Disentangling the AI Black-Box Model: From Direct to Indirect Influence
Bibliographic record
Abstract
Due to the widespread use of artificial intelligence and machine learning models across numerous domains that directly impact individuals’ lives, it is essential to ensure that these models are making their decisions in a non-discriminatory manner. That is, biases encoded in the underlying training data are not reflected in the output of the model. This is, however, often hard to control since the predictive accuracy of these models come at the cost of interpretability. The purpose of this thesis is to investigate and apply methods for interpreting black-box machine learning models to bring more transparency into their decision-making process by analyzing both the direct and indirect influence of input features on model predictions. We begin by studying the theoretical background of the SHAP (SHapley Additive exPlanations) framework, Shapley values, and their implementation in measuring direct influence. We explore the SHAP Python package for both local and global explanations and visualization of direct feature influence on both synthetic and real-world datasets. We then transition to studying indirect influence using the Disentangled Influence Audit procedure, which uses adversarial training to learn disentangled latent representations for auditing hidden dependencies. After presenting all the background information, we implement this procedure and evaluate these methods through numerical experiments on synthetic functions and real-world datasets including the Adult Income and the Montréal Housing datasets. In addition to these experiments, we also examine how “data-hungry” these auditing methods are by examining their convergence as a function of the number of samples required. Our results demonstrate the importance of auditing both direct and indirect pathways of influence to promote interpretability and fairness in complex models.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.010 | 0.063 |
| Meta-epidemiology (narrow) | 0.001 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.001 |
| Science and technology studies | 0.001 | 0.004 |
| Scholarly communication | 0.003 | 0.006 |
| Open science | 0.002 | 0.003 |
| Research integrity | 0.002 | 0.004 |
| Insufficient payload (model declined to judge) | 0.004 | 0.000 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".