💡 Two-Tower Deep Learning Model for Personalized Banking Recommendations

🚀 Deep Learning-Based Two-Tower Architecture for Personalized Financial Services

(Inspired by Meta’s Deep Learning Recommendation Model – DLRM)

---

📌 Project Overview

This project implements a deep learning-based Two-Tower architecture, inspired by large-scale industry models such as Meta’s DLRM (Deep Learning Recommendation Model), to deliver personalized banking product recommendations.

Each tower encodes different data modalities customer features and product attributes into a shared embedding space. This allows the system to efficiently match customers with suitable financial products based on learned latent representations.

By leveraging modern deep learning techniques, the system captures complex behavioral and financial patterns, enhancing recommendation accuracy and boosting customer engagement.

---

📂 Dataset

• Customer Features: Demographics, financial metrics, behavioral scores, credit info
• Product Features: Product type, risk level, yield, duration, terms
• Total Samples: 10,000+ customer-product data points

---

🏗️ Model Architecture

The model adopts a Two-Tower Deep Neural Network, inspired by large-scale recommender systems like Meta’s DLRM and Google’s YouTube Retrieval system.

🏛️ Customer Tower

• Inputs: Age, income, credit utilization, loan amount, digital engagement, etc.
• Architecture: Dense layers + Batch Normalization + Dropout + ReLU activations
• Output: Fixed-length customer embedding vector representing latent user preferences

🏦 Product Tower

• Inputs: Product category, risk level, expected yield, duration, financial requirements
• Architecture: Symmetrical to customer tower for parallel learning
• Output: Product embedding vector capturing financial product characteristics

🔗 Matching Layer

• Computes cosine similarity between customer and product embeddings
• High similarity → strong customer-product match
• Can be extended with cross features or interaction layers for complex relationships

---

🧠 Methodology

📊 1. Feature Engineering

• Customer Features: Standardized continuous features (e.g., income), encoded categorical features (e.g., gender, engagement tier)
• Product Features: Embedded categorical fields like risk category, duration type
• Normalization ensures stable and efficient learning

⚙️ 2. Model Training

• Implemented using TensorFlow Functional API for flexibility
• Loss Function: Contrastive loss or Triplet loss to maximize relevant pair similarity
• Negative Sampling used to improve discriminative learning

🎯 3. Recommendation Logic

• For a given customer, similarity scores are computed for all available products
• Top-K ranking used to retrieve the most relevant recommendations
• Optional filters: risk-tolerant vs conservative, interest-specific, etc.

📈 4. Evaluation

Evaluation metrics include:

• ✅ Mean Reciprocal Rank (MRR)
• ✅ Precision@K / Recall@K
• ✅ Normalized Discounted Cumulative Gain (nDCG)

---

🌐 Streamlit Interface

To make the model interactive and usable by business teams, the project includes a Streamlit dashboard.

🔧 Key Features:

• Upload or choose a customer profile
• Visualize customer attributes and financial behavior
• Get Top-5 personalized banking product recommendations
• Interactive insights:
  • Similarity scores
  • Matching product details
  • Customer-product embedding proximity
• SHAP visualizations for transparency and explainability

▶️ To Run the App:

streamlit run Recommendation_Dashboard.py