In cross-domain recommendation, the cold-start recommendation problem often arises in scenarios where users have interacted with items in a source domain but not in a target domain. A key challenge in this cross-domain recommendation setting is how to effectively transfer user preferences from the source domain to the target domain. Most existing transfer learning models address this challenge but typically require extensive computations and incremental operations, which limit their scalability and efficiency. To overcome these limitations, we propose a novel similarity-based framework, called Similarity-based Transfer Graph Convolution Network (SimTranGCN), designed specifically for cold-start users. Our approach combines item-KNN, deep learning, and graph convolutional models such as LightGCN. SimTranGCN first constructs a similarity matrix across domains, and then uses this matrix to infer user preferences in the target domain based on their interactions in the source domain. Empirical experiments demonstrate that SimTranGCN is highly competitive against existing methods, achieving state-of-the-art performance on two paired domain transfer tasks.

Reinforcement learning (RL) has demonstrated great potential to improve slate-based recommender systems by optimizing long-term user engagement. However, addressing the combinatorial action space in slate recommendations remains challenging. Recent work decomposes slate Q -values into item-wise Q -values, improving the tractability of value-based methods to learn the model. But in scenarios with a large item pool and a resource-intensive value function like deep neural networks, the action selection process still incurs substantial computational costs. Slow training might be tolerable, but high costs during action selection could hinder real-time deployment. To address this issue, this article introduces an actor method that reduces Q -function evaluations to a subset of items, significantly cutting inference time for practical deployment. The research suggests acquiring representations at both item and slate levels, strategically identifying a specific item subset for slate composition. The proposed methodologies are assessed over different simulated user engagement behaviors: users certain about preferences (“decisive” behavior) and those more exploratory or bored users, losing interest with repetitive content exposure (“explorative” behavior). Empirical evaluation shows that the proposed approach achieves comparable user engagement with a value-based policy across behaviors. Meanwhile, it notably enhances serving time while recommending diverse topic slates, thus demonstrating its potential effectiveness and efficiency in real-world applications.

Session-based slate recommendation systems curate ranked sets of items in real-time, adapting to evolving user interactions.Balancing relevance, diversity, and novelty remains challenging for reinforcement learning (RL) methods.Recent advances in large language models (LLMs) offer a new possibility to leverage their semantic reasoning capabilities to refine slate composition.In this work, we examine the impact of LLM-driven reasoning on slate generation by integrating LLMs with an RL-based slate recommender and evaluating in terms of accuracy, similarity, diversity, and novelty.We extend the RecSim framework with real-world interaction data and introduce a session-aware evaluation protocol that captures long-term engagement.Our analysis reveals that LLM reasoning enhances subcategory-level diversity while maintaining relevance, leading to increased user engagement.By visualizing category-level shifts in slate composition we uncover systematic patterns in how LLMs refine recommendation diversity.Although direct LLM use during inference may be hampered by computational demands and latency concerns, our experimental results demonstrate that integrating LLM modifications during training enables the model to internalize the nuanced characteristics of LLM reasoning without incurring inference overhead, thereby improving recommendation performance, serving time efficiency, and deployability.

Data privacy is a critical concern in today's data-driven world. To this end, Federated Learning (FL) has been researched extensively, as it allows sensitive data to be kept secure on local devices while training global Machine Learning (ML) models across multiple devices. Several FL topologies have been introduced to address different users' needs. FL fits well within the Recommender Systems (RS) domain, with decentralised large-scale datasets and user privacy issues, as it improves personalised recommendations and accuracy while keeping users' sensitive data secure.

Recommender systems have achieved remarkable success in various web applications, such as e-commerce, online advertising, and social media, harnessing the power of big data. To attain optimal model performance, recommender systems are typically trained on very large datasets, with substantial numbers of users and items. However, large datasets often present challenges in terms of processing time and computational resources. Coreset selection offers a method for obtaining a reduced yet representative subset from vast datasets, thereby enhancing the efficiency of training machine learning algorithms. Nevertheless, little research has been conducted to explore the practical implications of different coreset selection approaches on the performance of recommender systems algorithms. In this paper, we systematically investigate the impact of various coreset selection techniques. We evaluate the performance of the resulting coresets using inductive recommendation models which allow for consistent evaluations to be performed. The experimental results demonstrate that coreset methods are a powerful and useful approach for obtaining reduced datasets which preserve the properties of the large original dataset and have competitive performance compared to the time required to train with the full dataset.

Reinforcement learning (RL) has demonstrated great potential to improve slate-based recommender systems by optimizing recommendations for long-term user engagement. To handle the combinatorial action space in slate recommendation, recent works decompose the Q-value of a slate into item-wise Q-values, using an item-wise value-based policy. However, the common case where the value function is a parameterized function taking state and action as input results in a linearly increasing number of evaluations required to select an action, proportional to the number of candidate items. While slow training may be acceptable, this becomes intractable when considering the costly evaluation of the parameterized function, such as with deep neural networks, during model serving time. To address this issue, we propose an actor-based policy that reduces the evaluation of the Q-function to a subset of items, significantly reducing inference time and enabling practical deployment in real-world industrial settings. In our empirical evaluation, we demonstrate that our proposed approach achieves equivalent user session engagement to a value-based policy, while significantly reducing the slate serving time by at least 4 times.

In this paper, we propose a novel communication-efficient algorithm for distributed matrix factorisation. Our goal is to find a good trade-off between the communication overhead and the overall model training time. In our setting, the training data is distributed across multiple servers that aim to learn a joint machine learning model. In contrast to standard distributed computation, due to privacy concerns, the participating servers are not allowed to share raw data, however, sharing of the non-personal model parameters is allowed. We investigate the draw-backs of traditional strongly coordinated distributed techniques and compare them to weakly coordinated gossip approaches. The advantage of strongly coordinated approaches is that the learning process closely mimics that of a centralised algorithm and hence this approach can keep the overall training time at a minimum. However, this is at the expense of a large communication footprint of the algorithm. On the other hand, the weakly coordinated gossip approach offers a communication efficient solution that can take a large amount of training time to reach a good accuracy. As a solution, we develop a hybrid approach combining the above two approaches. We apply the hybrid approach on a latent factor model solving a top-N recommendation problem and we show that the hybrid approach achieves good accuracy in relatively short training time with minimal communication overhead particularly on very sparse data.

Federated learning (FL) is quickly becoming the de facto standard for the distributed training of deep recommendation models, using on-device user data and reducing server costs. In a typical FL process, a central server tasks end-users to train a shared recommendation model using their local data. The local models are trained over several rounds on the users' devices and the server combines them into a global model, which is sent to the devices for the purpose of providing recommendations. Standard FL approaches use randomly selected users for training at each round, and simply average their local models to compute the global model. The resulting federated recommendation models require significant client effort to train and many communication rounds before they converge to a satisfactory accuracy. Users are left with poor quality recommendations until the late stages of training. We present a novel technique, FedFast, to accelerate distributed learning which achieves good accuracy for all users very early in the training process. We achieve this by sampling from a diverse set of participating clients in each training round and applying an active aggregation method that propagates the updated model to the other clients. Consequently, with FedFast the users benefit from far lower communication costs and more accurate models that can be consumed anytime during the training process even at the very early stages. We demonstrate the efficacy of our approach across a variety of benchmark datasets and in comparison to state-of-the-art recommendation techniques.

Nowadays we commonly have multiple sources of data associated with items. Users may provide numerical ratings, or implicit interactions, but may also provide textual reviews. Although many algorithms have been proposed to jointly learn a model over both interactions and textual data, there is room to improve the many factorization models that are proven to work well on interactions data, but are not designed to exploit textual information. Our focus in this work is to propose a simple, yet easily applicable and effective, method to incorporate review data into such factorization models. In particular, we propose to build the user and item embeddings within the topic space of a topic model learned from the review data. This has several advantages: we observe that initializing the user and item embeddings in topic space leads to faster convergence of the factorization algorithm to a model that out-performs models initialized randomly, or with other state-of-the-art initialization strategies. Moreover, constraining user and item factors to topic space allows for the learning of an interpretable model that users can visualise.

Conventional approaches to matrix factorisation (MF) typically rely on a centralised collection of user data for building a MF model. This approach introduces an increased risk when it comes to user privacy. In this short paper we propose an alternative, user-centric, privacy enhanced, decentralised approach to MF. Our method pushes the computation of the recommendation model to the user's device, and eliminates the need to exchange sensitive personal information; instead only the loss gradients of local (device-based) MF models need to be shared. Moreover, users can select the amount and type of information to be shared, for enhanced privacy. We demonstrate the effectiveness of this approach by considering different levels of user privacy in comparison with state-of-the-art alternatives.

Recommender systems (RS) share many features and objectives with reinforcement learning (RL) systems. The former aim to maximise user satisfaction by recommending the right items to the right users at the right time, the latter maximise future rewards by selecting state-changing actions in some environment. The concept of an RL gym has become increasingly important when it comes to supporting the development of RL models. A gym provides a simulation environment in which to test and develop RL agents, providing a state model, actions, rewards/penalties etc. In this paper we describe and demonstrate the PyRecGym gym, which is specifically designed for the needs of recommender systems research, by supporting standard test datasets (MovieLens, Yelp etc.), common input types (text, numeric etc.), and thereby offering researchers a reproducible research environment to accelerate experimentation and development of RL in RS.