Customer Purchase Prediction in Steel Manufacturing: A Machine Learning Sales Decision System Based on Behavioral Data

This project is an end-to-end customer purchase prediction system for the steel manufacturing industry, maintained by ashokmedasani. The source code is available on GitHub (link: https://github.com/ashokmedasani/steel-client-purchase-prediction) and was released on June 8, 2026. Based on the actual business scenario of Alpha Steel Company, it integrates 8 machine learning algorithms to predict purchase volumes by analyzing customer platform behavioral data (such as website visits, negotiation interactions, tonnage confirmation, etc.). It transforms traditional experience-driven B2B sales forecasting into data-driven, providing a practical technical solution for the digital transformation of the manufacturing industry. The project supports an automated analysis pipeline and offers Gradio Web interface and Hugging Face deployment options.

As a typical B2B industry, the steel manufacturing sector faces pain points such as long order cycles, complex customer behaviors, large market fluctuations, and difficulty in sales resource allocation. After implementing a Web order system, Alpha Steel Company has accumulated rich customer behavioral data (website visits, negotiation records, tonnage confirmation, platform function usage, etc.), which provides the foundation for the machine learning prediction of this project. The aim is to solve the problem of "how to use behavioral data to predict procurement activities".

Feature Engineering: Uses the STEEL CLIENTS dataset, covering multi-dimensional features such as negotiation interactions (number of negotiations, annual sessions), transaction behaviors (confirmed tonnage, Web application sessions), logistics preferences (distribution center changes, delivery/self-pickup preferences), and tool usage (Excel tools, customer catalogs).

Model Comparison: Implements 8 algorithms, including linear models (OLS, Ridge, Lasso, Polynomial Regression), tree models (Decision Tree, Random Forest, Gradient Boosting), and neural networks. The optimal solution is selected through multi-model comparison.

Automated Pipeline: Generates visual analysis (correlation matrix, model performance comparison, feature importance, etc.), structured reports (model_results.csv), and model persistence files (best_model.pkl), reducing the threshold for non-technical users to use.

Core Dependencies: Python ecosystem (Pandas/NumPy for data processing, Scikit-Learn for machine learning, Matplotlib for visualization, Joblib for model persistence, Gradio for Web interface).

Local Execution: After installing dependencies, execute steel_clients_pipeline.py to run the complete pipeline, and start app.py to open the Web application.

Hugging Face Deployment: Create a Gradio Space, upload relevant files for automatic deployment, supporting zero-configuration usage.

Empowering Sales Teams: Customer segmentation, resource allocation, and seizing follow-up opportunities;

Inventory Optimization: Demand forecasting, production plan adjustment, supply chain coordination;

Customer Success: Churn warning, cross-selling, personalized services.

Improvement Suggestions: Data level (introduce time-series/external data, customer lifecycle features); Model level (try XGBoost/LightGBM, LSTM, model integration); Engineering level (API interface, automatic retraining, A/B testing); Business level (churn prediction, price sensitivity analysis, recommendation system).

Industry Insights: Demonstrates the digital transformation path for traditional manufacturing (data collection → feature engineering → model building → business integration → continuous optimization), providing a reference architecture for B2B industries such as chemical and building materials.

This project is a typical case of B2B machine learning application. The keys to success are designing effective features based on in-depth business scenarios, avoiding algorithmic bias through multi-model comparison, focusing on interpretability to gain business trust, and reducing the usage threshold with an automated pipeline. It provides an excellent reference implementation for applying machine learning in traditional industries.