Sberbank Data Science Day live stream November 10

Hello!

A big conference of Sberbank Data Science Day will be held on November 10 (tomorrow!) In Moscow at the October Cinema Center, where the winners will be awarded SDSJ 2018, speeches by a large number of international and Russian experts in the field of Data Science, the ML section and the use of artificial intelligence in science and business. And many more interesting things!

You can watch the live broadcast here . Under the cut and on the website program. We also tell how the winners of the Sberbank Data Science Journey were rated.
')

Program

The conference is divided into several thematic blocks, here is the schedule:

Main hall

11:00 - 11:30 Opening of the conference.
11:30 - 12:30 Panel discussion "Data Analysis and Artificial Intelligence Technologies in the Digital Economy"
12:30 - 13:15 "Biologically-based methods and architectures in deep learning." Sergey Bartunov, Deep Mind
13:15 - 14:00 "Conversational Agents as an Intelligent Digital Companion to Understand Human Emotion and Express". Soo-Young Lee, KAIST
3:00 pm - 3:45 pm "Scalable automatic machine learning." Andrey Spiridonov, H2O
15:45 - 16:30 Panel discussion “Trend for Innovation: Using DS / AI and Improving Customer Experience”
17:15 - 18:00 Award ceremony for the winners of the competitions Sberbank Data Science Journey and KlassikAI (competition on versification with the help of artificial intelligence)

Hall "Science"

12:30 - 13:45 .DS / AI technology: AutoML
13:45 - 14:45 DS / AI technology: Computer Vision
14:45 - 15:45 .DS / AI technology: Natural Language Processing (NLP)
15:45 - 16:30 .DS / AI technology: Reinforcement Learning
16:30 - 17:15 .DS / AI technology: Speech Analytics

Hall "Business" (Hall 1)

12:30 - 13:45 Application DS / AI in banking and finance
13:45 - 15:00 The use of DS / AI in medicine and bioinformatics
15:00 - 16:15. Application of DS / AI in banking and finance
16:15 - 17:15 .Brainwriting: we create a platform for AI research

Hall "Business" (Hall 2)

12:30 - 14:45 DS / AI application in retail
14:45 - 16:30 Application DS / AI in industry
16:30 - 17:15 DS / AI in media and telecom

Hall "Community"

12:30 - 13:15 Poster Session Lightning Talk posters presentation
13:15 - 15:00 Presentation of open projects in the field of DS / AI "AI Open Projects"
3:00 pm - 3:45 pm Analysis of the competition decisions ClassicAI
15:45 - 17:15 Sberbank Data Science Journey competition review

Winners Sberbank Data Science Journey

This year we offered to solve problems using AutoML technology. Until the end of November 3, participants unloaded their decisions, and in the next 12 hours they selected the best from their decisions. Now the choice is for the jury. At the conference, we will award the winners of the Sberbank Data Science Journey.

Participants were provided with ready-made data sets from Sberbank. All 24 datasets involved in the competition were collected by various departments: the retail unit, the risk unit and the technology unit. All of them were specially prepared and impersonal. The basis is information such as:

• Approved limit share
• Card delivery time
• Different types of scoring
• Feedback on the offer card
• Response to other product offerings
• ATM breakdowns
• Information about cash withdrawals at ATMs
• Account balances and other information

For evaluating decisions, the following dataset groups were selected: check (open for participants), public (hidden from participants, but you can see the result during the competition), private (set, which summarizes the competition)

In each such set there are three regression problems and five binary classification. Solutions worked on data sets of various sizes: from 1MB and 300 lines to 1GB and 1 million lines. Before the start of the competition, the jury prepared datasets, the testing system has already checked them automatically, and now you can see the results on the website (taking into account the restrictions associated with intrigue).

Decisions were taken in the format of archives with a code. The participants had to build an algorithm that implements a full cycle of solving a machine learning problem automatically, receiving data as input and returning a ready answer at the output.

The decisions of the participants were to fit into the specified restrictions:

• resources available to the solution
• the solution does not have access to Internet resources
• maximum size of a packed and unpacked archive with a solution: 1 GB
• The archive is unpacked into a file system in RAM (ramfs) available to the write solution.
• the rest of the container is read only.
• CSV with a set of data does not exceed 3 GB
• Restrictions are needed in order to achieve a fair comparison, putting the participants in equal technical conditions.

Here is what the evaluation system in this competition is:

1. For each task (dataset), the test-specific sample is considered to be a task-specific metric (RMSE for regression, ROC-AUC for binary classification).
2. For each task (dataset), the value of the participants' metrics is translated into a common scale according to the following scheme. For the best metric solution (among all the sent and successfully tested solutions) 1 point is given, the baseline solution is estimated at 0 points. Participants who are on the metric between the best and baseline solutions receive a proportional number of points between 0 and 1. Decisions on the quality of the bottom baseline are estimated at 0 points. If the best solution and the baseline solution coincide, then all participants receive 0 points. If the participant's decision gives an error on the task or does not pass through the time limit, then it receives 0 points for this task.
3. The final result of each participant is counted as the sum of the results for each task after being converted into a common scale. In general leaderboard participants are ranked by the final result.

The results of the competition are available here .

In addition to the main standings, the participants competed for the prize in the nomination “The Best Public Solution”. Throughout the competition, they published their approaches to solving the AutoML problem on GitHub, and the winners were determined by the number of GItHub stars.

At the conference there will be a separate section dedicated to SDSJ'18, where the winners will talk about their decisions and answer all questions.

Once again we leave a link to the online broadcast of the conference, so that all those interested can watch Sberbank Data Science Day.