ds-sale-prediction-package 0.1.6

This package provides a comprehensive toolkit for data preparation, feature extraction, model validation, hyperparameter optimization, and results visualization created for the Kaggle competition Predict Future Sales. The competition's objective is to predict total sales for each product and store in the upcoming month.

ds_sales_prediction_package

The package is designed to optimize the machine learning pipeline for time-series sales prediction task for the Kaggle competition Predict Future Sales. It includes the following key components:

DataPreparation Prepares a combined dataset from both training and test data. - Processes training data after DQL and EDA steps. It includes target value clipping, dataset expansion using a Cartesian product, and base feature selection. - Processes test data by assigning the date_block_num value equal 34 and merges it with the preprocessed train data. - Optimizes data types based on feature values to reduce memory usage.

FeatureExtractor Generates a diverse set of features to enhance predictive model performance. - Descriptive features: (e.g. item_category_id, city, shop_type, shop_cluster_umap, shop_cluster_pca) from dict based on csv files (items, and train dataset after EDA) - One-hot encoded features (item_category_id). - TF-IDF encoded features (shop_type). - Binary features (e.g. is_moscow) - Time-based features (e.g., months_since_last_sale, months_since_first_sale). - Lag features, including lagged item_cnt_month and item_price

ModelValidator Facilitates model validation and training using time-series cross-validation. - Split the dataset into training and validation sets based on time-series cross-validation with either expanding or sliding windows. - Train and evaluate models across multiple folds, calculating RMSE for both training and validation sets. - Retrain the model on the combined training and validation data, and make predictions on the test set.

ExplainabilityLayer Explains the results of model predictions through various plots. - Supports various scikit-learn, LightGBM, XGBoost, and CatBoost models. - Includes tools for error analysis, visualizing true vs. predicted values, and identifying patterns in residuals.

optimize_model Utilizes hyperopt for hyperparameter optimization, enabling efficient model tuning to achieve optimal performance.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Installation

pip install ds_sales_prediction_package