Semi-supervised Learning with Network Embedding on Ambient RF Signals for Geofencing Services [ICDE 2023]

This is the official implementation of the paper:

Semi-supervised Learning with Network Embedding on Ambient RF Signals for Geofencing Services
Weipeng Zhuo, Ka Ho Chiu, Jierun Chen, Jiajie Tan, Edmund Sumpena, Sangtae Ha, S.-H. Gary Chan, Chul-Ho Lee
IEEE International Conference on Data Engineering (ICDE), 2023

Note: all configurations (required for running) are in config.py.

Current config.py is the lab dataset.

To run the shopping mall experiment, change the 'data_name' in config.py to 'mall'. 'id' in config.py shall also be changed correspondingly. For mall, due to the large number of signal records, we split the dataset into parts, thus additional numbering of filename ('parts') in the config.py would be required here.

To run the UJI experiments, change the 'data_name' in config.py to 'uji{building_id}'. 'id' in config.py shall also be changed correspondingly.

1. Create a new folder input (replacing the old one). Put the input under folder input, please check the given file for suggested format and naming. In each row, 0 stands for 'in' and 1 stands for 'out'

2. Run prep_graph.py, the output would be folders in emb_output. This file processes the signal records and generates corresponding bipartite graph. In training, the signal records are used to construct an initial bipartite graph. In testing, new MAC nodes and signal-record nodes will be added into the graph.

3. Run bisage.py (requires modifying parameters in config.py, in our example, "lab_train" and "lab_test"), the output would be folders in emb_output/{your data name}/embedding. This file processes the bipartite graph and generates the embedding for each node using our proposed representation learning algorithm BiSAGE. In training, it generates node embeddings and learns weight matrices through minimizing the loss function. In testing, the embedding of each node is quickly inferenced through aggregation.

4. Run hbos.py, the result would be in folder result, log files would be in the folder update_output. This file builds the initial histogram for outlier detection and generates the raw outlier score for each embedding using the HBOS outlier detection algorithm.

5. Run sigmoid_scaling.py, the result would be in folder result. This file applies sigmoid scaling to the outlier scores and detects outliers if any. It also updates the histograms built in the HBOS model if there are any highly confident in-boundary signal records.