Initializing Services in Interactive ML Systems for Diverse Users

2023-12-19

Neural Information Processing Systems

AI Summary

This work addresses the multi-service initialization problem for heterogeneous users in interactive machine learning: under initial preference data scarcity and a non-convex aggregate loss function, user-adaptive service selection induces learning dynamics highly sensitive to initialization. To tackle this, we propose the first theoretically grounded randomized initialization algorithm, extending the k-means++ paradigm to non-convex multi-service learning with bandit feedback. Our method jointly performs stochastic service sampling and implicit preference inference to assign services, guaranteeing that the expected aggregate loss is within an $O(log K)$ factor of the global optimum. Empirical evaluation on real-world and semi-synthetic datasets demonstrates that our approach significantly outperforms uniform and naive random initialization baselines in convergence speed and final performance.

Abstract

This paper investigates ML systems serving a group of users, with multiple models/services, each aimed at specializing to a sub-group of users. We consider settings where upon deploying a set of services, users choose the one minimizing their personal losses and the learner iteratively learns by interacting with diverse users. Prior research shows that the outcomes of learning dynamics, which comprise both the services' adjustments and users' service selections, hinge significantly on the initialization. However, finding good initializations faces two main challenges: (i) Bandit feedback: Typically, data on user preferences are not available before deploying services and observing user behavior; (ii) Suboptimal local solutions: The total loss landscape (i.e., the sum of loss functions across all users and services) is not convex and gradient-based algorithms can get stuck in poor local minima. We address these challenges with a randomized algorithm to adaptively select a minimal set of users for data collection in order to initialize a set of services. Under mild assumptions on the loss functions, we prove that our initialization leads to a total loss within a factor of the globally optimal total loss with complete user preference data}, and this factor scales logarithmically in the number of services. This result is a generalization of the well-known $k$-means++ guarantee to a broad problem class, which is also of independent interest. The theory is complemented by experiments on real as well as semi-synthetic datasets.

Problem

• Initialize ML services for diverse users
• Overcome bandit feedback challenges
• Avoid suboptimal local solutions

Innovation

• Randomized algorithm for user selection
• Adaptive data collection initialization
• Logarithmic scaling in services