articles

The speed, precision and recall of information search in e-commerce are critical indicators for business success. A large number of academic studies are aimed at increasing these indicators through more efficient utilization of hardware and the development of machine learning models with new “layers”, training data and loss functions. In this study, the author focused on the practical task of speeding up the search by using knowledge about the nature of the load and data. Widely used Approximate Nearest Neighbors methods use artificial clusters to reduce the processing time of a search query. At the same time, the recall of the list of candidates found worsens. This approach is justified by its universality in relation to data. But in the case of an electronic online trading platform, the data are products and their modalities – name, description, product class, images, which makes it possible to use this knowledge about data to create more effective search structures and algorithms. In conditions of high dynamics of changes in product data, it is also necessary to take into account the speed, accuracy and completeness for offline and online processes. Therefore, the author considered the task of forming the completeness and accuracy of search results within the framework of an end-to-end process, and not only as an retrieval phase. As a result, the author received an improvement in the recall and precision of the retrieved product information by more than 50% without reducing the speed of search query processing.