Data Science meetup at Avito office on June 24
On June 24, we gather Data Science specialists in our Moscow office to exchange experiences in creating recommendation services. At the meeting, we will summarize the Avito contest held on the Dataring.ru site for building a recommendation system for announcements: we will reward the winners and ask them to tell us more about their decisions. In addition, the program contains interesting reports from representatives of Yandex.Zen, OZON.ru and, of course, Avito. Details under the cut!
At the beginning of the meeting, we will hear speakers who share their experience in creating recommendation services.
Zen Machine Learning
The first speakers will be speakers from Yandex: Evgeny Sokolov and Dmitry Ushanov. Eugene is the leader of the quality group of recommendations and content analysis of Yandex.Zen. Dmitry is a senior developer of the Zen service, and before that he was engaged in the development of linguistic components of search: search wizards and object answers.
Yandex.Zen is a personal recommendations service that aggregates news and media content from a large number of sites, and also allows authors to publish directly to the platform. At all stages of building recommendations, from collecting content and filtering it to ranking, machine learning is used. Recommended algorithms use two main types of signals: user feedback and semantic proximity of content. The speakers will analyze some unusual examples of accounting for these signals: how to use models with hidden variables for texts in matrix decompositions and how to correctly form factors based on them; how to take into account user clicks using sports rating systems; How to combine explicit and implicit user feedback.
Recommendations at OZON.ru
OZON.ru Leading Analyst Ksenia Boksha, without going deep into mathematics, will talk about the different types of recommendations that the marketplace provides to its users: related products, accessories, bundles, personalized recommendations, recommended categories of goods on request, recommended search queries, recommendations in the basket. In addition, from the report you will learn about the technology stack that the team uses, as well as about immediate plans.
What tasks does Avito's team of recommendations solve?
At one of the previous meetings, the head of the recommendations unit Vasily Leksin already said that under the hood are recommendations in Avito. This time, together with analyst Mikhail Kamenshchikov, they will talk about what tasks a unit of recommendations solves, why a contest was organized to build a recommender model, what results they expected to receive from the competition, what worked out and what didn’t. In addition, they will share their experience in participating in the RecSys Challenge 2017 .
Next, we will go on to summarize the results of the competition , reward the winners and ask them to tell us a little details about their approaches and decisions.
A few words about the contest itself. In this dating, participants were asked to build their recommendation system for ads based on an activity history of about 600,000 users for 6 days. After receiving the training sample, the participants had to predict events of user interaction with the ad, which could be of 4 types: a click on an ad, sending a message to the seller, adding the advertisement to favorites and requesting the seller’s contact. The correct predictions of events of different types had different weights.
The final results of the competition have already been published on the portal DataRing.ru, and the decisions of the winners are validated.
In order to get to the meeting, you need to register and receive confirmation . Please register under your real name and do not forget to bring your passport or driver’s license with you on the day of the event.
We will start the reports at 12:30 and plan to complete the meeting by 16:00. During the break, you can have a wok snack.
Address: Moscow, Lesnaya, 7th, 15th floor (BC “White Gardens”). Entrance from Lesnaya street. The nearest metro station is Belorusskaya Koltsevaya.