Fast top-k retrieval for model based recommendation

A crucial task in many recommender problems like computational advertising, content optimization, and others is to retrieve a small set of items by scoring a large item inventory through some elaborate statistical/machine-learned model. This is challenging since the retrieval has to be fast (few milliseconds) to load the page quickly. Fast retrieval is well studied in the information retrieval (IR) literature, especially in the context of document retrieval for queries. When queries and documents have sparse representation and relevance is measured through cosine similarity (or some variant thereof), one could build highly efficient retrieval algorithms that scale gracefully to increasing item inventory. The key components exploited by such algorithms is sparse query-document representation and the special form of the relevance function. Many machine-learned models used in modern recommender problems do not satisfy these properties and since brute force evaluation is not an option with large item inventory, heuristics that filter out some items are often employed to reduce model computations at runtime.

In this paper, we take a two-stage approach where the first stage retrieves top-K items using our approximate procedures and the second stage selects the desired top-k using brute force model evaluation on the K retrieved items. The main idea of our approach is to reduce the first stage to a standard IR problem, where each item is represented by a sparse feature vector (a.k.a. the vector-space representation) and the query-item relevance score is given by vector dot product. The sparse item representation is learnt to closely approximate the original machine-learned score by using retrospective data. Such a reduction allows leveraging extensive work in IR that resulted in highly efficient retrieval systems. Our approach is model-agnostic, relying only on data generated from the machine-learned model. We obtain significant improvements in the computational cost vs. accuracy tradeoff compared to several baselines in our empirical evaluation on both synthetic models and on a click-through (CTR) model used in online advertising.