This is the implementation of our paper "On Inferring User Socioeconomic Status with Mobility Records" (BigData22).
- Python >= 3.5 (Anaconda3 is recommended and 3.7 is tested)
- Tensorflow & scikit-learn (2.3.1 and 1.0.2 are tested)
- scikit-mobility
Please refer to the source code to install the required packages that have not been installed in your environment such as scikit-mobility. You can install these packages with pip in a shell as
pip install scikit-mobility
Download & unzip the dataset Geolife to the folder ./Data.
Run features/indicators.py, then the trajectories are separated by user number and stored into the directory ./user (e.g., ./user/000); additionally, the basic indicators of the users' features as well as their stay points are obtained; then the stay points are also seperated weekly with the format "user/weeks/userno" (e.g., user/weeks/000) while the features are summed up and dumped in a binary file with the name user/weeks/feature_in_bj. You can change the parameter stay point duration (MINS) and cellsize (CELLSIZE) here to gain a better performance.
python indicators.py
Run Vor_prices.py to obtain the economy information (house price) of the whole Beijing city, which is adopted in get_inputs.py to gain the label of each item in our set.
python Vor_prices.py
Run get_inputs.py to get the input vectors and the labels, which are stored in the directory user/input.
python get_inputs.py
Run spatial_pretrain.py and temporal_activity_pretrain.py to pretrain the embedding vectors with Skip-gram model, and the checkpoints are stored in ./spatial_pre and ./temporal_activity_pre, which will be loaded by the deep network. The directories of spatial_pre and temporal_activity_pre will be generated automatically by running these scripts.
Run deep_net.py to pretrain the deep network, and the checkpoints are stored in ./ckpt_deep_i (i is the number of classes of the label ranging from 2 to 5) , which will be loaded by DeepSEI model. The directory of ckpt_deep_i will also be generated automatically.
python spatial_pretrain.py
python temporal_activity_pretrain.py
python deep_net.py
Run DeepSEI.py to train and test our classification model: DeepSEI. The checkpoints of generated models will be stored in the folder ./ckpt_i (i is the number of classes of the label ranging from 2 to 5) automatically, and the checkpoints of each epoch are stored with the loss values as their filenames. The number of classes can be changed here to perform different tasks. The history of the training process is recorded in the directory ./logs, which can be displayed in TensorBoard. In addition, the variable history in function train also records the training process of each epoch (e.g., loss, accuracy, precision and recall). The training time and f1 score of each epoch are added in callback functions, and later printed.
python DeepSEI.py
The models are trained and stored in the Training process, and can be loaded by the function make_or_restore_model() in the script DeepSEI.py. This function adopts the latest created checkpoints to restore the model for classification, and you may also pick one with the best performance on the validation data (e.g., the model with the lowest loss) as your model from them. We provide a function classification(x) in the script DeepSEI.py with x as the input data, which loads the model and outputs the classification results.
Run Clustering.py to do the task of clustering. The inputs of our clustering model are the output vectors of one hidden layers in the classification model, so the clustering task has to be run after obtaining the checkpoints of classification model. The number of clusters is also changable according to the classification model ranging from 2 to 5.
python Clustering.py