Beyond Black Boxes: Implementing Recommender Systems That Generate Trust for Decision-Making in Sensitive Business Areas

There are sensitive business areas, such as the health and finance industries, where errors can cause severe problems. In such sectors, the accuracy and interpretability of the AI models that support decisions are pretty important. Nowadays, most AIbased systems act like black boxes, generating responses and supporting decision-making without providing clear explanations to end users. In this article, we propose an architecture for Recommender Systems that intends to generate trust among users of such systems. The main principle is to combine data from structured sources and data collected from experts as the basis for recommendations. A differentiator in our proposal lies in using expert systems to identify the most precise sequence of rules to reach a response. This would allow us to obtain the best routes and facts to be used in the fine-tuning process of LLM models using QLoRA techniques to reduce the computational cost of periodic retraining. Once the LLM model is calibrated, new counterfactual techniques (calculating and visualizing counterfactual feature importance values) are used to understand the impact of changes in the prompts entered for adjusting the LLM model. The results of this process provide feedback to the expert engine for future refinements of the LLM model. By combining all three expert systems, human expert knowledge, and counterfactual analysis techniques, greater accuracy would be achieved in the responses generated by the recommendation system, delivering interpretable explanations with a robust validation process.