Kaggle-Indoor solution

kaggle: Indoor Location & Navigation
https://www.kaggle.com/c/indoor-location-navigation/overview

Transcript

1. Indoor Location & Navigation
   Dataset
   Post Processing (PP)
   Training (2 stage)
   Hidden waypoints by LI
   Hidden waypoints by KF
   - Make wifi-based dataset both train & test
   - Remove bssid if time-diff (waypoint
   and last-seen timestamp) is more 10s
   - Minimum number of wifi=7
   - Interpolate hidden waypoints by Linear
   Interpolation(LI) and Kalman Filter(KF)
   - Calculate timediff (between waypoint
   and wifi group)
   BSSID
   RSSI
   Site id
   Floor
   Embedding
   Embedding
   xy
   floor
   CustomLoss
   MESLoss
   dim=64
   dim=64
   FC layer1(128→256)
   LSTM layer×2(256→128→16)
   FC layer2(xy:16→2, floor:16→1)
   1st stage
   2nd stage
   oof
   pred
   Add test data with pseudo labeling
   Remove train data if oof’s error is over 40m
   Loss
   - MSELoss-based
   - Given weight according to timediff
   - if timediff is large, weight become
   smaller (don’t learn too much)
   Ensemble
   Cost minimaization
   Snap to grid
   Device id leakage
   Repeats 6 times
   Repeated PP
   ■ Snap to grid
   ■ Cost minimaization
   80 pieces
   Delta correction by linear regression
   using sensor delta and target delta
   Automatically generate multiple
   patterns of extra grid.
   mean by timestamp
   Team:
   EXODIA REBORN at MOTOSUMIYOSHI
   Given waypoints
   Replace the predicted value of floor
   with the predicted value of another
   model (lightGBM and BiLSTM).
   Copyright 2021 @kuto_bopro
   4 pattern grid when do snap to grid
   ❶ snap to grid’s threshold=None / sparse extra grid
   ❷ snap to grid’s threshold=None / dense extra grid
   ❸ snap to grid’s threshold=None / edge extra grid
   ❹ snap to grid’s threshold=5 / only train grid
   model1
   repeated pp by
   4 pattern grid
   (❶〜❹)
   model2
   model3
   stacking
   light
   GBM
   final
   submission
   weighted mean
   ×3
   ×3
   Private LB: 16th (3.562)
   

   View Slide