T-COL: Generating Counterfactual Explanations for General User Preferences on Variable Machine Learning Systems

TL;DR

This work tackles interpretability with counterfactual explanations under two practical challenges: general user preferences and non-static ML systems. It introduces T-COL, an instance-based method that builds local greedy trees from a query and prototype cases to generate CEs that align with general user preferences through prototype screening and local optimization rules. The authors map preferences to CE properties, simulate users with LLM-based agents, and show that T-COL outperforms baselines in adaptability and robustness across five benchmark datasets, while achieving significant efficiency gains. The results suggest that robust, user-aligned CEs are feasible in dynamic ML deployments, enabling actionable explanations that remain valid as models evolve.

Abstract

To address the interpretability challenge in machine learning (ML) systems, counterfactual explanations (CEs) have emerged as a promising solution. CEs are unique as they provide workable suggestions to users, instead of explaining why a certain outcome was predicted. The application of CEs encounters two main challenges: general user preferences and variable ML systems. On one hand, user preferences for specific values can vary depending on the task and scenario. On the other hand, the ML systems for verification may change while the CEs are performed. Thus, user preferences tend to be general rather than specific, and CEs need to be adaptable to variable ML models while maintaining robustness even as these models change. Facing these challenges, we propose general user preferences based on insights from psychology and behavioral science, and add the challenge of non-static ML systems as one preference. Moreover, we introduce a novel method, \uline{T}ree-based \uline{C}onditions \uline{O}ptional \uline{L}inks (T-COL) for generating CEs adaptable to general user preferences. Moreover, we employ T-COL to enhance the robustness of CEs with specific conditions, making CEs robust even when the ML models are replaced. To assess subjectivity preferences, we define LLM-based autonomous agents to simulate users and align them with real users. Experiments show that T-COL outperforms all baselines in adapting to general user preferences.

Figures (6)

• Figure 1: The whole process of T-COL, with the triangular arrow indicating the flow between the several processes. The user wants a more robust CE, and according to the previous introduction, T-COL first selects the target class sample for which the ML model gives the highest probability as the prototype case. After that, a group of local greedy trees is constructed using the prototype case and the query sample. For each tree, a combination of local feature values is selected using rss and finally stitched into a CE.
• Figure 2: An example of a local greedy tree. The depth of the tree is equal to the number of elements in the local feature subset, and the nodes at each level store the feature values of the query sample and the prototype case, respectively. By traversing each path of the local greedy tree, several sets of local feature combinations can be obtained, and the optimal local feature combination can be selected according to the preset rules.
• Figure 3: An example of a simulated user experiment for preference selection. To ensure the fairness of the experiment, T-COL and the baseline methods were replaced with method A or method B. Each US-Agent is asked to choose their preferred method, A or B, based on their preferences and the properties of CEs. To facilitate statistical analysis, each agent's response is restricted to JSON format.
• Figure 4: The regional distribution of user research. The regions are grouped by continent.
• Figure 5: The industry distribution in user research. Others include industries with few participating users in the research, such as Healthcare or Social Security, Logistics and Transport, Professional Services (e.g., legal or consultancy services), Wholesale and Retail.