Expertise drift in referral networks

摘要

Learning-to-refer is a challenge in expert referral networks, wherein Active Learning helps experts (agents) estimate the topic-conditioned skills of other connected experts for problems that the initial expert cannot solve and therefore must seek referral to experts with more appropriate expertise. Recent research has investigated different reinforcement action selection algorithms to assess viability of the learning setting both with uninformative priors and with partially available noisy priors, where experts are allowed to advertise a subset of their skills to their colleagues. Prior to this work, time-varying expertise drift (e.g., experts learning with experience) had not been considered, though it is an aspect that may often arise in practice. This paper addresses the challenge of referral learning with time-varying expertise, proposing Hybrid, a novel combination of Thompson Sampling and Distributed Interval Estimation Learning (DIEL) with variance reset, first proposed in this paper. In our extensive empirical evaluation, considering both biased and unbiased drift, the proposed algorithm outperforms the previous state-of-the-art (DIEL) and other competitive algorithms e.g., Thompson Sampling and Optimistic Thompson Sampling. We further show that our method is robust to topic-dependent drifts and expertise level-dependent drifts, and the newly-proposed DIEL\(_{reset}\) can be effectively combined with other Bayesian approaches e.g., Optimistic Thompson Sampling and Dynamic Thompson Sampling and Discounted Thompson Sampling for improved performance.