Search Results for « Black-box models »

The Fragile Families Challenge provided an opportunity to empirically assess the applicability of black box machine learning models to sociological questions and the extent to which interpretable explanations can be extracted from these models. In this paper I use neural network models to predict high school grade-point average and examine how variations of basic network parameters affect predictive performance. Using a recently proposed technique, I identify the most important predictive variables used by the best-performing model, finding that they relate to parenting and the child’s cognitive and behavioral development, consistent with prior work. I conclude by discussing the implications of these findings for the relationship between prediction and explanation in sociological analyses.

Abstract Explainable AI for tabular models underpins decisions in finance, healthcare, and policy, yet today’s explanations are dominated by heuristics without statistical guarantees. We introduce Stat-XAI, a model-agnostic framework that converts explanations into testable statistical statements. For each feature, Stat-XAI assesses association with model predictions on held-out data via appropriate hypothesis tests and reports standardized effect sizes (e.g., $\eta^2$, $R^2$, Cramér's $V$), yielding compact, uncertainty-aware rankings. Across six synthetic datasets with known causal structure and two real benchmarks, Stat-XAI delivers stable, parsimonious attributions, filters spurious correlates, and achieves orders-of-magnitude lower runtime than SHAP while maintaining faithfulness. We quantify stability under perturbations and show that interaction testing clarifies when pairwise dependencies meaningfully alter importance. By elevating explanation from heuristic scoring to inferential analysis, Stat-XAI provides a rigorous, reproducible pathway for trustworthy tabular AI—supporting scrutiny, governance, and human decision-making where reliability matters most.

Abstract Genome scale metabolic models provide detailed mechanistic descriptions of cellular metabolism, whereas black box growth models capture physiological behaviors using a small number of effective parameters. However, the quantitative relationship between these two modeling scales remains unclear. In this study, we develop a thermodynamic framework that connects black box growth models with thermodynamically constrained genome scale metabolic models. By linking black box model parameters with Gibbs energy dissipation rates derived from genome scale metabolic models, we demonstrate that coarse grained physiological descriptions can be obtained from detailed metabolic networks while preserving their thermodynamic foundation. We validate this framework in both Escherichia coli and yeast, showing that the resulting black box models reproduce key physiological behaviors, including growth dynamics, biomass yield, and overflow metabolism observed experimentally. Our results indicate that black box growth models and genome scale metabolic models are connected through shared thermodynamic constraints, revealing a consistent thermodynamic basis across different levels of metabolic description. This framework provides a general approach for integrating detailed metabolic networks with simple black box growth models and enables efficient multiscale modeling of cellular metabolism.

Black carbon (BC) is an important component of particulate matter (PM) in urban environments. BC is typically emitted from gas and diesel engines, coal-fired power plants, and other sources that burn fossil fuel. In contrast to PM, BC measurements are not always available on a large scale due to the operational cost and complexity of the instrumentation. Therefore, it is advantageous to develop a mathematical model for estimating the quantity of BC in the air, termed a BC proxy, to enable widening of spatial air pollution mapping. This article presents the development of BC proxies based on a Bayesian framework using measurements of PM concentrations and size distributions from 10 to 10,000 nm from a recent mobile air pollution study across several areas of Jordan. Bayesian methods using informative priors can naturally prevent over-fitting in the modelling process and the methods generate a confidence interval around the prediction, thus the estimated BC concentration can be directly quantified and assessed. In particular, two types of models are developed based on their transparency and interpretability, referred to as white-box and black-box models. The proposed methods are tested on extensive data sets obtained from the measurement campaign in Jordan. In this study, black-box models perform slightly better due to their model complexity. Nevertheless, the results demonstrate that the performance of both models does not differ significantly. In practice, white-box models are relatively more convenient to be deployed, the methods are well understood by scientists, and the models can be used to better understand key relationships.

Feature-effect curves for black-box predictors require repeated model evaluations and become difficult when inputs are dependent, high-dimensional,or available only through a noisy query interface. Accumulated Aggregated D-Optimal Designs (A2D2E) stabilize local slopes with a full 2p factorial hy-percube. This study replaces that hypercube with a folded Hadamard design requiring fewer than 4pevaluations per anchor. Unlike an ordinary Hadamard fraction, the paired construction cancels even-order curvature and retains the O(δ2) truncation order of the full design. A finite-sample error decompositionseparates binning, surrogate, curvature, sampling, and query-noise terms and motivates a support-constrained pilot rule. On an eight-dimensionalofficial benchmark, the proposed estimator incurs a paired ORMSE increase of 0.000543 while reducing queries 16-fold. In a 30-dimensional noisy-oracleexperiment with 60–64 final calls per anchor, folded Hadamard attains RMSE 0.02737 versus 0.15496 for coordinate differences at a common coordinate step,but 0.02737 versus 0.02825 at a common Euclidean radius. A support-matched pilot gives 0.00685 versus 0.00721. A sparse simulation and 100-split analysesof two public data sets further delimit accuracy, support, and computational tradeoffs. The method is therefore a scalable approximation to A2D2E forjoint finite-radius slope estimation, not a universal replacement for PD, ALE, or coordinate differences.

Abstract This study investigates the dispositional tendencies of large language models(LLMs) by developing and validating a survey instrument tailored to their uniquecharacteristics. While prior research has adapted human psychometric frameworkssuch as the Big Five to LLMs, these approaches face conceptual andmethodological limitations. To address this gap, this study introduced three traits(convergent-divergent reasoning, novel-conventional problem solving, and ethicalorientation) that more directly capture the functional dispositions of LLMs.A forced-choice questionnaire was co-designed with multiple models to assessthese traits, and its predictive validity was tested against open-ended behavioraltasks. Results demonstrate that questionnaire responses align with observed tendencies,with moderate-to-strong correlations for divergent and novel reasoning,and consistent expression of virtue ethics across models. These findings suggestthat LLMs exhibit measurable, context-dependent dispositions that can be systematicallyassessed. The proposed framework contributes to ongoing efforts tointerpret model behavior, inform prompt design, and advance the theoreticaldiscourse on personality-like constructs in artificial systems.