Source-linked AI summary
ProToMEx: Rapid, Interpretable Explanations via Structured Representations
Athina Georgara, Adarsh Valoor, Sarvapali D. Ramchurn
TL;DR
Feature-attribution explainers can miss how combinations of features form coherent reasons for a classifier's predictions. ProToMEx uses probabilistic topic models to represent those patterns and generate global and local explanations; its online explanations are reported faster than SHAP and LIME.
Problem
Feature-attribution methods can identify important individual features but may not reveal the coherent combinatorial logic behind a classifier's decision.
Source: §1 · p. 1Method
ProToMEx converts classifier instance-label pairs into weighted documents and uses a topic model to represent feature-value patterns as reasons for predictions.
Source: §1 · p. 2, Document Transformation · p. 4Results
Online explanations were reported at least 30-40x faster than SHAP and 20x faster than LIME, with explanation quality described as similar to SHAP.
Source: §5.3 · p. 9, §6 · p. 12Takeaways & Limitations
For repeated explanations on public datasets, the initial time investment is offset, and ProToMEx becomes the most time-efficient method after 2600 instances.
Source: §5.3 · p. 9Evidence that ProToMEx identifies influential features missed by SHAP is explicitly described as nonconclusive.
Source: §5.3 · p. 11
Abstract
from arXiv · showhide
Existing post-hoc explainers for machine learning classifiers primarily focus on feature attribution, assigning importance scores to individual features. While valuable, this approach struggles to articulate the complex, combinatorial patterns that often drive a model's decision-making process. To overcome this limitation, we introduce ProToMEx, a new paradigm for explainability that leverages Probabilistic Topic Models (PTMs). Our model-agnostic framework learns latent ''topics'' that represent distinct, high-level reasons for a classification, moving beyond simple feature importance to reveal underlying semantic structures. ProToMEx naturally provides both global explanations of a model's overall behaviour and local explanations that can disentangle multiple co-existing reasons for a specific prediction. We demonstrate empirically that ProToMEx not only produces explanations of comparable fidelity to popular methods like SHAP and LIME but also drastically reduces the amortised computational cost of generating local explanations, making it highly suitable for real-time applications. Specifically, we show that ProToMEx is ~30-40x faster than SHAP and LIME over standardised tabular datasets and synthetic datasets.
1 Introduction
Feature-attribution explainers can identify influential features without showing how they combine into a coherent reason. ProToMEx uses latent topics to represent such reasons and provide global and local explanations.
1 Introduction
- Feature rankings can leave the combinatorial logic behind a prediction ambiguous, motivating explanations of why features matter together.The introduction frames this as a gap in methods such as SHAP and LIME.Source: §1 · p. 1
- A single prediction may reflect multiple distinct reasons, which a simple feature ranking can conflate.The introduction illustrates this with alternative combinations of low income and high debt or unstable employment history.Source: §1 · pp. 1–2
- ProToMEx models latent topics as distinct, interpretable combinations of feature-values associated with class labels, serving as global summaries and local explanations.A topic's activation for an instance indicates which reasoning patterns apply to its prediction.Source: §1 · p. 2
- ProToMEx translates instance-label pairs into weighted textual documents, bridging structured tabular data and the bag-of-words format required by PTMs.The authors describe this as the first successful adaptation of PTMs for post-hoc explanation of any black-box classifier.Source: §1 · p. 2
- The proposed explainer provides global class patterns and local patterns active for an instance, with the aim of disentangling multiple reasons for a prediction.The introduction presents topic-based reasoning and the data-translation method as its contributions.Source: §1 · p. 2
2 Related work
Related work includes feature-attribution methods, local rules, and approaches to explaining feature patterns. ProToMEx differs by representing patterns of features associated with class labels as reasons for outcomes.
2 Related work
- LRP, LIME [16], and Chen and Song [5] identify important features or subsets, using backpropagation, local linear weights, or mutual-information-based selection.The section describes these as feature-oriented explanation approaches.Source: §2 · p. 2
- SHAP assigns marginal contributions to input features using Shapley values, but its computation considers feature combinations and can become demanding as feature count grows.The paper contrasts its representation of joint feature knowledge with SHAP's importance weights.Source: §2 · p. 2
- Anchors are high-precision local rules under which a prediction remains consistent when the rules are satisfied.Source: §2 · p. 2
- The FCA-based approach of Radja et al. identifies properties related to outcomes in multi-label classification models.Source: §2 · p. 3
- ProToMEx identifies feature patterns associated with class labels and treats a selection of features that together justify an outcome as a reason.Source: §2 · p. 3
3 Background
The background establishes the classifier and topic-model concepts used by ProToMEx. It describes supervised classification as learning a mapping to labels and PTMs as discovering latent themes in document collections.
3.1 ML Classifiers
- A classifier approximates a rule mapping observations, represented by feature vectors, to one of a set of class labels.The background denotes the observation domain as a product of feature domains.Source: §3.1 · p. 3
3.2 Probabilistic Topic Modelling
- Probabilistic Topic Models are unsupervised generative models that discover latent semantic themes in a corpus of documents.A topic is a multinomial distribution over the vocabulary.Source: §3.2 · p. 3
- LDA models each document as a mixture over latent topics and estimates document-topic mixtures and shared topic-word distributions from an observed corpus.The model draws document topic proportions from a Dirichlet prior before generating words.Source: §3.2 · p. 3
4 ProToMEx: A Probabilistic Topic Model Explainer
ProToMEx trains a topic model on documents constructed from a black-box classifier's predictions, then infers topics for new instances. The resulting topics provide local explanations for individual predictions and global patterns for classes.
4 ProToMEx: A Probabilistic Topic Model Explainer
- ProToMEx learns topics representing coherent co-occurrences of feature-values and class labels, then uses them for global and local explanations of a black-box classifier.The framework is model-agnostic and is designed to disentangle multiple reasons for a classification.Source: §4 · p. 3
4.1 Offline Phase: Corpus Construction and Topic Modelling
- The methodology separates one-time offline corpus construction and topic-model training from online explanation generation for new instances.Source: §4 · p. 3, §4.1 · p. 4
Synthetic Dataset Generation.
- The offline phase samples instances from the domain, obtains labels from the classifier, and uses the resulting synthetic dataset as an empirical representation of its behavior.Source: Synthetic Dataset Generation · p. 4
Document Transformation.
- Each instance-label pair becomes a bag-of-words document containing class-label and discretized feature-value tokens, with characteristic feature-values weighted more strongly.The label token is repeated to assert its importance.Source: Document Transformation · pp. 4–5
4.2 Online Phase: Explanation Generation
- After training, ProToMEx generates explanations efficiently by transforming a new instance and its predicted label into a document and inferring its topic mixture.Source: §4.2 · p. 5, Local Explanations · p. 5
Local Explanations.
- A local explanation consists of topics with non-negligible probability in the instance's topic mixture, each highlighting a distinct feature-value combination that supports the predicted class.The set can reveal multiple independent lines of reasoning for the prediction.Source: Local Explanations · p. 5
Global Explanations.
- A global explanation for a class consists of topics strongly associated with its label, each representing a prototypical feature-value profile learned by the classifier.Source: Global Explanations · p. 5
5 Empirical Evaluation
The evaluation compares ProToMEx with SHAP and LIME on fidelity and computational performance across public and synthetic datasets. Synthetic datasets use class-related feature patterns as ground-truth explanations and vary feature and label characteristics.
5 Empirical Evaluation
- The evaluation benchmarks ProToMEx against SHAP and LIME on explanation fidelity and computational performance across public and synthetic binary and multi-class datasets.Source: §5 · p. 5
Synthetic Datasets.
- Synthetic datasets vary feature counts and class labels, with numerical and categorical features and patterns generated for each label.Source: Synthetic Datasets · p. 6
- Synthetic pattern similarity combines feature-level similarity with a scaling factor that accounts for pattern coverage and alignment.Feature similarity is based on domain overlap for matching features; coverage and alignment affect the permitted pattern similarity.Source: Synthetic Datasets · p. 6
- Generated data points are sampled from patterns associated with their labels, and each generating pattern is treated as a sufficient explanation.Source: Synthetic Datasets · p. 6
- The synthetic-data split assigns 60% to classifier training, 35% to ProToMEx training, and 5% to testing.Source: Table 1 · p. 7
Loan Approval Dataset.
The loan-approval dataset contains 45,000 data points with 13 categorical and continuous features and two approval outcomes.
Loan Approval Dataset.
- The loan-approval dataset has 45,000 data points and 13 features: five categorical and eight continuous, with loan approval or rejection as the two classes.About 22.23% of data points are approved and 77.77% rejected.Source: Loan Approval Dataset · p. 7
Lung Cancer Dataset.
The lung-cancer dataset contains 3,000 data points, 15 features, and two health-status classes.
Lung Cancer Dataset.
- The lung-cancer dataset has 3,000 data points, 14 categorical features and one continuous feature, split between healthy and malignant lungs.Healthy lungs account for 49.4% of data points and malignant lungs 50.6%.Source: Lung Cancer Dataset · p. 7
Wine Quality Dataset.
The wine-quality dataset has 21,000 data points and three quality classes, while the air-quality dataset has 5,000 data points and four air-quality classes.
Wine Quality Dataset.
- The wine-quality dataset contains 21,000 data points, 11 continuous features, and low-, mediocre-, and high-quality classes.The classes account for 28.57%, 42.86%, and 28.57% of data points, respectively.Source: Wine Quality Dataset · p. 7
Air Quality Dataset.
- The air-quality dataset contains 5,000 data points, nine continuous features, and four classes ranging from good to hazardous.The classes are good (40%), moderate (30%), poor (20%), and hazardous (10%).Source: Air Quality Dataset · p. 7
5.2 Setup
The evaluation partitions each dataset into training, corpus, and testing sets, then repeats the explanation procedure across 100 random seeds. Its tables cover synthetic-dataset runtime comparisons and similarity between discovered topics and original patterns.
5.2 Setup
- Synthetic datasets use 60% training, 35% corpus, and 5% testing data; public datasets use 70%, 28%, and 2%, respectively.Source: §5.2 · pp. 7–8
- Table 3 reports average time required for ProToMEx versus SHAP on synthetic datasets.Source: Table 3 · p. 8
- Table 4 compares similarity between ProToMEx topics and original patterns.Source: Table 4 · p. 8
- Each run trains a five-layer ReLU neural-network classifier, uses corpus-set predictions to train ProToMEx’s LDA model, and generates test-instance explanations with ProToMEx, SHAP, and LIME.The procedure is repeated 100 times with different random seeds, and reported results are averages across runs.Source: §5.2 · p. 8
5.3 Results
Across public datasets, ProToMEx generates online explanations faster than SHAP and LIME; synthetic evaluations find its topics align with generated patterns and local topics often match ground-truth reasons.
5.3 Results
- On synthetic datasets, ProToMEx was up to 25× faster than SHAP during online explanations, after a one-time corpus-generation and LDA-training cost.Results were averaged over 100 runs.Source: §5.3 · p. 8, Figure 1 · p. 11
- ProToMEx topics reached at least 0.62 average similarity to the patterns used to generate synthetic datasets, which tended to be subsets of discovered topics.Similarity depends on dataset feature count and the number of patterns used.Source: §5.3 · p. 9
- Most-probable relevant topics had good ground-truth similarity, but the most-similar topics scored higher, indicating probability rankings were not always the closest match.The local evaluation considers the most probable topic, the most similar relevant topic, and probability-weighted similarity across relevant topics.Source: §5.3 · p. 9
- 30–40× faster than SHAP and 20× faster than LIME online on public datasets; ProToMEx became most time-efficient after 2,600 Loan, 1,225 Wine, 125 Air, and 90 Lung Cancer instances.The comparison includes a one-time corpus-generation and LDA-training overhead.Source: §5.3 · p. 9, Table 5 · p. 10
- In most instances across datasets, ProToMEx missed fewer than 3 influential features identified by either SHAP or LIME.Tables 6 and 7 compare average influential-feature counts and the number of instances by how many influential features ProToMEx missed.Source: Table 6 · p. 10, Table 7 · p. 10, §5.3 · p. 10
- 0 ProToMEx instances missed more than 6 features on Wine and Air, versus 9 and 3.37 Poor Sampling SHAP instances, respectively; the authors caution that results are not conclusive.The comparison forced SHAP to miss the top three features through mean-value obfuscation and reduced sampling to n=10.Source: §5.3 · pp. 10–11, Table 8 · p. 12
6 Conclusions
ProToMEx uses probabilistic topic modelling, specifically LDA, to explain classifiers through latent relations between input features and outputs. In a preliminary evaluation, it produced explanations of similar quality to SHAP on many instances and generated single-instance explanations at least 40× faster.
6 Conclusions
- ProToMEx generated single-instance explanations at least 40× faster than SHAP.The evaluation also found it provided explanations of similar quality to SHAP for 75% of instances in three datasets, missing fewer than 2 influential features.Source: §6 · p. 12
- For 75% of instances in three datasets, ProToMEx explanations were of similar quality to SHAP, missing fewer than 2 influential features.Source: §6 · p. 12
- ProToMEx identified more influential features than Poor Sample SHAP, which misses influential features.Source: §6 · p. 12
- The explainer uses LDA to extract latent relations between input features and classifier outputs, after translating input data into documents.The topics can serve as both global and local explanations.Source: §6 · p. 12
- The evaluation was preliminary; the authors identify user studies as a future step to assess human interpretability and utility.Source: §6 · p. 12