Update

We are excited to announce AgentMove, an LLM-based agentic framework designed for zero-shot mobility prediction. Leveraging the world knowledge and sequential modeling capabilities of LLMs, AgentMove paves the way for a promising new direction in mobility prediction.

DeepMove

PyTorch implementation of WWW'18 paper-DeepMove: Predicting Human Mobility with Attentional Recurrent Networks link

Datasets

The sample data to evaluate our model can be found in the data folder, which contains 800+ users and ready for directly used. The raw mobility data similar to ours used in the paper can be found in this public link.

Requirements

• Python 2.7
• Pytorch 0.20

cPickle is used in the project to store the preprocessed data and parameters. While appearing some warnings, pytorch 0.3.0 can also be used.

Project Structure

• /codes
  • main.py
  • model.py # define models
  • sparse_traces.py # foursquare data preprocessing
  • train.py # define tools for train the model
• /pretrain
  • /simple
    • res.m # pretrained model file
    • res.json # detailed evaluation results
    • res.txt # evaluation results
  • /simple_long
  • /attnlocallong
  • /attnavglong_user
• /data # preprocessed foursquare sample data (pickle file)
• /docs # paper and presentation file
• /resutls # the default save path when training the model

Usage

1. Load a pretrained model: > python > python main.py --model_mode=attn_avg_long_user --pretrain=1 >

The codes contain four network model (simple, simplelong, attnavglonguser, attnlocallong) and a baseline model (Markov). The parameter settings for these model can refer to their res.txt file.

|modelincode | modelinpaper | top-1 accuracy (pre-trained)| :---: |:---:|:---: |markov | markov | 0.082| |simple | RNN-short | 0.096| |simplelong | RNN-long | 0.118| |attnavglonguser | Ours attn-1 | 0.133| |attnlocallong | Ours attn-2 | 0.145|

1. Train a new model: > python > python main.py --model_mode=attn_avg_long_user --pretrain=0 >

Other parameters (refer to main.py): - for training: - learningrate, lrstep, lrdecay, L2, clip, epochmax, dropoutp - model definition: - locembsize, uidembsize, timembsize, hiddensize, rnntype, attntype - history_mode: avg, avg, whole

Citation

If you find this work helpful, please cite our paper. latex @inproceedings{feng2018deepmove, title={Deepmove: Predicting human mobility with attentional recurrent networks}, author={Feng, Jie and Li, Yong and Zhang, Chao and Sun, Funing and Meng, Fanchao and Guo, Ang and Jin, Depeng}, booktitle={Proceedings of the 2018 world wide web conference}, pages={1459--1468}, year={2018} }