New Trends in Artificial Intelligence for Recommender Systems and Collaborative Filtering

In recent times, recommender systems (RSs) have been attracting a lot of attention from the research community because of their groundbreaking applications. This has led to software-intensive companies like Amazon, Netflix, Spotify, or Google relying on RSs to organize their huge catalogue of products and to offer highly attractive items to their users.

In a highly connected society, consumers are exposed to a wide offering of products to be consumed, a large number of advertisements to attract new purchases, and a huge amount of data about the setup of these purchased items. Furthermore, this overload of information is even more overwhelming if we also consider multi-source data to which we are exposed to daily, like traffic information, financial trading information, or news, among others. Moreover, the inclusion of social networks in our lives has opened a new landscape for offering data, since social network users are intensive consumers of ever-changing new content. For this reason, it is crucial to provide intelligent systems capable of managing this large amount of data, sorting it according to the preferences and likes of the users, and offering a small portion of highly relevant content to consumers. For this purpose, RSs were developed with the aim of addressing this information overload problem.

To address these problems, in recent years, the RS community has proposed new, astonishing, and very innovative solutions. Currently, the area is experiencing an exciting revolution of traditional Collaborative Filtering (CF) methods, based on Matrix Factorization and K-Nearest Neighbors, with the incorporation of cutting-edge technologies. Neural Networks, Deep Learning, Model Explainability, and Fair Prediction, among others, are making their way into the realm of RSs, importing techniques from other areas of artificial intelligence to provide novel approaches.

In this Special Issue, we aim to widen the boundary of knowledge in CF-based RSs with new proposals incorporating cutting edge trends in artificial intelligence. In addition, we have collected exciting novel applications of RS techniques to address new challenges in real-world problems.

1. Theoretical Advances towards New Recommender Systems

One of the main research lines we have investigated in this Special Issue is the development of new RS models and technology. These novel approaches have allowed us to address old problems in radically new ways.

Some of the works collected in this issue have proposed disruptive changes in the nature of the prediction algorithm. In this direction, in [1], the authors introduce the novel concept of ‘black sheep neighbors’, which are groups of similar users that have a discordant opinion with respect to the general consensus about controversial items. Enhancing the role of these black sheep neighbors, the new models can outperform the baselines in terms of accuracy and efficacy.

In a similar vein, in [2], the authors propose to import techniques from natural language processing to the RS world. In particular, they propose to apply a relaxation algorithm to handcraft a more sophisticated vector embedding of users and items into the latent space. The accuracy of these embeddings is known to be crucial for the success of the RS, since they aim to capture, in only a few factors, the essence of a user and an item. Only if the factors are well computed do the comparisons between users’ and items’ factors accurately characterize the affinity, which is a necessary condition for achieving a good performance for these methods. Hence, in [2], these techniques inspired by the classical word vector embedding are used to provide a new versatile framework to compute these latent factors, leading to a compelling performance in real-life multi-purpose datasets.

This idea of exploiting the similarities between users has also been studied in [3]. There, the authors propose an innovative solution to incorporate relevant user feedback to improve the CF process. Inter-coder agreement techniques are also used to evaluate the performance of the suggested method in the context of hotel recommendations, with a clear improvement with respect to the state of the art.

Following these ideas of studying the intrinsic similarities between users, in this Special Issue, we have included paper [4]. In this paper, a novel sampling criterion for neighbors is introduced to speed up the training of Graph Convolutional Neural Networks (GCNNs) applied to RSs. Despite the effectiveness and accuracy of GCNNs, their computational complexity makes them unfeasible for tasks in real-world RSs. Instead of using the classical recursive approach, in this paper, the authors apply Kullback–Leibler divergence to compare the interaction probability distribution of the vertices. The results demonstrate a noticeable drop in the computational time while preserving the accuracy of the model.

Other graph-based approaches to RSs have been explored in this issue. For example, in [5], the authors propose to use an attributed graph-based representation to encompass several user features in location recommendation problems. Using this additional information, the RS is able to provide a much more accurate latent factor embedding in these highly sparse settings. The empirical results support these findings, with a substantial improvement in the accuracy for multiple localization recommendation datasets.

Another graph-based RS was explored in [6]. In this case, a knowledge graph approach is taken to enhance the attention aggregation network. Thus, the authors propose the KANR model, which is composed of three stages: an attention aggregation network to gather user features, a knowledge graph embedding to push these features into a latent space, and a final fusion component to provide the final predictions.

Finally, in this Special Issue, we have also explored new ways of looking at the prediction problem from a regression point of view. In [7], a novel approach using the Dirichlet distribution was developed. Classically, the prediction problem for rating filling is addressed in an RS as a classification problem, in which one of the possible rating values must be assigned to each unknown rating. However, recent works have shown that the system may benefit from looking at this problem as a regression problem in such a way that the output of the RS should not be a single prediction, but a whole probability distribution. In this way, the mode of the distribution must be seen as the predicted value, but the actual value of the probability of the mode may be also interpreted as a reliability measure of the prediction. In this manner, the aim of [7] was to create DirMF, a matrix-factorization-based approach in which the outputs of the model are the parameters of the Dirichlet distribution, whose samples lead to the desired rating probability distributions for users and items.

2. Recommender Systems in Action

The study of the applicability of these techniques to real life is as important as developing theoretical models for RSs. In this spirit, complementary to the importance of the theoretical work conducted in the above-mentioned papers, in this Special Issue, we have also conducted a thorough study of use cases of RSs in real-life problems. We should not forget that eventually the goal of these systems in particular, and of artificial intelligence in general, is to provide new solutions to social problems that alleviate information overexposure and improve people’s lives through technology.

With this idea in mind, in paper [8], the authors apply RSs to online education. The amount of educational resources available on the internet is huge and this trend has accelerated even more in pandemic and post-pandemic scenarios. To match learners with online courses, in [8], the authors create DECOR, a system able to accurately recommend educational resources, as tested in several real-world datasets.

In the same spirit of democratizing access to online resources for the population, in [9], the authors propose Negative-items Mixed Collaborative Filtering (NMCF), a method to improve accessibility to e-government resources. By emphasizing the learning of positive items’ latent features, the system is able to outperform the state-of-the-art algorithms on a real e-government dataset.

Furthermore, devoted to online resources, but now in the context of software development, we find [10]. There, the authors propose to apply RSs to suggest code repositories to software developers on the GitHub platform. Applying deep learning techniques, the method is able to provide accurate sequential recommendations of repositories, also introducing the first dataset in this direction. We expect that this work will open new perspectives on this interesting topic.

The problem of localization recommendation is also addressed in [11]. In this paper, a new method to build user profiles based on their Twitter interactions, personal preferences, and travel habits is introduced. This model, called clustering-based profiling, has been successfully applied to recommendations for visits to the city of Barcelona, with a noticeable improvement in performance relative to baselines.

In addition, a very exciting cross-approach was explored in [12]. In this work, the authors apply Generative Adversarial Networks (GANs) to boost the performance of financial portfolio management. To be precise, the GAN is trained to generate, given an investor profile, the ideal financial advisor to manage its assets. To do so, the features of the financial advisor are encoded into a grayscale image so that the goal of the GAN is to generate images that represent the recommended advisor. This technique has been empirically shown to be very compelling, with a remarkable increase in the obtained return compared to the baselines.

Finally, we close this Special Issue with a survey on the use of RSs in the real estate market, presented in [13]. The recommendation task in real estate platforms presents several specificities that make it a particularly challenging problem. High sparsity, many cold users, and strong search constraints are some of the problems that this type of system must address. In this work, the authors analyzed 26 research articles to evaluate and compare existing solutions in the literature, classifying them according to their methodology. Furthermore, prospective future research is proposed.

3. Conclusions

In this Special Issue, we have contributed to the expansion of the frontiers of knowledge in recommender systems, both by proposing new methods and algorithms to issue predictions and by analyzing their practical perspective. The contributions collected are playing a crucial role in broadening existing methods and discovering innovative applications of recommender systems. The researchers in this field are a highly engaged global community, determined to tackle new problems with the final goal of improving people’s lives. We expect that this Special Issue will serve as a remarkable testament to their determination, and we will persist in advancing our efforts to meet society’s growing needs for novel solutions to longstanding problems.