Didi eta (estimate the travel time)

ETA - Estimate the Travel Time, Time consuming estimation , Estimated travel time , Estimated time of arrival

1 Learning to Estimate the Travel Time,2018

Information ：

1. KDD 2018： Didi put forward WDR The model is significantly improved ETA Prediction accuracy
   
2. sound of dripping water ETA Interpretation of the thesis ：WDR Model

Overview of the paper ：
3. Treat the problem as a space-time regression problem
4. This article requires a given driving route , Then predict the travel time on the route . The next paper does not require a given route , Just know the starting point and the end point .

The overall service architecture is shown in the figure below ：

Algorithm evolution path

1. traditional method ： Passage time = The time of each section + Traffic light time . This rule-based , There are many limitations .
2. Into a regression problem ： Regression model , Didi has considered two popular models in the industry ,Tree Based model and Factorization Machine. But the regression model is limited by input , Statistics are used , Unable to use the details of the road .
3. The regression problem of this paper + Deep learning ： Maximize information lossless , The model looks right link Modeling of sequence information .

The overall structure of the model is shown in the figure below ：

features ：

• spatial information：
  • First match the trajectory to the road network , Get sections and intersections .
  • And then extract features , For example, the length of the road , Width , Road grade , Number of lanes , Road network number and other information , And what happened POIs Information about points of interest .
• temporal information：
  • a year, a month and a day, the holiday indicator and rush hour indicator,
• traffic information
  • We build a traffic monitor and prediction system that provides us the real-time traffic speed estimation in each road segment in the traffic network in every two minutes. Every time 2 Update in minutes .
  • such as the real-time estimated speed, the average speed and free-flow speed, et al.
• personalized information：
  • Different people have different driving habits , Can extract driver profile, rider profile and vehicle profile Other characteristics .
  • For example, drivers have novices , an old hand ; Cars have electric cars , Fuel car , Trucks, etc ; Riders have bicycles , Electric bikes, etc .
• augmented information:
  • Other supporting information , such as the weather information and traffic restriction, et al.
    The total dimension ： Hundreds of characteristics of various categories , There are millions of dimensions .
    After sophisticated feature engineering, we obtain a set of features in hundreds of categories and millions of dimensions.

2 Multi-task Representation Learning for Travel Time Estimation,2018

summary ：

• The road network structure is used , And transcendental spatiotemporal information , Also used. Path Information
• This article does not require the designated driving route . Just give the starting point and the ending point . This is better for scenes with many possible driving routes .

The overall structure of the model is as follows ：

1. The overall structure ： Unsupervised chart embedding + Link embedding network + spatial embedding network +temporal embedding network + Other characteristics -> Deep residual network -> Get the results
2. This model produces meaningful representation that preserves various trip properties in the real-world and at the same time leverages the underlying road network and the spatiotemporal prior knowledge.

• The effective representation methods of different trajectories are given , Various attributes of tracks can be preserved
• Besides , The road network is also used , Transcendental information such as time and space

3. Further-more, we propose a multi-task learning framework to utilize the path information of historical trips during the training phase which boosts the performance.

• The training phase uses the path information of the historical track , Used to improve the effect .

5. pytorch Open source implementation address ：https://github.com/neuguotian/deep-eta-murat