Recommender systems usually observe only positive user‑item interactions, leaving explicit negative feedback unavailable. Deep models need negative signals, and existing sampling methods simply treat unobserved pairs as negatives, which does not explain why a user is uninterested nor guarantees business relevance. We therefore define implicit negative candidate discovery: mining interaction patterns from observed behavior that can justify labeling an unobserved pair as negative, and encoding these patterns as symbolic rules. Each rule is scored by support, informativeness, and product relevance, then ranked by evidential strength. An LLM interprets the retained rules using business objectives and domain knowledge, and the interpretations are merged with statistical evidence in a final report. Experiments on an industrial B2B platform and five public recommendation datasets show that our method yields higher precision than baselines; in the industrial task, symbolic selection improves downstream test PR‑AUC by 12.5% compared with random selection of four negatives per positive. The results demonstrate that negative candidate validity can be evaluated separately from downstream recommendation performance, enabling evidence‑based, business‑aligned, and explainable negative selection, which enhances interpretability and model training in sparse, skewed real‑world settings.
Review