The dataset represents purchasers' activity timeline. Each transaction may be encoded by a pair of one categorical (generalized shop/brand/product code) and one numerical (priced quantity) variable. Further we combine a number of these transactions within a certain time span into one multinomial variable (denoted here as aggregated purchase vector, APV).
The variety of such transactional data can be effectively handled via a combination of a trx-encoder (which provides concatenated embeddings for pairs of cat/num variables) and a seq-encoder (provides RNN-encoded embeddings of the whole event sequence).
# https://www.kaggle.com/c/acquire-valued-shoppers-challenge/data
# download and preprocess dataset (and a sample from it) into event sequences with
# properly defined state and corresponding target variables
# useful statistics will be saved in <make_dataset.log>
cd ./scenario_shoppers/
sh bin/make_dataset.shThe main goal of the task is to represent a long range target variable (e.g. CLTV) via modeling APV-distribution over shorter time span. In practice one may consider 120-day CLTV and its evaluation by 12 one-step predictions of corresponding 10-day APV.
# scanning over loss-functions and hyperparameters
# compute faster overriding conf variable {data_path} by "data_path=data/sample"
sh bin/scan_loss.sh
sh bin/scan_hparams.sh
# after those one must analyze <cv_result.json> and insert the actual checkpoint
# numbers {*_CKPT} into <bin/apply.sh> for every benchmark and one-step predictorsAs long as one-step predictor is learned one can evaluate CLTV as the expectation of sum of APV over continuum of possible 12-step trajectories. We compute the expectation by Monte-Carlo method.
# ILMC requires a sufficient number of {repeats} and proper {chunk} size
# ILMC results saved in <monte_carlo.csv>; metrics (MAE, BucketAccuracy, RankAUC,
# Jensen-Shannon divergence, MAPE, SMAPE, R2) saved in <monte_carlo.json>
sh bin/apply.sh