user distribution, we can finally obtain better users’ preference information and provide higher accuracy in recommendation. We evaluate the CFWGAN model on three real-world datasets, and the empirical results show that our method is competitive with or superior to state-of-the-art approaches on the benchmark top-N recommendation task.

A correlative denoising autoencoder to model social influence for top-N recommender system

In recent years, there are numerous works been proposed to leverage the techniques of deep learning to improve social-aware recommendation performance. In most cases, it requires a larger number of data to train a robust deep learning model, which contains a lot of parameters to fit training data. However, both data of user ratings and social networks are facing critical sparse problem, which makes it not easy to train a robust deep neural network model. Towards this problem, we propose a novel Correlative Denoising Autoencoder (CoDAE) method by taking correlations between users with multiple roles into account to learn robust representations from sparse inputs of ratings and social networks for recommendation. We develop the CoDAE model by utilizing three separated autoencoders to learn user features with roles of rater, truster and trustee, respectively. Especially, on account of that each input unit of user vectors with roles of truster and trustee is corresponding to a particular user, we propose to utilize shared parameters to learn common information of the units that corresponding to same users. Moreover, we propose a related regularization term to learn correlations between user features that learnt by the three subnetworks of CoDAE model. We further conduct a series of experiments to evaluate the proposed method on two public datasets for Top-N recommendation task. The experimental results demonstrate that the proposed model outperforms state-of-the-art algorithms on rank-sensitive metrics of MAP and NDCG.

TNAM: A tag-aware neural attention model for Top-N recommendation

Recent work shows that incorporating tag information to recommender systems is promising for improving the recommendation accuracy in social systems. However, existing approaches suffer from less reasonable assignment of tag weights when constructing the user profiles and item characteristics in real-world scenarios, resulting in decreased accuracy in making recommendations. The above issue is specifically summarized into two aspects: 1) the weight of a target item is mainly determined by number of one certain type of tags, and 2) users place equal focus on the same tag for different items. To tackle these problems, we propose a novel model named TNAM, a Tag-aware Neural Attention Model, which accurately captures users’ special attention to tags of items. In the proposed model, we design a tag-based neural attention network by extracting potential tag information to overcome the difficulty of assigning tag weights for personalized users. We combine user-item interactions with tag information to map sparse data to dense vectors in higher-order space. In this way, TNAM acquires more interrelations between users and items to make recommendations more accurate. Extensive experiments of our model on three publicly implicit feedback datasets reveal significant improvements on the metrics of HR and NDCG in Top-N recommendation tasks over several state-of-the-art approaches.

Information retrieval models for recommender systems

Information retrieval addresses the information needs of users by delivering relevant pieces of information but requires users to convey their information needs explicitly. In contrast, recommender systems offer personalized suggestions of items automatically. Ultimately, both fields help users cope with information overload by providing them with relevant items of information. This thesis aims to explore the connections between information retrieval and recommender systems. Our objective is to devise recommendation models inspired in information retrieval techniques. We begin by borrowing ideas from the information retrieval evaluation literature to analyze evaluation metrics in recommender systems [2]. Second, we study the applicability of pseudo-relevance feedback models to different recommendation tasks [1]. We investigate the conventional top-N recommendation task [5, 4, 6, 7], but we also explore the recently formulated user-item group formation problem [3] and propose a novel task based on the liquidation of long tail items [8]. Third, we exploit ad hoc retrieval models to compute neighborhoods in a collaborative filtering scenario [9, 10, 12]. Fourth, we explore the opposite direction by adapting an effective recommendation framework to pseudo-relevance feedback [13, 11]. Finally, we discuss the results and present our conclusions. In summary, this doctoral thesis adapts a series of information retrieval models to recommender systems. Our investigation shows that many retrieval models can be accommodated to deal with different recommendation tasks. Moreover, we find that taking the opposite path is also possible. Exhaustive experimentation confirms that the proposed models are competitive. Finally, we also perform a theoretical analysis of some models to explain their effectiveness. Advisors : Álvaro Barreiro and Javier Parapar. Committee members : Gabriella Pasi, Pablo Castells and Fidel Cacheda. The dissertation is available at: https://www.dc.fi.udc.es/~dvalcarce/thesis.pdf.

Attentive Aspect Modeling for Review-Aware Recommendation

In recent years, many studies extract aspects from user reviews and integrate them with ratings for improving the recommendation performance. The common aspects mentioned in a user’s reviews and a product’s reviews indicate indirect connections between the user and product. However, these aspect-based methods suffer from two problems. First, the common aspects are usually very sparse, which is caused by the sparsity of user-product interactions and the diversity of individual users’ vocabularies. Second, a user’s interests on aspects could be different with respect to different products, which are usually assumed to be static in existing methods. In this article, we propose an Attentive Aspect-based Recommendation Model (AARM) to tackle these challenges. For the first problem, to enrich the aspect connections between user and product, besides common aspects, AARM also models the interactions between synonymous and similar aspects. For the second problem, a neural attention network which simultaneously considers user, product, and aspect information is constructed to capture a user’s attention toward aspects when examining different products. Extensive quantitative and qualitative experiments show that AARM can effectively alleviate the two aforementioned problems and significantly outperforms several state-of-the-art recommendation methods on the top-N recommendation task.

Modeling Side Information in Preference Relation based Restricted Boltzmann Machine for recommender systems

A majority of the collaborative filtering techniques exploit user-item rating information to generate recommendations of unseen items for a user. However, a user’s preference also depends on some extra information like item features, user attributes and others, which is known as side information. Further, according to recent studies, using preference relation as an alternative to absolute ratings often produces quality recommendations. This study proposes a collaborative filtering technique using Preference Relation based Restricted Boltzmann Machine for recommender system. The proposed method takes the preference relations of items as input and generates a ranking of items for any user. Using Conditional Restricted Boltzmann Machine, the side information of items along with preference relations are integrated into the model. Besides side information, the proposed method is also able to capture second order and higher order user-item interactions. Experimental verification of the proposed model is done using three datasets: MovieLens-1M, MovieLens-20M, and Book-Crossing, which are the most widely used datasets for testing recommender systems. Results obtained at different positions using standard ranking measures like, NDCG and MAP, indicate that the performance of the proposed method is better compared to related state-of-the-art collaborative filtering models for Top-N recommendation task.

UTSP: User-Based Two-Step Recommendation With Popularity Normalization Towards Diversity and Novelty

Information technologies such as e-commerce and e-news bring overloaded information as well as convenience to users, cooperatives and companies. Recommender system is a significant technology in solving this information overload problem. Due to the outstanding accuracy performance in top-N recommendation tasks, two-step recommendation algorithms are suitable to generate recommendations. However, their recommendation lists are biased towards popular items. In this paper, we propose a user based two-step recommendation algorithm with popularity normalization to improve recommendation diversity and novelty, as well as two evaluation metrics to measure diverse and novel performance. Experimental results demonstrate that our proposed approach significantly improves the diversity and novelty performance while still inheriting the advantage of two-step recommendation approaches on accuracy metrics.

Interest-Related Item Similarity Model Based on Multimodal Data for Top-N Recommendation

Nowadays, the recommendation systems are applied in the fields of e-commerce, video websites, social networking sites, which bring great convenience to people's daily lives. The types of information are diversified and abundant in recommendation systems; therefore the proportion of unstructured multimodal data such as text, image, and video is increasing. However, due to the representation gap between different modalities, it is intractable to effectively use unstructured multimodal data to improve the efficiency of recommendation systems. In this paper, we propose an end-to-end multimodal interest-related item similarity model (multimodal IRIS) to provide recommendations based on the multimodal data source. Specifically, the multimodal IRIS model consists of three modules, i.e., multimodal feature learning module, the interest-related network (IRN) module, and item similarity recommendation module. The multimodal feature learning module adds knowledge sharing unit among different modalities. Then, IRN learns the interest relevance between target item and different historical items respectively. Finally, the multimodal feature learning, IRN, and item similarity recommendation modules are unified into an integrated system to achieve performance enhancements and to accommodate the addition or absence of different modal data. Extensive experiments on real-world datasets show that, by dealing with the multimodal data which people may pay more attention to when selecting items, the proposed multimodal IRIS significantly improves accuracy and interpretability on top-N recommendation task over the state-of-the-art methods.