A cuisine-adaptive explainable AI framework for restaurant-specific meal recommendation: a Thai restaurant case study

Food recommender systems usually target individual foods or recipes, yet restaurant dining often calls for recommending a complete dish in relation to others already chosen within the same visit. This study develops a transparent-by-design, restaurant-specific and cuisine-adaptable framework for whole-dish recommendation and demonstrates it using transaction data from a Thai restaurant. The framework combines two complementary components: a behavioral layer that learns local co-ordering patterns from the restaurant’s own transactions, and a semantic layer in which culinary and sensory descriptors are defined for the cuisine being served. In the Thai application, three Thai food experts independently coded dishes by preparation-based culinary form and by selected sensory characteristics, distinguishing five gustatory categories from chemesthetic pungency/spiciness and representing texture and dominant aroma-source as reduced-dimensional descriptors. The Apriori algorithm identified recurrent co-ordering relationships using support, confidence, and lift. Under the final specification (minimum support = 0.10, minimum confidence = 0.20, lift > 1), 78 directional rules representing 39 unordered category pairs were retained. The highest-support pair was Spicy Salad–Fried Dish (support = 53.60%); directionally, Spicy Salad → Fried Dish had confidence = 77.54% and lift = 1.05, while Fried Dish → Spicy Salad had confidence = 72.91%. These explicit rules are combined with human-interpretable culinary and sensory attributes to explain why the system recommends one dish given another. The framework does not encode universal Thai pairing rules; rather, it learns restaurant-specific behavioral patterns while using cuisine-appropriate domain knowledge to make those patterns interpretable. The architecture is therefore designed to be adaptable to other restaurants and cuisines by retaining the transaction-learning mechanism while redefining the semantic coding scheme with relevant domain experts. Cross-restaurant and cross-cuisine transferability remains to be empirically validated.