Title: - AAAI: Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting
Year: - 2019
Venue: - AAAI
Main Problem
The author argues that the problem of forecasting traffic flow is dynamic and non-linear in nature. The spatial and temporals aspects of traffic flow are dynamically correlated to each other and affect the performance of traffic on a given highway or road. To be able to understand this non-linearity and correlation between spatial and temporal features of traffic flow, the authors propose a novel ASTGCN model.
Related Work
Author states that there have been many researches made in the domain of traffic flow forecasting but they all suffer from some flaws. Some of the few notable ones are: -
Forecasting Methods: - Some researchers tackle this problem in the means of a forecasting problem having many assumptions. The authors argues that not all of the assumptions can be statisfied by the data in hand and that they traffic data is too complex to satisfy that. Author also says there are methods that learn spatial-temporal features but their limitation is that the input must be 2D and not 3D grid data.
Convolutions on Graphs: - Graph Convolutions do work on grid data but is unaable to caputre dynamic spatial-temporal aspects of traffic.
Attention Mechanism - Recently attention mechanism are being used to capture dynamic correlations between two different features. Motivated by this the author proposes ASTGCN.
Proposed Approach
The author proposes a novel model named: - ASTGCN for the Spatial-Temporal Traffic Flow Forecasting problem. The model contains three independent compnents with the same stucture for three different time series - recent, daily-periodic and weekly-periodic. The three components share the same network structure and each of them consists of several spatial-temporal blocks and a fully-connected layer. There are a spatial-temporal attention module and a spatial-temporal convolution module in each spatial-temporal block. In the end, the outputs of the three blocks are merged.
Input
Spatial Time-Series data of traffic at a given road.
Output
Forecasted output of the traffic at the next interval.
The paper introduces a new approach to tackle the spatial-temporal dynamic correlations of traffic data.
The paper has good baselines to compare itself with.
The paper does not proposes a way where this method could be used for datasets having only one time.
The applicability of this approach on other generalized solutions is not good. They should have explained on how this method can be used in different domains.
Title: - AAAI: Attention Based Spatial-Temporal Graph Convolutional Networks for Traffic Flow Forecasting Year: - 2019 Venue: - AAAI
Main Problem The author argues that the problem of forecasting traffic flow is dynamic and non-linear in nature. The spatial and temporals aspects of traffic flow are dynamically correlated to each other and affect the performance of traffic on a given highway or road. To be able to understand this non-linearity and correlation between spatial and temporal features of traffic flow, the authors propose a novel ASTGCN model.
Related Work Author states that there have been many researches made in the domain of traffic flow forecasting but they all suffer from some flaws. Some of the few notable ones are: -
Proposed Approach The author proposes a novel model named: - ASTGCN for the Spatial-Temporal Traffic Flow Forecasting problem. The model contains three independent compnents with the same stucture for three different time series - recent, daily-periodic and weekly-periodic. The three components share the same network structure and each of them consists of several spatial-temporal blocks and a fully-connected layer. There are a spatial-temporal attention module and a spatial-temporal convolution module in each spatial-temporal block. In the end, the outputs of the three blocks are merged.
Input Spatial Time-Series data of traffic at a given road.
Output Forecasted output of the traffic at the next interval.
Dataset PeMSD4 PeMSD8
Code https://github.com/wanhuaiyu/ASTGCN
Pros and Cons
The paper introduces a new approach to tackle the spatial-temporal dynamic correlations of traffic data. The paper has good baselines to compare itself with. The paper does not proposes a way where this method could be used for datasets having only one time. The applicability of this approach on other generalized solutions is not good. They should have explained on how this method can be used in different domains.