TLDR: This research introduces the Contextual Budget Bandit (CBB) model and algorithms (COcc and Mitosis) to improve volunteer engagement in food rescue platforms while ensuring geographical fairness. CBB extends restless multi-armed bandits by allowing context-dependent budget allocation, enabling platforms to prioritize underserved communities. The Mitosis algorithm guarantees optimal budget allocation and significantly outperforms existing methods, demonstrating its effectiveness on both synthetic and real-world data, with potential applications in digital agriculture, peer review, and email campaigns.
Food waste is a global crisis, with billions of tons discarded annually while millions struggle with food insecurity. Volunteer-based food rescue platforms (FRPs) play a crucial role in bridging this gap by connecting surplus food from businesses to communities in need. However, these platforms face a complex challenge: how to keep volunteers engaged and maximize food rescued without inadvertently creating geographical disparities, where some communities are consistently overlooked.
Existing algorithms designed to boost volunteer engagement often exacerbate these imbalances. For instance, a central, popular downtown area might see a 90% rescue completion rate, while outer suburbs could drop to 40%. This highlights a critical need for smarter, fairer allocation strategies.
A new research paper, Contextual Budget Bandit for Food Rescue Volunteer Engagement, introduces an innovative solution: the Contextual Budget Bandit (CBB) model. Developed by Ariana Tang, Naveen Raman, Fei Fang, and Zheyuan Ryan Shi, this model extends the concept of restless multi-armed bandits (RMABs) – a common framework for online resource allocation – by incorporating context-dependent budget allocation. In simpler terms, it allows FRPs to adjust how many volunteers they notify in different regions based on specific geographical information and needs.
The core idea behind CBB is to allocate higher notification budgets to communities with lower match rates, directly addressing and alleviating geographical disparities. This approach offers flexibility in volunteer notifications without sacrificing overall food rescue performance.
How the Contextual Budget Bandit Works
Imagine a system that understands that a volunteer in a busy city center might behave differently than one in a quieter suburb. CBB takes this ‘context’ (like geographical location) into account. Instead of a fixed budget for volunteer notifications across all areas, CBB allows the budget to vary. For example, during a pest season, digital agriculture chatbots might assign a higher budget for pest control tips, as farmers are more likely to engage. Similarly, in food rescue, regions with fewer volunteers or lower engagement rates can be given a higher budget to ensure they receive adequate support.
The researchers developed two key algorithms to implement CBB:
-
COcc (Contextual Occupancy Index Policy): This is an empirically fast heuristic algorithm that provides a good approximation for budget allocation. It assigns budgets based on the average usage predicted by a relaxed optimization problem.
-
Mitosis Algorithm: Recognizing that COcc might not always be optimal, especially when active volunteers are scarce, the Mitosis algorithm was designed. It guarantees the computation of the optimal budget allocation. It uses a clever approach inspired by cell division, progressively exploring and refining budget allocations to find the best solution efficiently, even in complex scenarios.
The paper demonstrates that these algorithms significantly outperform traditional baselines, such as random selection, greedy approaches, and the standard Vanilla Whittle policy, on both synthetic and real-world food rescue datasets. Mitosis consistently achieves the highest rewards while being much more computationally efficient than a brute-force optimal search method called Branch And Bound.
Ensuring Fairness in Food Rescue
Beyond maximizing rescued food, the research also emphasizes fairness. The authors define a fairness index to ensure that an algorithm achieves rewards for each context (region) in proportion to its frequency. This means that underserved areas receive equitable attention. The COcc, Branch And Bound, and Mitosis algorithms can all be modified to incorporate these fairness constraints, allowing platforms to balance overall efficiency with equitable distribution.
Experiments show that fairness-aware Mitosis policies consistently achieve high rewards across varying fairness levels. Visualizations of volunteer allocations in Pittsburgh-area regions clearly illustrate this: without fairness constraints, the central downtown region dominates volunteer engagement. However, with fairness requirements, volunteer support shifts significantly towards suburban and under-served regions, like the northwestern area, correcting structural imbalances.
Also Read:
- Improving Online Planning with Robust Sparse Sampling
- Understanding Digital Ad Performance: Multimodal Forecasting with Explanations
Broader Applications
While grounded in food rescue, the CBB model and its algorithms have wide-ranging applicability. They can be used in:
-
Digital Agriculture: Optimizing nudges from chatbots to smallholder farmers based on urgent needs (e.g., pest control vs. watering tips).
-
Peer Review: Allocating reviewer invitations for different types of academic submissions (e.g., long papers vs. trendy topics).
-
Email Campaigns: Optimizing user engagement by dynamically adjusting email budgets based on campaign features and user attributes.
This research offers a powerful framework for optimizing resource allocation in dynamic, context-rich environments, promising more efficient and equitable outcomes across various domains.
