Machine Learning in the Work of Taxi Aggregators

The rapid development of technology has affected many areas of business, including taxis. If earlier, back in the early 2000s, to order a taxi it was necessary to call the dispatcher, now everything has become much simpler — you just need to open an application on your phone and enter the necessary addresses. Digitalization has made adjustments not only in how the taxi business communicates with the consumer, but also in the very structure of the work process. Before the introduction of digital technologies, all cars registered with the company underwent a physical control of safety and cleanliness of the car, but now the car inspection before sending it to the streets of the city has become remote.

There are two stages of digitalization of car inspections in taxi aggregators: before and after the introduction of machine learning. To pass a remote quality control, it is necessary to provide photographs of the interior and exterior of the car from different angles. Having received these photos, the service checks them using three performers who assess the external and internal cleanliness of the car, compliance with the necessary information (number, brand, color) with what is written about the car in the application. If everything is fine with the car, then the driver can continue to work further and accept orders, but if something is wrong with it, the car is sent for additional examination by service employees — assessors. However, as the business grows, so does the number of cars waiting for inspection, but often there are not enough resources for a quick check. That’s where machine learning help is needed.

The task of machine learning in the work of assessing the suitability of a car consists in automating checks without losing their quality. Here there are three quantities that are subject to automation:

• The share of flow that artificial intelligence can answer automatically;
• The share of drivers with cars unsuitable for various criteria, which are allowed to accept orders after inspection;
• The share of drivers with cars suitable for various criteria, which are not allowed to accept orders after inspection.

Of all the components of automation listed above, the first item is the easiest to automate, since remote inspection involves choosing an answer option about the condition of the car based on the photographs sent by the drivers — and this task can be handled by computer vision. The main difficulty for artificial intelligence in this task is that it cannot answer all questions at once: for example, to answer the question of whether the car number matches the one indicated by the driver in the application, a special Optical Character Recognition (OCR) model is needed. Moreover, sometimes there are errors in the model, which does not allow to accurately answer the question of what exactly is wrong with the car, and in a situation where the answer depends on the income of a particular taxi driver, this becomes critically important.

The solution to such a task is to divide one task into several subtasks. Most often, drivers send photographs of the exterior of the car, and teaching the model to answer step-by-step simple questions about whether the number matches what the driver indicated in the profile — thus, the model gives more accurate answers, and this helps to weed out cars with which something is wrong.

As a result, this can lead to a reduction in the load on assessors who deal with really difficult issues. Moreover, it significantly reduces the time interval required to detect a car unsuitable for the service. Finally, this technology does not cancel the fact that the business needs people: yes, indeed, artificial intelligence copes with some tasks, but at the same time, it significantly facilitates the work of a person, giving him the opportunity to solve more complex tasks.