Counterfactual-Empowered Causal Analysis and Trustworthiness Verification for Recommendation Results

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Recommender systems are widely used in e-commerce to alleviate information overload, yet the nonlinear decision mechanisms of neural recommenders make it difficult to verify whether an identified feature is genuinely responsible for a recommendation. To address this problem, this study proposes a counterfactual-empowered causal analysis and trustworthiness verification framework that operates on interpretable semantic attributes extracted from review text. DeepSeek is used as a constrained semantic annotator to extract aspect–opinion–sentiment triplets from Amazon Cell Phones and Accessories reviews, thereby constructing an explicit semantic intervention space. A symmetric multilayer perceptron (MLP) is trained as the recommendation predictor and is kept fixed during verification; the verifier assumes gradient access with respect to the semantic input representation but does not update the predictor parameters. Three mechanisms are introduced to improve counterfactual verification: (1) AdamW optimization with cosine-annealing scheduling for stable continuous perturbation search; (2) a gradient-guided greedy strategy that maps continuous perturbations to discrete human-understandable semantic attributes; and (3) Elastic Net regularization that jointly constrains explanation sparsity and perturbation magnitude. Experiments on the processed Amazon Cell Phones and Accessories dataset show that the proposed components improve Precision, Probability of Necessity (PN), Probability of Sufficiency (PS), and Feature Necessity Score (FNS) relative to the evaluated optimization and ablation variants. These results indicate that semantic counterfactual intervention provides a practical post-hoc mechanism for assessing whether recommendation rationales are causally effective rather than merely correlated with model outputs.