Addressing Stability in Classifier Explanations
Why this work is in the frame
A frame that forgets how it found something cannot be audited. These are the routes that admitted this work.
Bibliographic record
Abstract
Machine learning based classifiers are often a black box when considering the contribution of inputs to the output probability of a label, especially with complex non-linear models such as neural networks. A popular way to explain machine learning model outputs in a model agnostic manner is through the use of Shapley values. For our use case of abuse fighting in digital advertisements, one primary impediment of using Shapley values in explanations was a problem of instability. Specifically, the instability problem manifests as explanations for the same example varying greatly due to random sampling in the algorithm. We found it useful to view this problem explicitly as Monte Carlo integration in the form of averaging the model output while varying only a subset of features in the example to be explained. In turn, this guides the number of samples needed to achieve a stable estimate of individual Shapley values and unlocked the use of Shapley value based explainers for our models as well as classifiers in general, including neural networks.
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.
Full frame distilled prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. Learned from the 10,348 direct Codex labels and 10,348 direct Gemma labels. Candidate is the union of thresholded teacher heads; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels or direct frontier model labels.
Codex and Gemma teacher scores by category
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.001 | 0.002 |
| Meta-epidemiology (narrow) | 0.000 | 0.000 |
| Meta-epidemiology (broad) | 0.000 | 0.000 |
| Bibliometrics | 0.000 | 0.001 |
| Science and technology studies | 0.000 | 0.000 |
| Scholarly communication | 0.001 | 0.003 |
| Open science | 0.008 | 0.003 |
| Research integrity | 0.000 | 0.001 |
| Insufficient payload (model declined to judge) | 0.001 | 0.001 |
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 it