Helping the Homeless with AI Technology

Nochlezhka / VK

A research team from the HSE University Artificial Intelligence Centre led by Ivan Yamshchikov has developed a model to predict the success of efforts to rehabilitate homeless people. The model can predict the effectiveness of the work of organisations for the homeless with about 80% accuracy. The project was presented at a conference dedicated to the activities of social centres.

Homelessness in Russia has yet to be seriously studied. There are no reliable statistics on the number of homeless people in the country, and there are very few studies on the topic. The Nochlezhka (Homeless Shelter) project and the HSE Natural Language Laboratory—Yandex under the HSE AI Centre is one of the first attempts to apply machine-learning methods to study ways to rehabilitate homeless people.

Nochlezhka has operated an electronic Social Worker’s Multifunctional Profile (SWMP) system for several years. Social workers and lawyers use the system to log the support and services they provide to their clients. The database has a total of 12,891 unique clients. The HSE Natural Language Laboratory—Yandex has conducted a study into predicting contract outcomes using data from 3,219 clients’ cases with at least one contract. The model was trained and tested on a sample of 6,528 contracts with these clients.

'Client' and 'contract' are terms used in the Social Worker’s Multifunctional Profile. A contract refers to a service that an SWMP client can receive from a social worker. There are 43 contracts (services) in total, including temporary registration at Nochlezhka's address, passport renewal or receipt, etc.

‘The first task we had to address was predicting the success of the contract,’ says Anna Bykova, Analyst at the HSE Natural Language Laboratory—Yandex. ‘To teach a machine to do something, you need to prepare information. We identified attributes by client category based on comments from database of Nochlezhka's SWMP. We also selected contract statuses that could be considered successful (fulfilled contracts) and unsuccessful (contract not fulfilled for reasons attributed to the client)’.

Each client has an entry with 93 attributes in the database. However, the researchers believe that ‘you can't have too much data’. Any information helps to better predict the outcome of a contract and this, in turn, gives social workers more opportunities to help real people in difficult situations. The attributes were loaded into machine-learning models.

The most difficult contract to fulfil was ‘obtaining citizenship’, while the most feasible was ‘temporary registration at Nochlezhka’s address’

Despite the effectiveness of artificial intelligence, the researchers emphasise the importance of the human element in decision-making. ‘We design a tool and explain how to use it. It’s up to the specialists to interpret the result from an ethical perspective,’ explains Anna Bykova.

We are talking about people, so the decision must be made by a social worker

The researchers are planning to improve the model by selecting hyperparameters, using ensemble methods and various neural network architectures, and conducting experiments with synthetic data obtained through computer modelling. They are also planning to study data from other regions and investigate the impact of humanitarian projects on clients’ willingness to ask social workers for help in the future.

‘We want to test the hypothesis that a client who has visited one of Nochlezhka's service points (Warm Yourself, Night Bus, Night Shelter, Cultural Laundry, etc.) is more likely to decide to “leave the street” and turn to social workers for help. Using the SWMP terminology, this means that at least one “contract” will be associated with this particular client,’ says Nikolai Filippov, Analyst at the HSE Natural Language Laboratory— Yandex.