Towards Interpreting Multi-Objective Feature Associations

TL;DR

The paper addresses interpretability in multi-objective feature associations for agricultural outcomes by proposing an objective-specific feature interaction design that couples additive explanations with global sensitivity analysis to prune the combinatorial search. It develops a five-step explainable AI for combinatorial optimization framework and an EBCO method that uses a learned model, pruning, and sensitivity-based scoring to identify optimal feature-value combinations for multi-label predictions. The approach is validated on pre-harvest MDR and post-harvest poultry pathogen datasets, showing explanation-based selection can reduce pathogen presence with fewer iterations than dynamic programming. This work contributes a practical, scalable method to derive actionable farm-management patterns that minimize pathogen risk across multiple objectives.

Abstract

Understanding how multiple features are associated and contribute to a specific objective is as important as understanding how each feature contributes to a particular outcome. Interpretability of a single feature in a prediction may be handled in multiple ways; however, in a multi-objective prediction, it is difficult to obtain interpretability of a combination of feature values. To address this issue, we propose an objective specific feature interaction design using multi-labels to find the optimal combination of features in agricultural settings. One of the novel aspects of this design is the identification of a method that integrates feature explanations with global sensitivity analysis in order to ensure combinatorial optimization in multi-objective settings. We have demonstrated in our preliminary experiments that an approximate combination of feature values can be found to achieve the desired outcome using two agricultural datasets: one with pre-harvest poultry farm practices for multi-drug resistance presence, and one with post-harvest poultry farm practices for food-borne pathogens. In our combinatorial optimization approach, all three pathogens are taken into consideration simultaneously to account for the interaction between conditions that favor different types of pathogen growth. These results indicate that explanation-based approaches are capable of identifying combinations of features that reduce pathogen presence in fewer iterations than a baseline.

Figures (2)

• Figure 1: Selection of the best feature value combinations based on the importance of the features and objective criteria. To reduce the combinatorial optimization search space, we use popular explanation methods to find local explanations of each feature (root node) and value (child node). A threshold on feature relevance is used to prune combinations of features and values from the search space. The final combination selection is based upon the objective of the problem and the relevance of the feature combinations to the model prediction.
• Figure 2: Design diagram of multi-label combinatorial optimization. To predict food-borne pathogens in the first stage of the process, we build a multi-label multi-layer perceptron architecture based on our agricultural practices. The next step involves learning the relevance of a feature, value> relevance in the prediction using DeepShap lundberg2017unified which is combined with additive explanations (SHAP) and linear functional approximations (DeepLIFT). By using threshold-based pruning, the search space is optimized and computing time is reduced. In the final step of our study, our objective based selection is based on learning the global explanation of feature combinations.