The online advertising market, with its thousands of auctions run per second, presents a daunting challenge for advertisers who wish to optimize their spend under a budget constraint. Thus, advertising platforms typically provide automated agents to their customers, which act on their behalf to bid for impression opportunities in real time at scale. Because these proxy agents are owned by the platform but use advertiser funds to operate, there is a strong practical need to balance reliability and explainability of the agent with optimizing power. We propose a generalizable approach to optimizing bidding policies in production environments by learning from real data using offline reinforcement learning. This approach can be used to optimize any differentiable base policy (practically, a heuristic policy based on principles which the advertiser can easily understand), and only requires data generated by the base policy itself. We use a hybrid agent architecture that combines arbitrary base policies with deep neural networks, where only the optimized base policy parameters are eventually deployed, and the neural network part is discarded after training. We demonstrate that such an architecture achieves statistically significant performance gains in both simulated and at-scale production bidding environments compared to the default production bidding policy. Our approach does not incur additional infrastructure, safety, or explainability costs, as it directly optimizes the parameters of existing production routines without necessarily replacing them with black box-style models like neural networks.

Relevance feedback has been demonstrated to be an effective strategy for improving retrieval accuracy. The existing relevance feedback algorithms based on language models and vector space models are not effective in learning from negative feedback documents, which are abundant if the initial query is difficult. The probabilistic retrieval model has the advantage of being able to naturally improve the estimation of both the relevant and non-relevant models. The Dirichlet compound multinomial (DCM) distribution, which relies on hierarchical Bayesian modeling techniques, is a more appropriate generative model for the probabilistic retrieval model than the traditional multinomial distribution. We propose a new relevance feedback algorithm, based on a mixture model of the DCM distribution, to effectively model the overlaps between the positive and negative feedback documents. Consequently, the new algorithm improves the retrieval performance substantially for difficult queries. To further reduce human relevance evaluation, we propose a new active learning algorithm in conjunction with the new relevance feedback model. The new active learning algorithm implicitly models the diversity, density and relevance of unlabeled data in a transductive experimental design framework. Experimental results on several TREC datasets show that both the relevance feedback and active learning algorithm significantly improve retrieval accuracy.

Relevance feedback, which traditionally uses the terms in the relevant documents to enrich the user's initial query, is an effective method for improving retrieval performance. The traditional relevance feedback algorithms lead to overfitting because of the limited amount of training data and large term space. This paper introduces an online Bayesian logistic regression algorithm to incorporate relevance feedback information. The new approach addresses the overfitting problem by projecting the original feature space onto a more compact set which retains the necessary information. The new set of features consist of the original retrieval score, the distance to the relevant documents and the distance to non-relevant documents. To reduce the human evaluation effort in ascertaining relevance, we introduce a new active learning algorithm based on variance reduction to actively select documents for user evaluation. The new active learning algorithm aims to select feedback documents to reduce the model variance. The variance reduction approach leads to capturing relevance, diversity and uncertainty of the unlabeled documents in a principled manner. These are the critical factors of active learning indicated in previous literature. Experiments with several TREC datasets demonstrate the effectiveness of the proposed approach.

The classical probabilistic models attempt to capture the Ad hoc information retrieval problem within a rigorous probabilistic framework. It has long been recognized that the primary obstacle to effective performance of the probabilistic models is the need to estimate a relevance model. The Dirichlet compound multinomial (DCM) distribution, which relies on hierarchical Bayesian modeling techniques, or the Polya Urn scheme, is a more appropriate generative model than the traditional multinomial distribution for text documents. We explore a new probabilistic model based on the DCM distribution, which enables efficient retrieval and accurate ranking. Because the DCM distribution captures the dependency of repetitive word occurrences, the new probabilistic model is able to model the concavity of the score function more effectively. To avoid the empirical tuning of retrieval parameters, we design several parameter estimation algorithms to automatically set model parameters. Additionally, we propose a pseudo-relevance feedback algorithm based on the latent mixture modeling of the Dirichlet compound multinomial distribution to further improve retrieval accuracy. Finally, our experiments show that both the baseline probabilistic retrieval algorithm based on the DCM distribution and the corresponding pseudo-relevance feedback algorithm outperform the existing language modeling systems on several TREC retrieval tasks.