A Novel Approach to Balance Convenience and Nutrition in Meals With Long-Term Group Recommendations and Reasoning on Multimodal Recipes and its Implementation in BEACON

Vansh Nagpal, Siva Likitha Valluru, Kausik Lakkaraju, Nitin Gupta, Zach Abdulrahman, Andrew Davison, Biplav SrivastavaarXiv 2024 · DOI · arXiv

Introduction

Despite the widespread knowledge that nutrition is vital for long-term health, nearly 40% of the global population fails to adhere to national dietary guidelines. This failure is largely attributed to the inherent trade-off between nutrition (health goals) and convenience (cost, speed, and taste). Existing digital tools often exacerbate this problem: manual trackers are labor-intensive, while Large Language Models (LLMs) frequently hallucinate nutritional data or suggest unbalanced diets for users with specific health conditions like diabetes.

BEACON (Balanced Eating and Convenience Optimization Network) is designed to bridge this gap. It acts as a "trusted companion" that reasons over structured recipe data to provide realistic, multi-day meal plans that respect both health constraints and the user's need for variety and ease.

System Architecture and Methodology

BEACON's effectiveness stems from a three-pronged technical approach:

  1. Multimodal Data Representation (R3): The system moves beyond unstructured text by utilizing the Rich Recipe Representation (R3) format. This JSON schema captures ingredients, nutrients, atomic cooking steps, and the functional "role" of a dish (e.g., Main, Side, or Dessert).
  2. LLM-Assisted Data Pipeline: To scale the recipe database, the researchers developed pipelines to convert plain-text recipes into the R3 format. While LLMs like Mixtral-8x7B were effective at preserving the semantic "essence" of recipes, the study found they required strict guardrails (such as specific temperature settings and few-shot prompting) to consistently produce valid JSON structures.
  3. Relational Boosted Bandit Engine: The core of BEACON is a contextual bandit algorithm (referred to as M2). Unlike standard recommenders that treat food items in isolation, this engine uses predicate logic to evaluate complex "meal configurations" (e.g., a main course paired with a side and a drink) against a user's health profile and taste preferences.
The BEACON system architecture, illustrating the flow from the R3 recipe database and user inputs to the recommendation engine and final evaluation.

Experimental Setup and Datasets

The system was evaluated using a curated dataset of 52 recipes in the R3 format. This included 25 recipes from the RecipeQA dataset and 27 new recipes representing real-world "convenience" choices, such as items from major fast-food chains and culturally significant Soul Food dishes. Performance was measured across three distinct user configurations (varying in dietary strictness) and three time horizons (1, 3, and 5 days).

Key Results and Insights

The researchers evaluated the system using three primary metrics: User Constraint (uc), Meal Item Duplicate (dm), and Meal Coverage (mc).

  • Superior Constraint Satisfaction: The M2 (Relational Boosted Bandit) algorithm consistently outperformed random (M0) and sequential (M1) baselines. In the most constrained user scenarios, M2 achieved a User Constraint score of 0.875 to 0.992, demonstrating its ability to navigate complex dietary requirements.
  • The Power of the "Nudge": Rather than strictly forbidding "unhealthy" foods, BEACON uses visual progress bars to show users how a specific choice (like a fast-food meal) impacts their daily nutritional targets, such as carbohydrate limits for diabetics. This "nudge" approach respects user autonomy while promoting healthier long-term patterns.
Performance metrics across different user configurations (c1, c2, c3) and time frames, showing the M2 algorithm's superior ability to satisfy user constraints compared to baselines.

Conclusion and Future Work

BEACON demonstrates that by combining structured multimodal data with relational learning, recommendation systems can provide meal plans that are both nutritionally sound and practically convenient. While the current study utilized a limited dataset of 52 recipes, future work will focus on scaling the R3 database to thousands of items, expanding the array of tracked allergens and micronutrients, and conducting long-term human-in-the-loop studies to evaluate real-world user adherence.

Summary updated

Abstract

A common decision made by people, whether healthy or with health conditions, is choosing meals like breakfast, lunch, and dinner, comprising combinations of foods for appetizer, main course, side dishes, desserts, and beverages. Often, this decision involves tradeoffs between nutritious choices (e.g., salt and sugar levels, nutrition content) and convenience (e.g., cost and accessibility, cuisine type, food source type). We present a data-driven solution for meal recommendations that considers customizable meal configurations and time horizons. This solution balances user preferences while accounting for food constituents and cooking processes. Our contributions include introducing goodness measures, a recipe conversion method from text to the recently introduced multimodal rich recipe representation (R3) format, learning methods using contextual bandits that show promising preliminary results, and the prototype, usage-inspired, BEACON system.
BibTeX
@misc{nagpal2024novel,
  author = {Vansh Nagpal and Siva Likitha Valluru and Kausik Lakkaraju and Nitin Gupta and Zach Abdulrahman and Andrew Davison and Biplav Srivastava},
  title = {{A Novel Approach to Balance Convenience and Nutrition in Meals With Long-Term Group Recommendations and Reasoning on Multimodal Recipes and its Implementation in BEACON}},
  year = {2024},
  eprint = {2412.17910},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2412.17910},
}
Figure