huaweiukuniversitychallenge2021's Introduction

HuaweiUKUniversityChallenge2021

Team MatchLab submission to Huawei UK University Challenge 2021

Authors

Team Name: MatchLab

Team members:

Axel Barroso Laguna, Imperial College London, Email: [email protected]
Mikolaj Jankowski, Imperial College London, Email: [email protected]
Michal Nazarczuk, Imperial College London, Email: [email protected]

Method

Task 1

MAC address vendor lookup and removing invalid APs (probably mobile hotspots)
Removing rare MAC addresses (less than 10 occurrences)
Removing outlier samples in the training set (>100m between fingerprints)
Input to the model: pairs of signal powers for common MAC addresses (up to 10) for both fingerprints, MAC similarity index indicating the total number of common MAC addresses, absolute difference between fingerprint indices
Model: 5-layer NN with linear layers, PReLU activations and batch normalization layers
Trained for 10 epochs with batch size equal to 1024, SGD optimizer and sum of L1 and MSE as a loss function

Task 2

Initialize clusters from elevations (all trajectories between fingerprints specified in elevations considered as one of the initial clusters)
Merge clusters based on threshold of estimated WiFi distance between closest points in two clusters

Replicating the results

To replicate our results it is advised to use the newest versions of PyTorch, Pandas, Numpy, and tqdm. For task 1 simply run python train_task1.py and it will produce a submission file called my_submission.csv. For task 2, run python elevation_clustering_task2.py, which produces file submission.csv.

Recommend Projects