Sberbank of Auto ML

Auto ML is the system of algorithms which quickly and independently creates application solutions on the basis of models of machine learning.

2019: Development and testing of the Auto ML model

Sberbank developed model of machine learning Auto ML. It was said on February 14, 2019 by the vice chairman of the board of Sberbank Anatoly Popov. As he explained, Auto ML represents an algorithm which is able to create other models. And they, in turn, already solve applied problems — for example, predict solvency of the client at issuance of credit or help to separate law-abiding clients from violators.

Sberbank developed an algorithm which works as date scientist

As Anatoly Popov emphasized, the algorithm allows to optimize one of important steps of creation of the model of machine learning applied in business, namely — creation of baseline-model (the first version of model which then develops already with participation of the person). The quality of the baseline-models created by an algorithm Auto ML is comparable to quality of the model created in manual. At the same time the speed of work of an algorithm at 10-15 times exceeds the speed of work of the person, assured the vice-president of Bank board.

Development was already tested by Sberbank within a pilot project in January, 2019: algorithms Auto ML were applied to creation of several baseline-models (the first versions) of the class Next BestAction (targeting of campaigns of sales).

According to experts of bank, the received results prove a possibility of use of technology of automatic modeling for fast formation of basic models of data processing and start of campaigns of sales of corporate and investment business of Sberbank.

One of opportunities for increase in efficiency of all business processes in bank — implementation of artificial intelligence. However creation of tens of thousands of models to cover all aspects of activity, is almost unreal task if to creation and implementation of models to apply only manual work date scientists and developers. Therefore we implement at ourselves one of the most modern approaches to work with models of machine learning in the world — Auto ML. The system of algorithms which quickly and independently creates application solutions on the basis of models of machine learning — Anatoly Popov, the vice chairman of the board of Sberbank concluded.

• A system represents connected by data streams of ML model (machine learning) which work at one of largest in Russia Hadoop a cluster (more 15th petabyte of data).
• At the same time models are integrated into industrial business processes of Bank – crediting, targeting of sales, work of call center, the procedure compliance, etc.
• The system of models processes data on 2 million legal entities and about 100 million physical persons.

For creation of ML models manual approach is actively used and Auto ML is piloted.

Auto ML is an algorithm which is able to create other models. And they, in turn, already solve applied problems — for example, predict solvency of the client at issuance of credit or help to separate law-abiding clients from violators.

In 1 quarter 2019 the pilot was carried out: algorithms Auto ML were applied to creation of several baseline-models (the first versions) of the class Next BestAction (targeting of campaigns of sales).

The algorithm allows to optimize one of important steps of creation of the model of machine learning applied in business, namely creation of baseline-model (the first version of model which develops then already with participation of the person). The quality of the baseline-models created by an algorithm Auto ML is comparable to quality of the model created manually. At the same time the speed of work of an algorithm at 10-15 times exceeds the speed of work of the person.

The received results prove a possibility of use of technology of automatic modeling for fast formation of basic models of data processing and its use for start of campaigns of sales of KIB.

One of the most commercially effective applications in our opinion is a compliance.

The solution program, allowing to apply the differentiated strategy of work with clients in whose activity transactions of doubtful character are set is developed and started. Not to block law-abiding clients whose behavior nevertheless very much reminds behavior unfair, using ML models it was succeeded to minimize blocking of such clients and to allow fair clients to return to their regular activity.

One of the main activities of Bank – issuance of credits. In decision making process about issuance of credit the data obtained through the SIEI (System of Interdepartmental Electronic Interaction) channel are used. Through SIEI since the beginning of 2019 more than 4 million clients of SP and FL within 115 Federal Laws are verified.

One more of demanded services - opening of accounts for SPINNING TOPS/SP. Since the beginning of year more than 200 thousand checks of clients through SIEI are performed. It significantly releases time of specialists of the operational center and brings closer transition of Bank to digital channels of government relations within program implementation "Digital economy of the Russian Federation".

Virtual operators on the basis of NLP models. Chat-bots, as well as other programs with elements of artificial intelligence, well continue our line of non-bank services. Today non-bank services of Sberbank enjoy the greatest popularity at businessmen as help entrepreneurs to resolve the majority of the vital issues. In KIB are implemented and the voice virtual operator Anna, a text chat a bot and the assistant operator of call center develop. Already now it is possible to tell that virtual operators reduce time of the answer to the client's question by 4-9 times.

Projects implementation happens teams of 8-10 people working in agile. The direction of ML models and artificial intelligence one of the most interesting and perspective in KIB. In a year we extended from 30 to 90 employees.