future-sales-prediction-2024 2.4.13

A package for feature extraction, hyperopt, and validation schemas

Future Sales Prediction 2024

Future Sales Prediction 2024 is a Python package designed for building robust time-series sales prediction models. The package integrates preprocessing, feature engineering, hyperparameter optimization, and model training workflows, leveraging DVC for data versioning and Google Cloud Storage for seamless data access.

Project Status: Completed

Features

• Data Handling: Tools to preprocess raw datasets and optimize memory usage.
• Feature Engineering: Generate and refine features for predictive modeling.
• Hyperparameter Tuning: Automate parameter optimization with Hyperopt.
• Model Training: Time-series cross-validation and training for regression models.
• Validation: Validate data integrity to ensure quality and consistency.
• Data Versioning: DVC integration for easy data retrieval from Google Cloud.

Installation

Install the package using pip:

pip install future_sales_prediction_2024

Usage Guide

• Step 1: Authenticate with Google Cloud Before fetching data, authenticate with Google Cloud:

Option A: Use Google Cloud SDK: gcloud auth application-default login

Option B: Use a Service Account key file: export GOOGLE_APPLICATION_CREDENTIALS=/path/to/credentials.json

• Step 2: Pull the Data Step 2: Pull the Data

Option A - locally:

• Use the pull_data.py script to clone the repository, fetch DVC-tracked data, and save it to the current directory:

• pull_data --repo https://github.com/YPolina/Trainee.git --branch DS-4.1

Option B - using online-service(Google Colab, Kaggle and etc.)

• !pull_data --repo https://github.com/YPolina/Trainee.git --branch DS-4.1

This will:

Clone the repository. Pull datasets tracked via DVC from Google Cloud Storage. Save datasets in a folder called data_pulled in the current working directory.

• Step 3: Explore the Codebase and Build Models After fetching the data, you can explore and use the following modules:

---

Modules and Functions

Data Handling

File: future_sales_prediction_2024/data_handling.py

prepare_full_data(items, categories, train, shops, test) -> pd.DataFrame Merges raw datasets into a single comprehensive dataset (full_data.csv), available after dvc pull.

reduce_mem_usage(df) -> pd.DataFrame Optimizes memory usage by converting data types where applicable.

Feature Engineering

File: future_sales_prediction_2024/feature_extraction.py

Class: FeatureExtractor Extracts features for predictive modeling.

Initialization Parameters: full_data: Full dataset containing all columns. train: Training data for aggregating revenue-based features. Output: Returns a processed dataset (full_featured_data.csv), stored in preprocessed_data after dvc pull.

Class: FeatureImportanceLayer Analyzes feature importance using baseline and tuned models.

Initialization Parameters:

X: Feature matrix. y: Target vector. output_dir: Directory for saving feature importance plots. Key Methods:

fit_baseline_model(): Trains a baseline model for feature importance based on RandomForestRegressor. plot_baseline_importance(): Visualizes baseline model feature importance. fit_final_model(): Trains a final model with optimized hyperparameters - model-agnostic. Parameters:

• Model (XGBRegressor by default)
• params: Model hyperparameters (Optional)
• use_shap(bool): Use SHAP values if the model doesn't provide native feature importance plot_final_model_importance(): Visualizes feature importance for the final model.

Output of plot_baseline_importance and plot_final_model_importance: feature_importance_results/baseline_importance.png and feature_importance_results/final_model_importance.png

Hyperparameter Tuning

File: future_sales_prediction_2024/hyperparameters.py

hyperparameter_tuning(X, y, model_class, param_space, eval_fn, max_evals=50) -> dict Performs hyperparameter optimization using Hyperopt for models like XGBRegressor or RandomForestRegressor.

Parameters:

X: Feature matrix. y: Target vector. model_class: Model class (e.g., XGBRegressor). param_space: Search space for hyperparameters. eval_fn: Evaluation function for loss metric. max_evals: Number of evaluations. Returns: Best hyperparameters as a dictionary.

Model Training

File: future_sales_prediction_2024/model_training.py

tss_cv(df, n_splits, model, true_pred_plot=True) Performs time-series cross-validation and calculates RMSE. Returns Mean RMSE for all splits

df: DataFrame with features and target variable. n_splits: Number of cross-validation splits. model: Regression model (e.g., XGBRegressor). data_split(df) -> Tuple[np.ndarray, ...] Splits the data into training, validation, and test sets.

train_predict(X, y, X_test, model_, model_params=None) -> np.ndarray Trains the model with provided features and predicts outcomes.

Validation

File: future_sales_prediction_2024/validation.py

Class: Validator Ensures data quality by checking types, ranges, duplicates, and missing values.

Initialization Parameters:

column_types: Expected column data types (e.g., {'shop_id': 'int64'}). value_ranges: Numeric range for each column (e.g., {'month': (1, 12)}). check_duplicates: Whether to check for duplicate rows. check_missing: Whether to check for missing values. Method: transform(X) Validates a DataFrame and returns a confirmation message if successful.

Conclusion:

This package is a modular and flexible solution for streamlining data science workflows. It provides data scientists and ML engineers with reusable tools to focus on solving domain-specific problems.