# 29th Place Solution and Initial Solution

29th Solution Overview:
Model Architecture (from tubo's GitHub Repository):
2DCNN Feature Extractor -> Unet -> Unet Decoder
LSTM Feature Extractor -> 1DCNN -> LSTM Decoder
Input Features:
enmo, anglez, anglez_diff, hour, weekday, and a noise flag.
Noise is flagged if identical (enmo + anglez) occur within a 5-minute span in the same series_id.
EDA indicated that the day before a holiday typically shows later onset times, prompting the inclusion of weekday and holiday features.
Input steps of 5760 and 11520 were used, with four models based on ensemble model and input step pairs.
Target Size:
The target size is identical to the input size.
Training Targets:
Events with Gaussian soft labels (2 channels)
Sleep flag (1 channel)
Inference Process:
Chunks are created by sliding the input_step/4.
For example, in the 5760 model, chunk steps begin from 0, 1440, 2880, etc.
All chunks are predicted and averaged by step.
In CNN-based models, edge predictions are trimmed by 12.5%.
Postprocessing:
Use only events prediction (sleep flag is not used).
Utilizes find_peaks from SciPy signals, and applies a weighted average to steps near detected signals based on the prediction scores.
For instance, if the candidate step is 5000, the weighted average of steps 4998-5002 is calculated.
Final solution is just a small modification from tubo's approach, so let me share my initial solution. (I took most of my competition time in this approach…)
Initial Solution (LB 0.677):
Based on a two-stage model:
Stage 1: Chunk Detection by binary classification
Segmenting the dataset into 1-hour chunks with weak labels to identify useful segments. Label 1 is added if an event occurs within a chunk.
Stage 2: L1 Regression
Pinpointing the event position within the candidate chunks from Stage 1, followed by post-processing.
Challenges and Learning Points:
As a newcomer to neural networks, my approach lacked certain elements:
Random Dataloader: No ideas to implement a random chunk dataloader. Each epoch started from the same position, leading to overfitting and lack of robustness.
Model Exploration: I used a very basic model (a single MLP, attention layer, LSTM, and head). Exploring more sophisticated models, as seen in tubo's work, would have been beneficial.
Concluding Thoughts:
This competition was a tremendous learning experience. My thanks to Kaggle for this opportunity, and to all the participants for their contributions and shared insights