Abstract
In streaming scenarios, models must learn continuously, adapting to concept drifts without erasing previously acquired knowledge. However, existing research communities address these challenges in isolation. Continual Learning (CL) focuses on long-term retention and mitigating catastrophic forgetting, often without strict real-time constraints. Stream Learning (SL) emphasizes rapid, efficient adaptation to high-frequency data streams, but typically neglects forgetting. Recent efforts have tried to combine these paradigms, yet no clear algorithmic overlap exists. We argue that large in-context tabular models (LTMs) provide a natural bridge for Streaming Continual Learning (SCL). In our view, unbounded streams should be summarized on-the-fly into compact sketches that can be consumed by LTMs. This recovers the classical SL motivation of compressing massive streams with fixed-size guarantees, while simultaneously aligning with the experience-replay desiderata of CL. To clarify this bridge, we show how the SL and CL communities implicitly adopt a divide-to-conquer strategy to manage the tension between plasticity (performing well on the current distribution) and stability (retaining past knowledge), while also imposing a minimal complexity constraint that motivates diversification (avoiding redundancy in what is stored) and retrieval (re-prioritizing past information when needed). Within this perspective, we propose structuring SCL with LTMs around two core principles of data selection for in-context learning: (1) distribution matching, which balances plasticity and stability, and (2) distribution compression, which controls memory size through diversification and retrieval mechanisms. © 2026 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).

Abstract
Forecasting the covariance matrix of asset returns is central to portfolio construction, risk management, and asset pricing. However, most existing models struggle at medium-term horizons, several weeks to months, where shifting market regimes and slower dynamics prevail. We propose a novel deep learning framework that integrates three-dimensional convolutional neural networks, bidirectional long short-term memory, and multihead attention to capture complex spatiotemporal patterns in asset return dynamics. Using daily data on 14 exchange-traded funds from 2017 to 2023, we demonstrate that our model improves out-of-sample covariance forecasts by reducing Euclidean and Frobenius distance metrics by up to 20% compared with classical benchmarks such as shrinkage estimators and GARCH-type models. These gains persist across distinct market regimes, including bull and bear periods, and remain robust across various forecast horizons and under both raw and excess return specifications. Portfolio simulations based on global minimum variance strategies reveal that the proposed model consistently delivers lower volatility and moderate turnover, even under no-short-selling constraints. This balance between risk reduction and trading efficiency underscores the economic relevance of the forecasts, particularly for institutional investors managing portfolios at medium-term horizons.

Abstract

Abstract
Illicit activities in the waste management network, such as waste laundering, misreporting, or trade of stolen waste pose serious environmental and regulatory challenges. Detecting these behaviours is challenging, because they often emerge from higher-order interactions among multiple entities, and are not continuous over time. Furthermore, these activities often manifest as triangles in the network, and the participation of individuals in these waste transfer structures is additionally suspicious. Traditional anomaly detection methods, which rely on first-order relationships or static analyses, struggle to capture these complex, temporally dynamic patterns. To address this challenge, we propose a Conditional Motif-Based Graph Convolutional Network (CM-GCN) that integrates condition-driven triangular motifs directly into the GCN message-passing mechanism. The CM–GCN learns structural embeddings that encode both local graph topology and node attributes–based connectivity to triangular motifs. To detect sudden or sporadic changes, these weekly embeddings are processed by a Long Short–Term Memory Variational Autoencoder (LSTM–VAE), which models temporal behaviour and identifies anomalies through spikes in reconstruction error. Experiments on one year of Portuguese waste transport data demonstrate that the proposed approach effectively highlights companies with known illicit behaviour. The CM–GCN–LSTM–VAE outperformed a standard GCN–LSTM–VAE that ignores motif structure. Results are comparable to, and slightly improve upon, an LSTM–VAE trained on a manually engineered triangle–based feature. This demonstrates that higher–order structural representations learned by the model provide a more informative signal, while simple pairwise relationships contribute little to the detection of complex behaviours.

Abstract
The typical offline protocol to evaluate recommendation algorithms is to collect a dataset of user-item interactions and then use a part of this dataset to train a model, and the remaining data to measure how closely the model recommendations match the observed user interactions. This protocol is straightforward, useful and practical, but it only provides snapshot performance. We know, however, that online systems evolve over time. In general, it is a good idea that models are frequently retrained with recent data. But if this is the case, to what extent can we trust previous evaluations? How will a model perform when a different pattern (re)emerges? In this paper we propose a methodology to study how recommendation models behave when they are retrained. The idea is to profile algorithms according to their ability to, on the one hand, retain past patterns - stability - and, on the other hand, (quickly) adapt to changes - plasticity. We devise an offline evaluation protocol that provides detail on the long-term behavior of models, and that is agnostic to datasets, algorithms and metrics. To illustrate the potential of this framework, we present preliminary results of three different types of algorithms on the GoodReads dataset that suggest different stability and plasticity profiles depending on the algorithmic technique, and a possible trade-off between stability and plasticity. We further discuss the potential and limitations of the proposal and advance some possible improvements. © 2026 Copyright held by the owner/author(s).

Abstract
Counterfactual explanations (CFs) help users understand and act on black-box machine learning decisions by suggesting minimal changes to achieve a desired outcome. However, existing methods often ignore individual feasibility, leading to unrealistic or unactionable recommendations. We propose a personalized CF generation method based on cluster-specific fine-tuning of Generative Adversarial Networks (GANs). By grouping users with similar behavior and constraints, we adapt immutable features and cost weights per cluster, allowing GANs to generate more actionable and user-aligned counterfactuals. Experiments on the German Credit dataset show that our approach achieves a 6x improvement in prediction gain and a 30% reduction in sparsity compared to a baseline CounterGAN, while maintaining plausibility and acceptable latency for online use.