Transformer architectures, capable of capturing sequential dependencies in the history of user interactions, have become the dominant approach in sequential recommender systems. Despite their success, such models consider sequence elements in isolation, implicitly accounting for the complex relationships between them. Graph neural networks, in contrast, explicitly model these relationships through higher order interactions but are often unable to adequately capture their evolution over time, limiting their use for predicting the next interaction. To fill this gap, we present a new framework that combines transformers and graph neural networks and aligns different representations for solving next-item prediction task. Our solution simultaneously encodes structural dependencies in the interaction graph and tracks their dynamic change. Experimental results on a number of open datasets demonstrate that the proposed framework consistently outperforms both pure sequential and graph approaches in terms of recommendation quality, as well as recent methods that combine both types of signals.

Offline evaluation of recommender systems is often affected by hidden, under-documented choices in data preparation. Seemingly minor decisions in filtering, handling repeats, cold-start treatment, and splitting strategy design can substantially reorder model rankings and undermine reproducibility and cross-paper comparability. In this paper, we introduce SplitLight, an open-source exploratory toolkit that enables researchers and practitioners designing preprocessing and splitting pipelines or reviewing external artifacts to make these decisions measurable, comparable, and reportable. Given an interaction log and derived split subsets, SplitLight analyzes core and temporal dataset statistics, characterizes repeat consumption patterns and timestamp anomalies, and diagnoses split validity, including temporal leakage, cold-user/item exposure, and distribution shifts. SplitLight further allows side-by-side comparison of alternative splitting strategies through comprehensive aggregated summaries and interactive visualizations. Delivered as both a Python toolkit and an interactive no-code interface, SplitLight produces audit summaries that justify evaluation protocols and support transparent, reliable, and comparable experimentation in recommender systems research and industry.

The ranking of recommendation algorithms is a challenging problem since model performance is sensitive to dataset characteristics such as sparsity, sequential structure, and scale. This drives a demand for a proper methodology for fair comparison between algorithms. Naive aggregation of performance metrics (e.g., averaging NDCG over benchmarks) can yield misleading rankings, undermining practical selection. To address this problem, we introduce a novel, data-driven ranking methodology based on Bradley-Terry (BT) model. We demonstrate that the obtained ranking depends on key dataset statistics. Additionally, we propose a novel metric for evaluating ranking consistency and demonstrate robustness of our ranking to incomplete data. Finally, we introduce a dataset-specific methodology for ranking algorithms on unseen datasets without running the models, relying on extensions of the Bradley-Terry framework, including BT trees and BT models with covariates.

The growing demand for energy-efficient, high-performance AI systems has led to increased attention on alternative computing platforms (e.g., photonic, neuromorphic) due to their potential to accelerate learning and inference. However, integrating such physical components into deep learning pipelines remains challenging, as physical devices often offer limited expressiveness, and their non-differentiable nature renders on-device backpropagation difficult or infeasible. This motivates the development of hybrid architectures that combine digital neural networks with reconfigurable physical layers, which effectively behave as black boxes. In this work, we present a framework for the end-to-end training of such hybrid networks. This framework integrates stochastic zeroth-order optimization for updating the physical layer’s internal parameters with a dynamic low-rank surrogate model that enables gradient propagation through the physical layer. A key component of our approach is the implicit projector-splitting integrator algorithm, which updates the lightweight surrogate model after each forward pass with minimal hardware queries, thereby avoiding costly full matrix reconstruction. We demonstrate our method across diverse deep learning tasks, including: computer vision (deep convolutional network on the CIFAR-10 dataset), audio classification (ECAPA-TDNN model on the UrbanSound8K dataset), and language modeling (a decoder-only transformer model on the FineWeb corpus). Notably, across all modalities, the proposed approach achieves near-digital baseline accuracy and consistently enables effective end-to-end training of hybrid models incorporating various non-differentiable physical components (spatial light modulators, microring resonators, and Mach-Zehnder interferometers). This work bridges hardware-aware deep learning and gradient-free optimization, thereby offering a practical pathway for integrating non-differentiable physical components into scalable, end-to-end trainable AI systems.

Sequential self-attention models usually rely on additive positional embeddings, which inject positional information into item representations at the input. In the absence of positional signals, the attention block is permutation-equivariant over sequence positions and thus has no intrinsic notion of temporal order beyond causal masking. We argue that additive positional embeddings make the attention mechanism only superficially sensitive to sequence order: positional information is entangled with item embedding semantics, propagates weakly in deep architectures, and limits the ability to capture rich sequential patterns. To address these limitations, we introduce a kernelized self-attention mechanism, where a learnable positional kernel operates purely in the position space, disentangled from semantic similarity, and directly modulates attention weights. When applied per attention block, this kernel enables adaptive multi-scale sequential modeling. Experiments on standard next-item prediction benchmarks show that our positional kernel attention consistently improves over strong competing baselines.

Cross-domain recommendation systems (CDR) leverage knowledge from multiple source domains to address data sparsity in collaborative filtering. Meanwhile, Graph Convolutional Networks have achieved significant progress for single-domain recommendations, capturing deep relationships between users and items by propagating through their interaction graph. In this research, we extend UltraGCN, a recent graph-based recommendation model that approximates infinite-layer graph convolutions through constraint loss optimization, to optimize an arbitrary number of constraints from various relation graphs. This new feature enables UltraGCN to operate in a cross-domain setting by transferring user relations across domains and introducing a new constraint derived from the user–user graph in the source domain. Moreover, we modify the implementation of UltraGCN to compute side constraints (user–user, item–item) more efficiently to accommodate various computational infrastructures. Experiments on several real-world datasets demonstrate that our model, named UltraGCDR , surpasses recent cross-domain and single-domain models. The proposed model benefits from the inherited computational efficiency of UltraGCN with an asynchronous setup; i.e. constraints are transferred once before training begins, then the model optimizes locally. These features make our model highly suitable for industrial applications, especially when transitioning from single-domain to cross-domain environments. Finally, we provide the source code of our model and the experimental setup to ensure the reproducibility of the results and facilitate future contributions.

Sequential recommendation models must navigate sparse interaction data popularity bias and conflicting objectives like accuracy versus diversity While recent contrastive selfsupervised learning SSL methods offer improved accuracy they come with tradeoffs large batch requirements reliance on handcrafted augmentations and negative sampling that can reinforce popularity bias In this paper we introduce BT-SR a novel noncontrastive SSL framework that integrates the Barlow Twins redundancyreduction principle into a Transformerbased nextitem recommender BTSR learns embeddings that align users with similar shortterm behaviors while preserving longterm distinctionswithout requiring negative sampling or artificial perturbations This structuresensitive alignment allows BT-SR to more effectively recognize emerging user intent and mitigate the influence of noisy historical context Our experiments on five public benchmarks demonstrate that BTSR consistently improves nextitem prediction accuracy and significantly enhances longtail item coverage and recommendation calibration Crucially we show that a single hyperparameter can control the accuracydiversity tradeoff enabling practitioners to adapt recommendations to specific application needs

In this work, we present a fast and effective Linear approach for updating recommendations in a scalable graph-based recommender system UltraGCN. Solving this task is extremely important to maintain the relevance of the recommendations under the conditions of a large amount of new data and changing user preferences. To address this issue, we adapt the simple yet effective low-rank approximation approach to the graph-based model. Our method delivers instantaneous recommendations that are up to 30 times faster than conventional methods, with gains in recommendation quality, and demonstrates high scalability even on the large catalogue datasets.

Cold-start challenges in recommender systems necessitate leveraging auxiliary features beyond user-item interactions. However, the presence of irrelevant or noisy features can degrade predictive performance, whereas an excessive number of features increases computational demands, leading to higher memory consumption and prolonged training times. To address this, we propose a feature selection strategy that prioritizes the user behavioral information. Our method enhances the feature representation by incorporating correlations from collaborative behavior data using a hybrid matrix factorization technique and then ranks features using a mechanism based on the maximum volume algorithm. This approach identifies the most influential features, striking a balance between recommendation accuracy and computational efficiency. We conduct an extensive evaluation across various datasets and hybrid recommendation models, demonstrating that our method excels in cold-start scenarios by selecting minimal yet highly effective feature subsets. Even under strict feature reduction, our approach surpasses existing feature selection techniques while maintaining superior efficiency.

Retrieval-augmented generation has emerged as one of the most effective approaches for code completion enhancement, especially when repository-level context is important. However, adding this extra retrieved context significantly increases sequence length, raises prefill cost, and degrades time-to-first-token (TTFT), which slows down inference – a critical limitation for interactive settings such as IDEs. In this work, we introduce CoRoVA, a framework that compresses context into compact, semantically rich representations that remain interpretable to code LLMs. This improves generation quality while reducing prompt augmentation to only a few compressed single-token vectors. Our approach requires training only a small projector module and introduces negligible additional latency, yet it significantly improves the prediction quality of code LLMs. Our experiments show that CoRoVA enables a 20-38% reduction in TTFT on completion tasks compared to uncompressed RAG.

We consider the task of learning from both positive and negative feedback in a sequential recommendation scenario, as both types of feedback are often present in user interactions. Meanwhile, conventional sequential learning models usually focus on considering and predicting positive interactions, ignoring that reducing items with negative feedback in recommendations improves user satisfaction with the service. Moreover, the negative feedback can potentially provide a useful signal for more accurate identification of true user interests. In this work, we propose to train two transformer encoders on separate positive and negative interaction sequences. We incorporate both types of feedback into the training objective of the sequential recommender using a composite loss function that includes positive and negative cross-entropy as well as a cleverly crafted contrastive term, that helps better modeling opposing patterns. We demonstrate the effectiveness of this approach in terms of increasing true-positive metrics compared to state-of-the-art sequential recommendation methods while reducing the number of wrongly promoted negative items.

In modern recommender systems, experimental settings typically include filtering out cold users and items based on a minimum interaction threshold. However, these thresholds are often chosen arbitrarily and vary widely across studies, leading to inconsistencies that can significantly affect the comparability and reliability of evaluation results. In this paper, we systematically explore the cold-start boundary by examining the criteria used to determine whether a user or an item should be considered cold. Our experiments incrementally vary the number of interactions for different items during training, and gradually update the length of user interaction histories during inference. We investigate the thresholds across several widely used datasets, commonly represented in recent papers from top-tier conferences, and on multiple established recommender baselines. Our findings show that inconsistent selection of cold-start thresholds can either result in the unnecessary removal of valuable data or lead to the misclassification of cold instances as warm, introducing more noise into the system.

In this paper, we propose a new geometric approach for knowledge graph completion via low rank tensor approximation. We augment a pretrained and well-established Euclidean model based on a Tucker tensor decomposition with a novel hyperbolic interaction term. This correction enables more nuanced capturing of distributional properties in data better aligned with real-world knowledge graphs. By combining two geometries together, our approach improves expressivity of the resulting model achieving new state-of-the-art link prediction accuracy with a significantly lower number of parameters compared to the previous Euclidean and hyperbolic models.

Graph-based recommender systems have emerged as a powerful paradigm for personalized recommendations. However, their reliance on full model retraining to incorporate new users or new interactions creates scalability barriers. The task becomes infeasible in real-life recommender systems due to excessive time and resource costs involved. To address this limitation, we propose a fast and efficient method for updating graph-based recommender models without full model retraining on new data. Instead of changing all weights, we modify only small share of user representations who have new interactions. Our approach achieves a remarkable speedup of 700x over conventional model retraining approaches, drastically reducing computational overhead while maintaining the accuracy of the recommendations. Furthermore, we integrate our method into a multi-representation architecture that combines graph- and sequential-based methods to capture different user and item representations. Extensive experiments on diverse datasets demonstrate that our approach achieves state-of-the-art recommendation accuracy while maintaining the efficiency of incremental updates, outperforming existing methods in both speed and quality.

Modern sequential recommender systems, ranging from lightweight transformer-based variants to large language models, have become increasingly prominent in academia and industry due to their strong performance in the next-item prediction task. Yet common evaluation protocols for sequential recommendations remain insufficiently developed: they often fail to reflect the corresponding recommendation task accurately, or are not aligned with real-world scenarios. Although the widely used leave-one-out split matches next-item prediction, it permits the overlap between training and test periods, which leads to temporal leakage and unrealistically long test horizon, limiting real-world relevance. Global temporal splitting addresses these issues by evaluating on distinct future periods. However, its applications to sequential recommendations remain loosely defined, particularly in terms of selecting target interactions and constructing a validation subset that provides necessary consistency between validation and test metrics. In this paper, we demonstrate that evaluation outcomes can vary significantly across splitting strategies, influencing model rankings and practical deployment decisions. To improve reproducibility in both academic and industrial settings, we systematically compare different splitting strategies for sequential recommendations across multiple datasets and established baselines. Our findings show that prevalent splits, such as leave-one-out, may be insufficiently aligned with more realistic evaluation strategies.

Adaptive gradient methods like Adagrad and its variants are widespread in large-scale optimization. However, their use of diagonal preconditioning matrices limits the ability to capture parameter correlations. Full-matrix adaptive methods, approximating the exact Hessian, can model these correlations and may enable faster convergence. At the same time, their computational and memory costs are often prohibitive for large-scale models. To address this limitation, we propose AdaGram, an optimizer that enables efficient full-matrix adaptive gradient updates. To reduce memory and computational overhead, we utilize fast symmetric factorization for computing the preconditioned update direction at each iteration. Additionally, we maintain the low-rank structure of a preconditioner along the optimization trajectory using matrix integrator methods. Numerical experiments on standard machine learning tasks show that AdaGram converges faster or matches the performance of diagonal adaptive optimizers when using rank five and smaller rank approximations. This demonstrates AdaGram's potential as a scalable solution for adaptive optimization in large models.

Linear contextual bandits, especially LinUCB, are widely used in recommender systems. However, its training, inference, and memory costs grow with feature dimensionality and the size of the action space. The key bottleneck becomes the need to update, invert and store a design matrix that absorbs contextual information from interaction history. In this paper, we introduce Scalable LinUCB, the algorithm that enables fast and memory efficient operations with the inverse regularized design matrix. We achieve this through a dynamical low-rank parametrization of its inverse Cholesky-style factors. We derive numerically stable rank-1 and batched updates that maintain the inverse without directly forming the entire matrix. To control memory growth, we employ a projector-splitting integrator for dynamical low-rank approximation, yielding average per-step update cost $O(dr)$ and memory $O(dr)$ for approximation rank $r$. Inference complexity of the suggested algorithm is $O(dr)$ per action evaluation. Experiments on recommender system datasets demonstrate the effectiveness of our algorithm.

Recent studies have demonstrated the potential of hyperbolic geometry for capturing complex patterns from interaction data in recommender systems. In this work, we introduce a novel hyperbolic recommendation model that uses geometrical insights to improve representation learning and increase computational stability at the same time. We reformulate the notion of hyperbolic distances to unlock additional representation capacity over conventional Euclidean space and learn more expressive user and item representations. To better capture user-items interactions, we construct a triplet loss that models ternary relations between users and their corresponding preferred and nonpreferred choices through a mix of pairwise interaction terms driven by the geometry of data. Our hyperbolic approach not only outperforms existing Euclidean and hyperbolic models but also reduces popularity bias, leading to more diverse and personalized recommendations.

In recent years, self-attentive sequential learning models have surpassed conventional collaborative filtering techniques in next-item recommendation tasks. However, Euclidean geometry utilized in these models may not be optimal for capturing a complex structure of behavioral data. Building on recent advances in the application of hyperbolic geometry to collaborative filtering tasks, we propose a novel approach that leverages hyperbolic geometry in the sequential learning setting. Our approach replaces final output of the Euclidean models with a linear predictor in the non-linear hyperbolic space, which increases the representational capacity and improves recommendation quality.

Scalability issue plays a crucial role in productionizing modern recommender systems. Even lightweight architectures may suffer from high computational overload due to intermediate calculations, limiting their practicality in real-world applications. Specifically, applying full Cross-Entropy (CE) loss often yields state-of-the-art performance in terms of recommendations quality. Still, it suffers from excessive GPU memory utilization when dealing with large item catalogs. This paper introduces a novel Scalable Cross-Entropy (SCE) loss function in the sequential learning setup. It approximates the CE loss for datasets with large-size catalogs, enhancing both time efficiency and memory usage without compromising recommendations quality. Unlike traditional negative sampling methods, our approach utilizes a selective GPU-efficient computation strategy, focusing on the most informative elements of the catalog, particularly those most likely to be false positives. This is achieved by approximating the softmax distribution over a subset of the model outputs through the maximum inner product search. Experimental results on multiple datasets demonstrate the effectiveness of SCE in reducing peak memory usage by a factor of up to 100 compared to the alternatives, retaining or even exceeding their metrics values. The proposed approach also opens new perspectives for large-scale developments in different domains, such as large language models.