Data and Code for: "Auditing Construct Validity and Explanation Faithfulness of an Attribute-Based Product Recommender Across Four E-Commerce Data Sources"

Data and code for: "Auditing construct validity and explanation faithfulness of an attribute-based product recommender across four e-commerce data sources" (Herrera-García, Caro Alvarez, Ramírez Mackenzie, Zuñiga Romero). The study audits a deployed attribute-based recommendation index on two independent fronts: (1) construct validity against real purchase behavior across four e-commerce data sources (Amazon, MercadoLibre, Falabella, Wish), and (2) faithfulness of LLM-generated justifications against the SHAP attribution of the underlying model, under baseline, attribution-guided, and global-ablation prompting conditions. Contents:- Cleaned Amazon Reviews 2023 working dataset and evaluation frame with verified-purchase counts- MercadoLibre and Falabella scraped ranking datasets (Colombian storefronts) Raw Amazon Reviews 2023 data (Hou et al., 2024) is not redistributed here due to source licensing; only the cleaned, derived datasets produced by the authors' processing pipeline are included. The Wish dataset (Kaggle, 2020) is not redistributed; the authors used the original source directly (see README for citation). This deposit is currently restricted while the associated manuscript is under peer review. Code (model training, SHAP pipeline, LLM audit prompts) will be added in a future version ahead of publication.