Spatial temporal machine learning

Comparison of Deep Neural Networks and Deep Hierarchical Models for Spatio-Temporal Data

Paper written by Senzhang Wang, Jiannong Cao, Fellow, IEEE, Philip S. Yu, Fellow, IEEE

Spatio-temporal data mining is becoming growingly important

It has broad applications in various domains including

  • environment and climate (wind prediction and precipitation forecasting)
  • public safety (crime prediction)
  • intelligent transportation (traffic flow prediction)
  • human mobility

Why traditional data mining skills perform poorly

  • ST data are in continuous space

  • Has both spatial and temporal properties

  • Not independently generated

  • Traditional methods rely heavily on feature engineering. Hard to process natural ST data in their raw form.

  • Compared with traditional methods, the advantages of deep leaning models for STDM are as follows:

    • Automatic feature representation learning

      • Do not require hand-crafted features
      • The spatial proximity and the long-term temporal correlations are complex and hard to be captured.
      • With CNN and RNN, these can be automatically learned from the raw data directly.
    • Powerful function approximation ability

Categorization of spatio-temporal data

Data types:

  • Event data: crime data
  • Trajectory data: gps, taxi
  • Point reference data
  • Raster data: air quality pm 2..5
  • Video: can be thought as raster data

Data representations: (different deep learning models require different data representations as input)

  • Sequence
  • Graph
  • Matrix: facilitate the utilization of CNN model
  • Tensor

Preliminary of Deep learning models

  • RBM: restricted boltzmamnn machines.

    image-20200920110332022

  • CNN: convolutional neural networks

    image-20200920110553106

  • GraphCNN: generalize CNN to graph structured data.

  • RNN and LSTM:

    RNN is designed to recognize the sequential characteristics and use patterns to predict the next likely scenario. They are widely used in the applications of speech recognition and natural language processing.

  • Seq2Seq:

    image-20200920123831792

  • Autoencoder and stacked AE:

    image-20200920123924475

image-20200920124022208

Framework

image-20200920124313551

  • Raw ST data will first be represented as a particular data format to fit the deep learning model.
  • RNN and LSTM models are good at handling sequence data with short-term or long-term temporal correlation, while CNN models are effective to capture the spatial correlation in the image like matrices.

ST data preprocessing

  • Usually one type of ST data instance corresponds to one typical data representations.
  • Trajectory and time series data can naturally be represented as sequence data.
  • Spatial map can be represented as a 2D matrix.
  • ST raster can be represented as a 2D matrix or 3D tensor
  • Trajectory data can also be represented as a matrix to apply CNN. For example, a city is first partitioned into grid cell regions. Then the ST field can be modeled as a matrix with each cell region representing an entry. If a trajectory paths over the cell region, the corresponding entry value is set to 1; otherwise, set to 0.
  • Spatial map is sometimes represented as a graph