The Yandex search engine, using machine learning technologies, has learned to predict the following user queries. Now the search offers useful queries that the user may not have had time to think about.
Predictive queries differ from other search engine features in that they do not offer the most popular queries based on statistics, but recommend those options that a person is most likely to click on. In order to find out such requests, the data of the previous session and the general search history of all users are used.
For example, if a person is looking for where to buy a snowboard, the search will suggest "How to choose a snowboard by height and weight." And for those who wish to purchase tickets to the Tretyakov Gallery, the system will recommend the request "When to get to the Tretyakov Gallery for free" or "How to get to the Tretyakov Gallery without a queue."
The database of potentially interesting queries is filtered using a machine learning algorithm based on the search for nearest neighbors (k-Nearest Neighbors). The system then chooses from hundreds of possible options the five most popular queries that the user is most likely to click on. The system learns this probability based on the response of users - now the system is already running and collecting feedback to improve the quality of recommendations.
According to the developers, this is a new level of interaction between the search engine and users, because in this way the system not only corrects typos and advises the most frequent requests, but learns to predict the interests of a person and offers him something new.
Source: 3dnews.ru