Welcome to Moscow Data Science Meetup September 1

On September 1, we are pleased to invite you to the next meeting of the Moscow Data Science community, where you can share practical experience in solving data analysis problems and chat with like-minded people. For one day, the program is very rich, there will be two sections of reports, and among the speakers are two speakers from neighboring countries. Also the guests of the mitap are waited by the tour of the Mail.Ru Group office and the drawing of souvenirs. Join now! The program of the meeting under the cut.

"Psycho-typing of users of social networks"
Mikhail Firulik, Mail.Ru Group

The psychological type is a structure, a frame of a person’s personality. The psychotype forms a person’s worldview, underlies his behavior, including influencing the perception of information. In Russia, in various industries, the three main concepts of personality typing have received the greatest practical application: Big5, MBTI, and socionics. The report will consider the features of the definition of these typologies and their modeling.
')
“Stochastic Computing Graphs in NLP”
Maxim Kretov, MIPT

Consider the formalism of stochastic computational graphs (graphs that contain sampling from distributions at nodes). For calculating the loss function gradient in such graphs, the usual error back-propagation algorithm is no longer suitable, and more complex methods must be applied. As an example - various ways to train a simple model seq2seq.

"Prediction of sequences in TensorFlow: the engineering side of the question"
Denis Dus, InData Labs

Let's discuss the features of constructing TensorFlow-graphs for effective learning of architectures working with sequences. We will try to answer the main questions that may arise from an engineer in the process of building a model: which data storage format to choose, how to organize the process of entering training and validation data into the model, how to work with sequences of different lengths, how to save the model for later use and others.

"Kaggle's Amazon from Space: the classification of satellite images"
Arthur Kuzin, Avito

In the report we will consider the techniques and tricks for learning deep convolutional neural networks on the example of the Planet kaggle-competition solution: Understanding the Amazon from Space. In this competition, the teams of ODS-community Russian Bears and ods.ai took 2nd and 7th place, respectively, from more than 900 participants. We also analyze the well-known top solutions.

"Multiple Time Series Forecasting"
Vitaly Radchenko, Ciklum

Let's talk about what features exist in the choice of metrics, validation, dataset formation and feature generation for the task of predicting multiple time series. Let us examine the “groups” of signs and approaches to modeling using examples of several real-life cases, as well as problems that should be avoided and what should be paid attention to in the first place.

Collection of participants and registration: 18:00
Reports start: 18:30
Address: Mail.Ru Group office, Leningradsky Avenue, 39, p. 79.

To participate you must register . For those who will not be able to attend in person, a video broadcast will be organized: section 1 , section 2 .