DEPI_Graduation_Project

Sales Forecasting and Demand Prediction

This repository hosts the code and documentation for a deep learning project focused on predicting customer next-action events (e.g., view, cart, purchase) in an online cosmetics retail environment. By analyzing sequences of user behavior, the goal is to develop a proactive system for identifying high-value users and mitigating customer churn in real-time.

The core solution is a Hybrid Convolutional Neural Network (CNN) and Bidirectional Gated Recurrent Unit (BiGRU) model, selected for its superior ability to capture both short-term intent and long-term temporal dependencies in the event stream.

---

💾 Dataset

The project utilizes a publicly available, real-world dataset.

• Dataset Name: Online Cosmetics Dataset
• Source: Kaggle
• Source link: https://www.kaggle.com/datasets/mkechinov/ecommerce-events-history-in-cosmetics-shop
• Time Period: October 2019 to February 2020 (5 months of data)
• Scale: Over 20 million event records
• Key Features: event_time, event_type, product_id, brand, price

Data Preprocessing Summary

• Feature Exclusion: user_id, category_id, and user_session were excluded to focus on generalized event patterns.
• Missing Data Handling: The category_code column was removed due to over 80% missing values (NaN).
• Temporal Engineering: event_time was decomposed into granular features (year, month, day, hour) for deep time-series analysis.

---

🧠 Model Development

Architecture: Hybrid CNN + BiGRU

The final model is a custom hybrid architecture combining two powerful sequential learning components:

• CNN Layer: Used first to extract local, high-level features and short-term patterns from the sequence of user actions (e.g., immediate viewing patterns).
• BiGRU Layer: Processes the features extracted by the CNN, reading the sequence in both forward and backward directions to capture the full temporal context and long-range dependencies of the user session.

This combination yielded the best performance in forecasting the user's next action compared to standalone LSTM, GRU, and CNN models.

---

Experiment Tracking

All model training runs, hyperparameters, and performance metrics were meticulously tracked using MLflow to ensure reproducibility and facilitate effective comparison between the benchmark architectures.

---

🚀 Deployment

The model is deployed via a user-friendly interface for immediate practical use.

• Deployment Tool: Gradio
• Functionality: Allows users to input a short sequence of recent events.
  The model processes it in real-time and outputs the predicted next event (e.g., purchase) along with a confidence score.

---

💰 Business Implications

The predictive capability of this model is directly linked to customer retention and revenue optimization.

• Proactive Churn Mitigation:
  The model serves as an early-warning system by flagging users whose predicted next action falls below the 'purchase' probability threshold, indicating a high risk of churn.

• Targeted Intervention:
  This real-time identification enables immediate personalized marketing actions (e.g., specialized discounts, personalized content recommendations based on favorite brands like irisk or runail) to re-engage the customer before they are lost.

• Increased CLV:
  By improving customer retention rates, the model provides a quantifiable return on investment by maximizing Customer Lifetime Value (CLV) and reducing costly customer acquisition spend.

---

🛠️ Future Work

To further enhance the model's performance and prepare it for enterprise-level scale, the following improvements are planned:

• Feature Enrichment:
  Integrate Embedding Layers for user_id and user_session to allow the model to learn and account for individual user and session biases.

• External Context Integration:
  Incorporate external data such as promotional schedules and seasonal economic indicators to improve predictive accuracy, particularly during volatile shopping periods.

• Scalable Deployment:
  Migrate the Gradio deployment to a robust microservice architecture (e.g., using Flask/FastAPI with Docker) to support high-volume, asynchronous, and concurrent real-time inference requests.