Internet resources: “Yandex” search becomes more accurate

“Yandex” has summed up first results of “MatrixNet” technology implementation, which provided training of the search engine.  Due to the new technology the search engine managed to win its former positions in the market - 62,9% (the last time “Yandex” had such a share was in 2006) – and improve the quality of  delivery on 5%.

Technologies of machine learning appeared already in ‘50s. Currently this approach is being widely implemented for images and voice recognition.

Today there are two most famous and largest schools of machine learning in the world: Support vector machines and Boosting. The first one registers and processes data on hundreds of factors, which characterize the object being studied by the computer. SVM algorithms work quite quick. The second approach is based on the analysis of objects considering thousands of factors, however calculations take more time. Therefore all the search engines until quite recently were making use of SVM learning methods.

"Yandex" was the first to turn to Boosting. It took a year and a half to work out a new system. In the end the company’s team succeeded to create "Matrixnet" technology, which can efficiently work with thousands of factors. The point is that, earlier the key data for the machine were evaluation of correspondence between the inquiry and the page, made by people. Due to the small number of evaluations the search engine began to retrieve and find not existing regularities, relying on evaluations of a few experts. The new system doesn’t rely on evaluations of people therefore it doesn’t commit errors.

According to Arkadiy Volozh, Director General of “Yandex”, due to "MatrixNet", the search engine has now become better in meeting rare and complicated inquiries.  Search results are delivered more accurately according to towns (search is working with 1250 towns of Russia and Ukraine), what is especially important, because 20% of all search inquiries depend on regions. The company’s team is sure that today competitors are unable to render the same opportunities to their users. But in the near future they are also going to mount something similar.