Aug 2026· Proceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.2· 0 citations· 9 references
TL;DR
Concept-Residual eXpansion (CRX), a concept-augmented framework that improves robustness by expanding the set of candidate predictive features by improving robustness to spurious correlations, is proposed.
Abstract
Models trained with empirical risk minimization (ERM) are prone to relying on spurious correlations to make predictions. A spurious correlation is a non-causal relationship in the training data between an attribute and the prediction target that does not generalize beyond the training environment. As a result, models can appear to achieve strong performance by exploiting these correlations, yet fail when the correlation changes or disappears. Despite their tendency to learn spurious correlations, the success of post-hoc mitigation methods in recent work suggests that ERM-trained models still retain useful, robust predictive features. However, core (non-spurious) features may be weak or entangled within the representation, making them difficult to identify. We propose Concept-Residual eXpansion (CRX), a concept-augmented framework that improves robustness by expanding the set of candidate predictive features. Starting from a frozen ERM representation, we augment the model's features with interpretable concept scores that describe the presence of task-relevant attributes and the surrounding context, together with residual features that capture the portion of the ERM features not expressed by the concepts. We then retrain a lightweight classifier on this expanded feature space, enabling it to leverage both structured semantic cues and complementary residual information. Across standard spurious correlation benchmarks, CRX consistently improves worst-group accuracy while maintaining competitive average performance. These results suggest that expanding the set of available features can substantially improve robustness to spurious correlations. The code and additional implementation details can be found at https://doi.org/10.5281/zenodo.20467988
It is shown that a usable signal is available after convergence, when loss no longer distinguishes the two populations, and applying a fixed perturbation to a converged model's inputs flips the predictions of the latter far more often than the former.
The performance of Large Language Models (LLMs) is fundamentally influenced by the distributional composition of multi-domain pre-training data. While manual heuristics were prevalent in early models, they increasingly fail to capture the intricate synergies between domains as data complexity grows. To overcome the issue, a dominant approach seeks to fit a proxy function mapping between domain weights and their corresponding validation losses, and then find the optimal domain weights to minimize validation losses. These methods rely on strong structural assumptions, such as rank invariance or scaling laws, which are often violated, resulting in non-negligible estimation bias. A promising approach is to directly optimize the weighting scheme from data. However, it suffers from unstable optimization trajectory and prohibitive computational overhead, limiting its potential to search better domain weights configurations. This paper presents a Bayesian domain weighting method to infer the weights from a Dirichlet distribution via introducing Gamma prior information learned from observations. Experimental results demonstrate that proposed method could achieve stable and efficient domain weights learning, and identifies optimal mixtures while consuming substantially less data than search-based function-fitting methods, revitalizing optimization-based domain weighting for large-scale applications.
Xiang Yuan, Kaiqing Lei, Zhenyu Jin et al.· 0 citations
The Normalised Sensitivity Ratio (NSR) is introduced, a post-hoc, model-agnostic diagnostic for this question under a structured-shift regime: environments differ primarily in the mean of spurious features while the causal mechanism and causal marginals remain stable, as in multi-site clinical data or multi-batch genomics.
Athanasios Vlontzos, G. Papanastasiou, Bernhard Kainz et al.· 0 citations
Over the last few years, there has been an increased interest in making machine learning models more interpretable. Although a great deal of effort goes into developing techniques for interpreting the interactions learned by a given model, fewer studies focus on assessing the quality of such explanations. Even fewer focus on how to adjust the model to produce explanations faithful to prior knowledge, a process known as explanation-guided learning. Furthermore, most approaches in this area focus on classification problems and usually assume prior knowledge about which input features or regions are most important. In this work, we introduce a new approach to steering neural networks based on partial dependence, such that their average response to certain features aligns with specific functional domain knowledge about the problem. We empirically demonstrate on a range of regression problems, including dynamical systems forecasting, that models whose training has been controlled using our method perform better than unconstrained models and are more data-efficient. Moreover, we highlight that interpretations obtained from the former actually align with the user-provided knowledge, whereas those obtained from the latter do not.
Yann Claes, P. Geurts, V. A. Huynh-Thu· 0 citations
A novel Ante-hoc Explainable AI framework designed to bridge the interpretability-accuracy trade-off in high-stakes financial prognosis, specifically within credit scoring systems, which provides a verifiable and robust solution for modern, regulatory-compliant financial environments.
Deniz NoorMohammadzadehMaleki, Mahdi Baghaei Oskouei, Alireza Taheri et al.· AI and Ethics· 0 citations
Prediction-based methods, including Large Language Models (LLMs) and other machine learning techniques, are often used to construct measures of political phenomena that are difficult to quantify directly, such as policy positions in manifestos or emotions expressed on social media. In many applications, these prediction-generated measures are used as explanatory variables in regression models, even though they are measured with error. This leads to biased estimates. In this paper, we propose a simple solution to these biases: instrumental variables constructed from multiple measures created on independent splits of the original data. This approach is theoretically valid, easy to implement, and does not require new data. Through simulations, we show that this approach recovers estimates close to the true values, even in relatively small samples, while the standard approach can produce substantial bias in practice. We illustrate the method by revisiting two applications: whether gendered speech affects legislative outcomes in the German Parliament, and whether political risk influences poverty alleviation programs in China.