energy-based attribution

This study addresses the challenge of effective attribution in large-scale hybrid cyber-physical Internet-of-Things systems, where traditional causal explanation methods struggle due to their reliance on explicit directed graphs and difficulties handling feedback loops and partial observability. To overcome these limitations, this work proposes a statistical mechanics–inspired undirected energy-based modeling framework that captures dependency structures among variables and analyzes shifts in the energy landscape to enable structured attribution without reconstructing a causal graph. The approach introduces a novel energy-landscape–based dependency-aware mechanism capable of reasoning about perturbation effects in systems with mixed continuous-discrete variables. Experiments on an industrial IoT platform demonstrate that the method significantly outperforms state-of-the-art graph-based approaches in attribution accuracy, robustness, and scalability, making it well-suited for high-dimensional cyber-physical and socio-technical systems.

This study addresses the lack of formal definitions in existing root cause analysis methods, which are often limited to root nodes in causal graphs or biased toward proximate causes. Within the potential outcomes framework, this work proposes the first counterfactual definition of root cause at the individual level and introduces a probabilistic measure—Probability of Root Condition (PRC)—to quantify the likelihood that a candidate set of variables constitutes a root cause for a specific outcome. Under standard causal assumptions, the authors derive an explicit identification formula for PRC by integrating causal mediation analysis with counterfactual reasoning, thereby establishing its identifiability. The effectiveness and practical utility of the proposed approach are demonstrated through two numerical examples, filling a critical gap in the formal theory of root cause analysis.

The causal origins of interpretable units—such as induction heads—in large language models remain poorly understood. This work proposes a scalable mechanistic data attribution framework that integrates influence functions with causal interventions to establish, for the first time, direct causal links between specific training examples and the emergence of such interpretable components. The study reveals that structured repetitive data plays a catalytic role in circuit formation and demonstrates a direct functional relationship between induction heads and in-context learning capabilities. By selectively intervening on a small set of high-influence training samples, the emergence of attention heads can be significantly modulated. Furthermore, the proposed data augmentation strategy consistently accelerates circuit convergence across different model scales.

Existing methods for root cause analysis in time series anomaly detection often rely on unrealistic feature perturbations and neglect temporal dynamics and cross-variable dependencies, leading to unreliable attributions. This work proposes a conditional attribution framework that identifies root causes by retrieving normal-state contexts similar to the anomalous context while preserving dependency structures, thereby establishing a more faithful baseline for attribution. The approach innovatively combines the latent space of a variational autoencoder with UMAP manifold embeddings to enable efficient, high-fidelity low-dimensional context retrieval, effectively avoiding out-of-distribution artifacts. Furthermore, it incorporates confidence-aware mechanisms and temporal evaluation metrics to enhance attribution reliability. Evaluated on the SWaT and MSDS benchmarks, the method substantially outperforms existing approaches, achieving significant improvements in root cause identification accuracy, temporal localization precision, and cross-model robustness.

This work addresses the limitation of existing explanation methods, which focus solely on prediction accuracy and fail to capture the actual decision-making value of predictions within downstream optimization systems. To bridge this gap, the authors propose a Decision Value Attribution (DVA) framework that, for the first time, extends Shapley values to joint prediction-optimization pipelines. By formulating a cooperative game whose value function is defined by operational objectives, DVA attributes decision value to input features (InfoDVA), optimization design choices (DesignDVA), and their interactions (DVI). The framework further distinguishes between pre-DVA and post-DVA evaluation paradigms to assess the alignment between model beliefs and realized operational performance. Experiments on electricity storage arbitrage and emergency medical service coverage demonstrate that conventional prediction explanations often misrepresent true decision value, and that optimization configurations critically modulate the decision relevance of predictive information.

This work addresses a critical limitation in existing root cause analysis methods for anomalies: their failure to distinguish between two fundamentally distinct sources—measurement errors and mechanism shifts—often leading to misdiagnosis. To resolve this, the paper proposes the first causal framework that explicitly models both anomaly types by treating them as implicit interventions on latent “true” variables and observed “measured” variables. A structural causal model (SCM) with latent variables is constructed, and maximum likelihood estimation is employed to simultaneously classify anomaly types and localize root causes. Theoretically, the approach is shown to be identifiable without requiring prior knowledge of the causal graph structure. Empirical evaluations demonstrate state-of-the-art performance in root cause localization, accurate anomaly-type classification, and robustness even when the underlying causal graph is unknown.

This work addresses the challenge of root cause localization in time-varying dynamic systems with lag effects and memory properties—such as energy systems—where anomalies exhibit complex temporal dependencies. For the first time, the authors extend a strictly causal root cause analysis framework to such systems, introducing two truncation strategies to manage infinite-time dependency graphs: one preserving the original causal mechanisms and the other employing mechanism approximation. By integrating causal graph modeling with a data generation approach tailored to energy consumption peak scenarios, the proposed method is evaluated in a simulated factory environment. Results demonstrate that, given sufficient lag order, the approach accurately identifies the spatiotemporal origins of anomalies, while also quantifying the performance trade-offs introduced by mechanism approximation.

This work addresses the limitation of existing methods that can either detect out-of-distribution samples or quantify uncertainty but struggle to pinpoint the specific causes of model failure. The authors propose a self-diagnosing model that jointly learns structured failure attribution signals alongside its primary predictions, extending scalar uncertainty into an attribution vector capable of distinguishing among four distinct failure modes: covariate shift, semantic shift, noisy corruption, and adversarial perturbations. The attribution vector is generated by a neural network and constrained via a consistency regularization term that aligns uncertainty estimates with attribution predictions. To evaluate the approach, the authors construct a benchmark dataset incorporating predefined shift mechanisms. Experimental results demonstrate that the method not only effectively detects anomalies but also accurately attributes their underlying failure types, significantly enhancing model interpretability and robustness.