A crucial task in many recommender problems like computationaladvertising, content optimization, and others is to retrieve a small setof items by scoring a large item inventory through some elaboratestatistical/machine-learned model. This is challenging since theretrieval 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 andrelevance is measured through cosine similarity (or some variantthereof), one could build highly efficient retrieval algorithms thatscale gracefully to increasing item inventory. The key componentsexploited by such algorithms is sparse query-documentrepresentation and the special form of the relevance function. Manymachine-learned models used in modern recommender problems donot satisfy these properties and since brute force evaluation is not anoption with large item inventory, heuristics that filter out some itemsare often employed to reduce model computations at runtime.
There are a two-stage approach where the first stage retrieves top-Kitems using our approximate procedures and the second stage selectsthe desired top-k using brute force model evaluation on the K retrieveditems. The main idea of our approach is to reduce the first stage to astandard IR problem, where each item is represented by a sparsefeature vector (a.k.a. the vector-space representation) and the query-item relevance score is given by vector dot product. The sparse itemrepresentation is learn to closely approximate the original machine-learned score by using retrospective data. Such a reduction allowsleveraging extensive work in IR that resulted in highly efficient retrievalsystems. Our approach is model-agnostic, relying only on datagenerated from the machine-learned model. We obtain significantimprovements in the computational cost vs. accuracy tradeoffcompared to several baselines in our empirical evaluation on bothsynthetic models and on a (CTR) model used in online advertising.
The Backend ViewNavigator cache reduces the number of servertransactions and associated network overhead when navigatingand reading Column Values information from the Documentsand Entries in a View. Performance gains are most profoundwhen accessing a View residing on a server from aclient, however retrieval from local Views will also be greatlyimproved.I hope this ppt will helpful for you but suggestions are stillwelcome from reader’s side.