PAVE: Lazy-MDP based ensemble to improve recall of product attribute extraction models

E-commerce stores face the challenge of missing and inconsistent attribute values in the product detail pages and have to impute them on behalf of their vendors. Traditional approaches formulate the problem of attribute extraction (AE) from product profiles as natural language tasks such as information extraction or text classification. Such models typically operate at high precision but may yield low recall especially on attributes with an open vocabulary due to 1) missing or incorrect information in product profiles, 2) generalization errors due to lack of contextual understanding, and 3) confidence thresholding to operate at high precision. In this work, we present PAVE: Product Attribute Value Ensemble, a novel reinforcement learning model that uses Lazy-MDP formalism to solve for low recall by aggregating information from a sequence of product neighbors. We train a policy network using Proximal Policy Optimization that learns to choose the correct value from the sequence. We observe consistent improvement in recall across all open attributes compared to traditional AE models with an average lift of 10.3% with no drop in precision. Our method surpasses simple aggregation methods like nearest neighbor, majority vote and binary classifier ensembles and even outperforms AE models for closed attributes. Our approach is scalable, robust to noisy product neighbors and generalizes well on unseen attributes.