Abstract
As robotic servers begin to replace human waitstaff, there is a clear need for intelligent systems to help consumers navigate complex menus. This paper presents a Multi-Attribute Relation Matrix Tri-Factorization (MARMTF) technique for a relation-driven food recommendation system. By integrating user ordering history with dish attributes like ingredients, spiciness, and price, the system predicts individual dietary preferences. Experimental results demonstrate that MARMTF effectively simplifies food choice tasks and outperforms traditional collaborative filtering methods, offering a robust solution for future automated restaurant services.
Introduction
The "paradox of choice" is particularly acute in restaurants with extensive menus. While human waiters often use intuition to guide customers, the trend toward robotic servers creates a "recommendation gap." Existing systems often focus narrowly on healthy dieting or simple collaborative filtering (e.g., "people who liked this also liked that"). However, eating behavior is driven by a complex interplay of taste, budget, and ingredient preferences. The challenge lies in mathematically fusing these heterogeneous data sources into a cohesive model that accurately predicts customer enjoyment.
Methodology: Multi-Attribute Relation Matrix Tri-Factorization
The core innovation is the Multi-Attribute Relation Matrix Tri-Factorization (MARMTF) technique. Unlike standard models that rely solely on user ratings, MARMTF constructs a large "block matrix" integrating multiple relationship types.
Relation Fusion
The system models five key entities: Users, Dishes, Price, Ingredients, and Spice Level. As illustrated in the relation fusion graph, these entities are interconnected through various matrices:
- User-Dish: Historical ratings (1–5 scale).
- Dish-Ingredient: Categorical data (meat, poultry, vegetables, etc.).
- Dish-Spiciness: Four levels from "not spicy" to "very spicy."
- Dish-Price: Three tiers (low, medium, high).
Optimization and Prediction
MARMTF uses tri-factorization to decompose the block matrix into three factors ($V$ and $B$), capturing hidden associations across diverse data types. This allows for simultaneous co-clustering of users and dishes based on multiple attributes. Multiplicative update rules are employed to minimize the difference between original sparse data and the reconstructed matrix, effectively filling in predicted ratings for untried dishes.
Experimental Results
The study utilized a dataset of 289 dishes from Chongqing, China, with 37 participants. The MARMTF model consistently outperformed industry-standard algorithms in terms of accuracy.
As shown in the convergence plot, the model achieves stable performance quickly. MARMTF reached a Root Mean Squared Error (RMSE) of 0.1693 and a Mean Absolute Error (MAE) of 0.1342. In comparison, popular models like SVD++ and Classical Matrix Factorization yielded significantly higher error rates (RMSE of 0.1805 and 0.2443, respectively).
Conclusion and Future Work
MARMTF demonstrates that incorporating background information like ingredients and price significantly stabilizes recommendations, especially when user rating data is sparse. While the current study focused on Sichuan cuisine, future iterations could incorporate seasonal contexts and health data (e.g., BMI) to provide even more personalized and nutritionally balanced recommendations. This system has great potential for deployment in future restaurants using non-human waitstaff.
Summary updated