June 7, 2018 Uber Technologies, Inc. filed an application for US Patent
No. 20180157984 , entitled
Predicting user state using machine learning .
A patent application describes a self-learning system (neural network) that constantly monitors how a particular person uses the Uber application — and identifies behavior that is unusual for that person. The system is trained on a number of input signals, including:
- the number of typos in the text;
- accuracy of clicking on links and buttons;
- device tilt angle;
- walking speed (probably, data is taken by GPS);
- data entry rate;
- day of the week and time of day;
- passenger's address.
Although the patent application does not mention the specific purpose of the system, but based on the parameters of the neural network training, it can be assumed who it should specifically identify. Apparently, we are talking about identifying very drunk passengers.
The patent application contains a table with examples of various values of input data - and the result that the system gives in the column “Identification of an Unusual User State” (1 or 0). Perhaps Uber has already really tried out the machine learning model - and in the patent application the real results of its work are indicated.
The results in the table are really similar to the results of a real neural network, because they are not always obvious. For example, user number 5 collects data quickly and accurately, while moving at an average speed. But the system still ascribes to it an “unusual” state (result 1). Probably, the call time and the day of the week (Saturday, 1:38 AM) have a lot of weight. In other words, on a Saturday at such a time it is very difficult to prove to the neural network that you are sober.
If the system is implemented, then upon receipt of the order, the driver may be warned as accompanying information that the order was potentially received from a passenger with “reduced adequacy”. Accordingly, the driver can either agree to such a trip, or refuse it.
It can also be assumed that the company will try to introduce a special increased tariff for the transport of drunk passengers.
True, immediately come to mind ways to deceive the system. For example, a user may ask a sober friend to make an order from an application so that the system does not offer him a ride at a higher rate. Special bots may appear - programs that independently call a taxi in Uber, for example, by the user's voice command. In this case, a number of input signals will give a distorted picture: the number of typos in the text will be zero, the accuracy of clicking on the links and buttons is ideal, and the time between opening the program and calling a taxi is minimal.
However, the other input parameters still can not be hidden. For example, if a person calls a taxi at 2 nights, being inside or near a pub, then a taxi driver can draw correct conclusions without any neural network.
The patent explicitly states that the findings of the neural network affect the service that Uber provides to the user. Drivers can be warned about the status of the passenger. If “the likelihood of an unusual passenger state is relatively high,” then his application will not go to the general pool, but will be sent only to drivers “with relevant experience or training,” the patent application states.
For Uber drivers, working with drunk passengers is one of the main shortcomings of their professional activities. If an ordinary taxi driver in the last century could even rejoice at such a passenger, because it promised him extra income, then in the automated route planning, tariffing and non-cash payments system Uber is very difficult to get extra money from a drunk passenger. Therefore, the introduction of an increased tariff in case of intoxication looks quite justified.
“It would be great if drivers got extra money for bringing drunk passengers. There is not much difference [between drunk and sober passengers], but the driver definitely has gray hair after dozens of such trips, ”
said Harry Campbell, author of the RideShare Guy blog.