mavillan/yolanda

Codebase for MeLi 2021 data challenge

Yolanda

This repository contains the codebase for my solution to the MeLi data challenge 2021.

TL;DR

The final model consists of two stacked models: the level-1 model is a single LightGBM trained to forecast the sold quantity, and level-2 model is a neural network that takes level-1 output as input (plus other features) and learns to minimize the RPS.

Model description

The model consist of a 2-level stacking model. The first model is trained to predict the sold_quantity value.

Details of level 1 model:

• LightGBM (single model)
• Target (1d) sold_quantity
• Tweedie loss: https://arxiv.org/pdf/1811.10192.pdf
• Extremely Randomized Trees: https://link.springer.com/article/10.1007/s10994-006-6226-1
• Trained with DART mode: https://arxiv.org/abs/1505.01866
• High cardinality categorically feats encoded with GLMM: http://contrib.scikit-learn.org/category_encoders/glmm.html

Details of level 2 model:

• 1DCNN: SoftOrdering1DCNN https://medium.com/spikelab/convolutional-neural-networks-on-tabular-datasets-part-1-4abdd67795b6
• Target (30d) discrete probs of stock out
• RPS loss + Probability regularization term
• Optimizer: AdaBelief: https://github.com/juntang-zhuang/Adabelief-Optimizer
• Scheduler: ReducerLROnPlateau

The level 2 model has a head of size 30, where the value at position i represents the probability of stock-out at day i+1.

@torch.jit.script
def rps_loss(preds, targets):
    return torch.mean(torch.sum((targets - preds)**2, dim=1))

Optiming only the loss above gives very noisy probility estimates. For that reason I included a term to penalize high variations between neighbor probabilities:

@torch.jit.script
def discont_loss(probs):
    return torch.mean(torch.sqrt(torch.sum(torch.diff(probs, dim=1)**2, dim=1)))

The final loss is rps_loss(preds, targets) + alpha * discont_loss(preds) with alpha=5 (barely tuned).

Stacking description

Submission generation

All the codes for data preparation and modeling are in the notebooks/ directory, however, only a subset of these notebooks are used to generate the final submission. Below is the list of the utilized notebooks as well as the order of execution:

Data preparation:

• Download and place the competition data in the data/ directory.
• Run notebooks/preproc-m1.ipynb -> prepares data for level-1 model.
• Run notebooks/eda.ipynb -> performs EDA, computes unpredictable.csv (skus with not enough information) and computes scales.csv (scale values for RMSSE).
• Run notebooks/preproc_assessment.ipynb -> computes skus_assess_m1.yaml (skus for assessment of level-1 model), the synthetic validation sets (validation_seed*.csv & validation_seed*\_harder.csv) and validation_m3.csv (the targets for the level-2 model).

Modeling:

• Run notebooks/encoders.ipynb -> trains the categorical encoders (GLMM) for stage1 and stage2 of level-1 model.
• Run notebooks/train_lgbm-m1-sm.ipynb -> train a LightGBM model for predicting the sold quantity at sku level. The stage1 is for generating the oof predictions, while stage2 is for generating the final predictions.
  • output-1: results/oof_preds_lgbm-m1.csv the oof predictions (for the period of 2021-03-02 to 2021-03-31).
  • output-2: results/preds_m1_lgbm_sub{SUB_NBR}.csv the predictions for the test period.
• Run notebooks/train_1dcnn-m3.ipynb -> take the oof predictions of level1 model and train a 1dcnn stacking model that learns to minimze the RPS. Then applies the trained model over the level-1 predictions on the test period.