This is a new post that discusses how AI is used in everyday life.
We gather various stories from ML engineers who develop real-world products and ask them about the nature of their work, common activities associated with the role, and specific technologies that enable them to perform their jobs more effectively.
The main task of the ML team in a taxi service is to maximize the number of trips. A more specific task is to predict demand and supply for the next 20 minutes. The prediction is made for a hexagonal area on the map. The division into hexagons is done using the library H3 from Uber.
Main methods
In 10 minutes, it will be the same as now. Despite its seeming rudimentary nature, this approach covers many basic needs.
Catboost receives a prediction from ARIMA as the first baseline, after which it is trained in the standard boosting paradigm, starting with the residuals obtained from the ARIMA prediction.
Those moments that we hunt for in this blog – when ML engineers find time not only for tools from a gentleman’s set (see the previous paragraph) but also for digging into an article. Such an honor was bestowed upon the work of comrade Jintao “Hexagon-Based Convolutional Neural Network for Supply-Demand Forecasting of Ride-Sourcing Services”
This paper proposes using a CNN on top of a hexagon map to predict the supply-demand gap (the number of unsatisfied orders).
The “local map” of hexagons is composed of the selected central hexagon and its neighbors. The map is transformed into a 2- or 3-dimensional tensor using one of the three methods proposed by the authors.
A model with multiple groups of Conv-layer + Batch-norm is used. The output is the prediction of gaps for all hexagons within the map fed to the network at the current iteration.
To use such a model in the inference mode for a specific hexagon, its prediction is taken from all possible maps that contain the necessary hexagon.
This approach was able to beat the strong ARIMA+Catboost baseline.