In question answering (QA) for multi-constraint recommendation, the model must find every item that satisfies a request with several requirements. Still, training data that pair natural-language questions with complete answer sets are scarce. Such data can be synthesized by computing answers from an item catalog and using users’ ratings and reviews to select requests, yet the result may reveal whose evidence was used. We develop a QA data synthesis framework that protects these users using user-level differential privacy (DP), bounding each user’s total contribution and adding calibrated noise to the support statistics used to weight requests. Using one construction template, we build QA datasets for Movie, Beauty, and Recipe from pre-provided, rule-cleaned, and model-inferred attributes. In ablations without noise, an adversary who knows all other users’ data identifies, without false positives, every tested member whose evidence supports a request, up to 96% of members. Under DP with privacy budget ε=4, it flags at most 3.0% of members of these three datasets at a 1% false-positive rate, and an audit with repeated mechanism runs finds no evidence of privacy loss above ε. Models trained on private data retain at least 97% of the non-private ablation’s macro-averaged F1 on shared test questions.

Question-answering recipe recommendation tasks lack datasets with ground-truth labels. Existing methods often generate datasets from simple combinations of ingredient or tag constraints, leading to unnatural expressions and unreasonable constraints that are hard to reflect users’ real preferences. This paper proposes a construction method for a benchmark-style QA dataset for natural-language recipe recommendation. Driven by real user behavior, our method integrates external user query and review data to extract fine-grained preference signals by LLMs and maps them onto a knowledge-graphbased tag schema, thereby constructing QA data consisting of preference and ingredient constraints. Besides, a greedy strategy for constraint selection is also used to control the answer size, while tagbased re-ranking is applied to improve data quality. The generated dataset RecipeADQA contains natural-language questions, structured constraints, and answer sets, resulting in 128,465 instances. Experiments compare RecipeADQA with baseline datasets from two perspectives: data quality and retrieval-recommendation performance. The results show that ours outperforms existing methods on the same recipe knowledge graph in coverage, higher question quality in naturalness and constraint rationality, and better queryability. Moreover, we design a noise-augmented retrieve-and-rerank model, in which a lightweight parser first retrieves candidate recipes, and a small language model then selects and ranks relevant recipes from a noisy candidate pool. The model trained on RecipeADQA outperforms the same model trained on the baseline datasets. The relevant construction code and data have been open-sourced at a GitHub repository https://github.com/MorimoLab/RecipeADQA. Note: The paper has been accepted by "The 22nd International Conference on Advanced Data Mining and Applications 2026 (ADMA 2026)".