Generative Adversarial Networks (GANs) have emerged as powerful tools in the realm of artificial intelligence, reshaping various domains, from image generation to text synthesis and music composition. In recent years, researchers have ventured into the realm of integrating GANs into recommendation systems, driven by the desire to elevate the quality of recommendations. This article presents an in-depth exploration of the current landscape surrounding the incorporation of GANs in recommendation systems. Researchers have harnessed the potential of GANs to craft highly personalized recommendations by incorporating user and item features. Notably, techniques like conditional GANs have been employed to consider user demographics, browsing history, and item attributes, enabling the tailoring of recommendations to individual preferences. Nevertheless, challenges have surfaced, including issues related to training stability, mode collapse, scalability limitations, and data privacy concerns in the application of GANs to recommendation systems. Diligent and persistent research endeavors are actively addressing these challenges with the overarching goal of not only overcoming the hurdles but also enhancing the stability and performance of GANs within recommendation systems. This article serves as a comprehensive guide to the current state of GANs in recommendation systems, ofering insights into their potential and the evolving landscape of research and development in this field.
making, saving time, and enhancing user satisfaction.
Personalization is the key to the efectiveness of recomIn today’s digital era, recommender systems have become mender systems. These systems analyze user preferences, essential tools that are changing the way we discover past interactions, and contextual information to provide products, services, and information. They are now in- tailored recommendations that match individual tastes tegrated into various aspects of our online lives, from and needs. While traditional recommendation methods e-commerce websites and streaming platforms to news like collaborative filtering and content-based filtering aggregators and social networks. These systems are de- have made significant improvements in recommendation signed to provide users with personalized recommenda- accuracy, they still face challenges, especially in ofertions that match their interests and preferences. This ing diverse and unexpected recommendations that go ability to sift through vast amounts of data and ofer beyond users’ known preferences. tailored suggestions has not only transformed user expe- Generative Adversarial Networks (GANs), a promising riences but also brought significant benefits to businesses, approach that has gained significant attention in recent including increased customer engagement, improved con- years. Introduced by Goodfellow and his colleagues in version rates, and enhanced customer satisfaction. 2014, GANs are a type of deep learning model known for
At their core, recommender systems aim to address their ability to generate realistic and novel data. While the overwhelming problem of information overload. As GANs were initially associated with tasks like creating the number of choices available continues to grow ex- images, they have piqued the interest of researchers for ponentially, users often find it challenging to navigate their potential to enhance recommender systems. through the sheer volume of options to discover what GANs have garnered attention in the field of recomtruly resonates with them. In this context, personalized mender systems due to their capability to capture comrecommendations play a vital role in simplifying decision- plex data patterns[1]. They ofer a unique solution to the challenge of recommendation diversity. GANs work 6th International Hybrid Conference On Informatics And Applied Math- by using a generator and discriminator network in an ematics, December 6-7, 2023 Guelma, Algeria adversarial manner. The generator learns to produce rec* Corresponding author. ommendations that resemble a user’s actual preferences, † These authors contributed equally. while the discriminator provides feedback to help the ($H.nSaEoRuIeDl.mI);amnaean@djuenl@ivl-aabngneadb.an.edtz((MN..SM.MA.NMAEAN);DseJrEiLd)i@labged.net generator create high-quality recommendations that are 0009-0006-6025-9374 (N. MANAA) distinct from what the user already knows. This dynamic © 2023 Copyright for this paper by its authors. Use permitted under Creative Commons License interaction between the generator and discriminator enAttribution 4.0 International (CC BY 4.0).
courages the generation of diverse recommendations, introducing users to new and unexpected items.
The application of GANs in recommender systems holds immense potential for improving recommendation In 2014, Goodfellow, et al.[5] introduced a groundbreakaccuracy. By leveraging deep learning and adversarial ing deep-learning technique called Generative Advertraining, GANs can push the boundaries of personalized sarial Networks (GANs). This innovative approach harrecommendations, enabling users to discover new and nessed the power of discriminative learners to construct relevant content that aligns with their evolving inter- a proficient generative learner, opening up new possibiliests. However, there are challenges to overcome, such ties in the field of artificial intelligence. as ensuring stable training, scalability, and safeguarding Generative Adversarial Networks (GANs) are a class of data privacy, to fully harness the benefits of GANs in structured probabilistic models that consist of two interrecommender systems. connected models engaged in an adversarial process.As
In this review article, we delve into recent advance- shown in Fig. 1 The first model, known as the Generator ments, methodologies, challenges, and future possibil- (G), is responsible for capturing the data distribution and ities in using GANs for recommender systems. We ex- generating synthetic data. The second model, known as plore how GANs can be applied to generate personal- the Discriminator (D), aims to discriminate between real ized recommendations, enhance diversity, tackle the cold- data samples and those generated by G. start problem, and handle sparse data. By examining the The training of GANs involves a two-player minimax strengths and limitations of GAN-based recommender game, where the Generator and Discriminator compete systems, our goal is to shed light on the potential of GANs against each other until reaching a Nash equilibrium. to revolutionize the recommendation landscape and cre- This equilibrium is achieved using a gradient-based opate new opportunities for research and innovation. timization technique called Simultaneous Gradient Descent. During training, G learns to generate data that 2. Background on Recommender closely resembles samples from the true data distribution, while D strives to correctly classify whether the data is Systems real or generated.
To update the parameters of both G and D, gradient signals are obtained from the loss incurred by comparing the distributions of real and generated data. This is typically achieved by calculating divergences between the two distributions using D as the discriminator. Through this iterative process, G and D continually improve their abilities until G is capable of generating synthetic data that is indistinguishable from real data, according to the Discriminator’s perspective [6].
Recommender systems are tools that help users overcome information overload by suggesting items that they may be interested in based on their preferences and behaviors [2]. These systems have become increasingly important in various domains, including e-commerce, social media, news, travel, and tourism [3]. The importance of recommender systems in these domains lies in their ability to improve user experience, increase user engagement, and drive revenue. By providing personalized recommendations, these systems can help users find relevant and interesting content more easily, leading to increased 4. Applications of GANs in satisfaction and engagement[2]. In e-commerce, recom- Recommender Systems mender systems can help increase sales by suggesting products that users are more likely to purchase[2]. In 4.1. Personalization Techniques tourism, these systems can help users plan their trips more eficiently by suggesting destinations, accommo- Conditional GANs: GANs that incorporate user and item dations, and activities that match their preferences[3]. features to generate personalized recommendations. The However, there are also challenges associated with rec- paper [7] presents an enhanced conditional GAN model ommender systems, such as the potential for biases and called c+GAN for generating relevant bottom item recfairness concerns. To ensure that these systems provide ommendations based on input top items. The c+GAN fair outcomes for all stakeholders involved in the rec- model incorporates a modified generator with both a ommendation process, it is important to evaluate them classical mean squared error (MSE) loss and a simplified from multiple perspectives and consider the potential perceptual loss using discrete cosine transform (DCT) coimpact on the environment and local communities[3]. eficients of the generated and target images. A simplified Additionally, there is ongoing research on how to incor- lensing technique is introduced to the discriminator to porate serendipity, or the discovery of unexpected and improve the stability of the generator training. The data novel items, into recommender systems to broaden user is clustered using a simple K-Means clustering technique preferences and improve satisfaction [4]. to enforce mode normalization across training batches. These methods result in a powerful technique that gen- clustering process is incorporated to convert the dense erates meaningful fashion items, which can be utilized GAN-generated samples into discrete and sparse vectors, for searching similar products in e-commerce platforms. which are required for creating each synthetic dataset. The generated datasets exhibit suitable distributions, ex4.2. Solve Imbalanced Data Problem pected quality values, and follow the desired evolutions compared to the source datasets.
In this work [8] proposes a hybrid GAN approach to [10] A tourism recommendation system, which readdress the data imbalance problem and improve the lies on technology, ofers suggestions to visitors based performance of recommendation systems. The authors on their preferences, previous travels, and experiences. implement a conditional Wasserstein GAN with gradient These systems gather information from diferent sources penalty to generate tabular data that includes both numer- such as web searches, user reviews, and travel history. ical and categorical values. To tackle the data imbalance However, current technologies face challenges when dealissue, an augmented auxiliary classifier loss is introduced ing with limited data for certain users or items, resultto encourage the model to generate data from the mi- ing in inaccurate recommendations. Another obstacle nority class. Additionally, the discriminator architecture in tourism is the diversity issue, where similarities are incorporates the concept of PacGAN, which processes prioritized over individual preferences. To address these multiple samples as input to overcome the mode collapse challenges, this system combines GAN (Generative Adproblem. The proposed model is evaluated based on the versarial Networks) and context-aware recommendation quality of the generated data and the performance of techniques. diferent recommendation models using the generated The primary objective of the system is to provide perdata compared to the original data. The study focuses on sonalized recommendations to travelers by considering GAN, imbalanced data, and oversampling techniques. various contextual factors. By utilizing GANs, the system identifies patterns and connections between diferent el4.3. Cold-Start Problem Solutions ements of a tourist’s environment and their preferences. Additionally, synthetic data is generated to supplement a. Synthetic Data Generation: GANs that generate syn- the original dataset, enabling the system to overcome thetic user-item interactions for cold-start users or items, issues related to cold-start and sparsity. Furthermore, enabling initial recommendations even with limited data. this approach contributes to the development of more The paper [9] introduces a novel method based on Gen- scalable recommendation systems. erative Adversarial Networks (GANs) for generating col- In this paper[11], the authors introduced the Graph laborative filtering datasets in a customizable manner. Convolutional Generative Adversarial Network (GCUnlike regular GANs, this method allows users to spec- GAN) as a solution to address the cold start problem in ify their desired number of users, items, samples, and recommendation systems. GCGAN combines the power stochastic variability. The proposed GAN model utilizes of GAN and graph convolution to efectively learn dodense, short, and continuous embedding representations main information by propagating feature values through of items and users, which enables accurate and eficient a graph structure. An important advantage of GCGAN learning compared to traditional approaches that rely is its ability to incorporate new nodes (Users) without on large and sparse input vectors.To extract the dense requiring retraining, as it leverages recommended user user and item embeddings, the authors employ a DeepMF features. By conducting experiments using the Moviemodel within the proposed architecture. Additionally, a Lens 1M dataset, we demonstrated that the proposed additional loss function for the generator. The perforGCGAN significantly outperforms the compared method mance of GANMF is evaluated using well-known datasets in terms of recommendation performance. We also ex- in the recommender systems community, demonstrating plored the impact of batch size and the number of graph improvements compared to traditional CF approaches convolution layers on the recommendation performance and other GAN-based models. An ablation study is conand observed the integration of nodes. The proposed ducted to analyze the efects of the architectural choices method is characterized by its capability to learn domain in GANMF, and a qualitative evaluation of the matrix information for both users and items, eliminating the factorization performance is provided. need for retraining the model when introducing new nodes. Moreover, it holds potential for application in var- 4.5. Improving accuracy of ious information recommendation services with similar conditions. Future research could focus on applying this recommendation method to such services.
Collaborative filtering for implicit feedback has seen successful applications of Generative Adversarial Networks 4.4. Handling sparsity and scalability (GANs). However, GANs encounter challenges in efectechniques tively capturing user interest distributions due to dififculties in feature characterization. To overcome this This paper[12] introduces a novel approach to improve issue, this paper [15] propose a collaborative filtering user recommendations by utilizing a generative network model called Improved Generative Adversarial Networks and a discriminative network in tandem. Additionally, (IGAN). In IGAN, we introduce an independent encoder an adversarial training strategy is employed to train the and generator to learn feature representations during model efectively. By leveraging the discriminative net- adversarial training. To further align with users’ interwork’s guidance, the generative network reaches an opti- est distributions and enhance recommendation accuracy, mal solution, leading to enhanced recommendation per- we incorporate the Kullback-Leibler (KL) loss and reconformance, particularly on sparse datasets. Furthermore, struction loss as penalty terms. we provide evidence demonstrating that our proposed As illustrated in Table 1, this table is designed to clasmethod substantially enhances precision. Recommender sify various techniques and approaches employed in systems face challenges in dealing with sparse interac- GAN-based recommender systems based on their obtion data and noisy data in real-world scenarios. Recently, jectives and functionalities. It serves as a foundational Generative Adversarial Network (GAN)-based recom- resource for advancing research and development in this mender systems have emerged as promising solutions ifeld, aiding in a more comprehensive understanding to tackle these issues. Negative sampling methods lever- of the diverse applications and methodologies utilized age the generator to extract informative signals from within GAN-based recommender systems. abundant unlabeled data, mitigating the data sparsity problem. However, they encounter challenges in the policy gradient training process due to sparse rewards. On 5. Challenges and Limitations of the other hand, vector reconstruction methods generate GANs in Recommender Systems user-related vectors to augment the data and improve robustness but involve redundant calculations and overlook While Generative Adversarial Networks (GANs) ofer item-specific information. To overcome the limitations promising opportunities for enhancing recommender sysof these approaches, The authors [13] propose a novel tems, there are several challenges and limitations that framework called Personalized Recommendation with need to be addressed. In this section, we discuss some Conditional Generative Adversarial Networks (PRGAN). of these challenges and their potential impact on GANThe framework considers both the user and the item based recommender systems. We also explore ongoing subset as conditions, formulating the generation of con- research eforts and potential solutions to overcome these ditional rating vectors as a user-item matching problem. limitations.
By doing so, we can control the sparsity of conditional rating vectors, simplifying the learning task for the dis- 5.1. Training Instability criminator. In [14] the authors present a new GAN-based approach called GANMF for the top-N recommendation GANs are notorious for their training instability, where problem in collaborative filtering. GANMF incorporates the generator and discriminator networks can enter a user and item latent factors using a matrix factorization cycle of chasing each other without convergence. In framework. Two unique issues in applying GAN to col- the context of recommender systems, this instability can laborative filtering are identified and addressed by using hinder the generation of accurate and reliable recommenan autoencoder as the discriminator and introducing an dations. Researchers have proposed various techniques to stabilize GAN training, such as adjusting the learning tive nature. Traditional evaluation metrics like accuracy rate, using diferent architectures, employing regulariza- or precision-recall may not capture the full picture of tion techniques, or incorporating auxiliary losses. the generated recommendations. Researchers are actively working on developing evaluation metrics that 5.2. Mode Collapse account for diversity, novelty, coverage, or serendipity in GAN-based recommender systems. These metrics aim to Mode collapse occurs when the generator fails to explore provide a comprehensive assessment of the quality and the entire item space, resulting in the generation of lim- efectiveness of recommendations generated by GANs. ited or repetitive recommendations. In recommender Ongoing research eforts focus on addressing these systems, mode collapse can lead to biased recommenda- challenges and limitations associated with GAN-based tions that focus only on popular or commonly selected recommender systems. Techniques such as progressive items. Addressing mode collapse requires strategies like training, self-supervised learning, adversarial regularimproving the diversity objectives, introducing regular- ization, and domain adaptation are being explored to ization techniques, or utilizing advanced GAN variants improve the stability and performance of GANs. Furthersuch as Wasserstein GANs or InfoGANs. more, collaborations between the recommender system community and the privacy research community aim 5.3. Scalability to develop privacy-preserving GAN architectures that protect user data while maintaining recommendation accuracy.
GAN-based recommender systems face scalability challenges when dealing with large-scale datasets or highdimensional item spaces. As the size of the data increases, training GANs becomes computationally expensive and 6. Conclusion time-consuming. Various techniques have been explored to improve scalability, such as parallelization, mini-batch This paper has provided an exploration of the role of Gentraining, distributed computing, or model compression. erative Adversarial Networks (GANs) in revolutionizing These approaches enable eficient training of GANs on recommender systems. We have witnessed the transforlarge-scale datasets, making them more applicable in mation of recommender systems from tools designed to real-world recommender systems. alleviate information overload to sophisticated engines that cater to users’ individual preferences and needs.
Recommender systems have become ubiquitous in our 5.4. Evaluation Metrics digital lives, enhancing user experiences across various domains, including e-commerce, social media, and travel.
Personalization lies at the heart of their efectiveness, as Assessing the performance of GAN-based recommender systems poses a unique challenge due to their generathese systems continuously analyze user behavior and proach to solve imbalanced data problem in recompreferences to generate tailored recommendations. How- mendation systems, IEEE Access 10 (2022) 11036– ever, traditional recommendation techniques still face 11047. challenges, especially in ofering diverse and serendipi- [9] J. Bobadilla, A. Gutiérrez, R. Yera, L. Martínez, Cretous recommendations. ating synthetic datasets for collaborative filtering
GANs, a disruptive technology that has captured the recommender systems using generative adversarattention of researchers and practitioners alike. Initially ial networks, arXiv (2023). doi:10.48550/arXiv. celebrated for their prowess in image generation, GANs 2303.01297. have proven to be equally transformative in the realm [10] E. E. Stephy, M. Rajeswari, Empowering tourists of recommender systems. These networks, driven by a with context-aware recommendations using gan, generator and discriminator, ofer a dynamic interaction in: 2023 Second International Conference on that fosters diversity in recommendations, introducing Electronics and Renewable Systems (ICEARS), users to novel content. 2023, pp. 1444–1449. doi:10.1109/ICEARS56392.
The applications of GANs in recommender systems are 2023.10085604. vast and promising. They enable the generation of per- [11] T. Sasagawa, S. Kawai, H. Nobuhara, Recommendasonalized recommendations, enhance diversity, tackle the tion system based on generative adversarial netcold-start problem, and address issues related to sparse work with graph convolutional layers, JACIII data. Leveraging deep learning and adversarial training, 25 (2021) 389–396. doi:10.20965/jaciii.2021. GANs extend the boundaries of recommendation sys- p0389. tems, empowering users to discover content that aligns [12] R. Yin, K. Li, J. Lu, G. Zhang, Rsygan: Generative with their evolving interests. adversarial network for recommender systems, in: 2019 International Joint Conference on Neural Networks (IJCNN), 2019, pp. 1–7. doi:10.1109/IJCNN.
References 2019.8851727.
[13] J. Wen, B.-Y. Chen, C.-D. Wang, Z. Tian, Prgan: [1] M. Gao, J. Zhang, J. Yu, J. Li, J. Wen, Q. Xiong, Rec- Personalized recommendation with conditional ommender systems based on generative adversarial generative adversarial networks, in: 2021 networks: A problem-driven perspective, Informa- IEEE International Conference on Data Mintion Sciences 546 (2021) 1166–1185. doi:10.1016/ ing (ICDM), 2021, pp. 729–738. doi:10.1109/ [2] jM..iAnls-G.2h0os2s0e.in0,9C.o0n1t3e.xt-aware recommender sys- [14] IEC.DDMe5rv1i6sh2a9j.,2P0.C21re.m00o0n8es4i., Gan-based matrix factems for real-world applications, Ph.D. thesis, Uni- torization for recommender systems, in: Proceedversité Paris Saclay (COmUE), 2019. ings of the 37th ACM/SIGAPP Symposium on Ap[3] A. Banerjee, P. Banik, W. Wörndl, A review on plied Computing, SAC ’22, Association for Comindividual and multistakeholder fairness in tourism puting Machinery, New York, NY, USA, 2022, pp. recommender systems, Front Big Data 6 (2023) 1373–1381. doi:10.1145/3477314.3507099.
1168692. doi:10.3389/fdata.2023.1168692. [15] X. Song, J. Qin, Q. Ren, J. Zheng, Igan: A collabora[4] D. Kotkov, J. A. Konstan, Q. Zhao, J. Veijalainen, tive filtering model based on improved generative Investigating serendipity in recommender systems adversarial networks for recommendation, Engibased on real user feedback, in: Proceedings of the neering Applications of Artificial Intelligence 124 33rd Annual ACM Symposium on Applied Computing, SAC ’18, Association for Computing Ma- (2023) 106569. doi:10.1016/j.engappai.2023. chinery, New York, NY, USA, 2018, pp. 1341–1350. 106569.
doi:10.1145/3167132.3167276. [5] I. Goodfellow, et al., Generative adversarial nets, in: Advances in Neural Information Processing Systems, Curran Associates, Inc., 2014. [6] D. Saxena, J. Cao, Generative adversarial networks (gans): Challenges, solutions, and future directions, ACM Comput. Surv. 54 (2022) 1–42. doi:10.1145/ 3446374. [7] S. Kumar, M. D. Gupta, +GAN: Complementary fashion item recommendation, arXiv (2019). doi:10.
48550/arXiv.1906.05596. [8] W. Shafqat, Y.-C. Byun, A hybrid gan-based ap