Credit Card Based Product Recommendation System

An end-to-end Credit Card Intelligence Pipeline covering:

1. Fraud Detection (Transaction-Level Risk Modeling)
2. Customer Segmentation (User-Level Behavioral Clustering)
3. Credit Card Recommendation System (Hybrid Neural Ranking Model)

This project simulates a production-grade banking system where risk, behavior, and product alignment are integrated into a single scalable framework.

---

1. Fraud Detection Module

Overview

The Fraud Detection module operates at the transaction level and identifies anomalous credit card activity under extreme class imbalance and real-world operational constraints.

Rather than relying on raw transaction attributes, the model captures behavioral deviations, which are central to modern fraud detection systems.

---

Feature Engineering Strategy

Fraud detection is driven by historical behavioral features, including:

• Transaction velocity (transactions in last 1h / 24h)
• Temporal gaps between transactions
• Deviation from user spending baseline
• First-time merchant and MCC indicators
• Time-of-day and channel usage patterns

These features model behavioral change instead of static transaction characteristics.

---

Model: LightGBM Classifier

A LightGBM classifier serves as the primary fraud engine due to:

• Strong performance on tabular financial data
• Ability to model non-linear interactions
• Scalability to millions of transactions
• Interpretability for compliance and audit use cases

The model outputs a continuous fraud risk score, which is mapped into business risk bands:

• Low
• Medium
• High
• Critical

Each band corresponds to operational actions:

• Approve
• Monitor
• OTP Verification
• Decline

---

Graph-Based Exploration

An experimental Graph Neural Network (GNN) was implemented on high-risk subsets to explore relational fraud propagation patterns.

This acts as an enrichment layer rather than the core production model.

---

Fraud Output Artifact

Final transaction-level output includes:

• fraud_score
• fraud_band
• fraud_action

This artifact feeds downstream segmentation and recommendation modules, ensuring risk-aware growth decisions.

---

2. Customer Segmentation (User-Level Behavioral Modeling)

Objective

The goal of segmentation is to group customers based on:

• Behavioral spending patterns
• Credit capacity
• Aggregated risk signals

Segmentation is performed at the user level, since business decisions are customer-centric.

---

Why User-Level Segmentation?

• Transaction-level clustering is unstable and noisy
• Card-level clustering fragments multi-card customers

Instead, transaction data, card data, and fraud signals are aggregated into a single user-level feature table (1 row per user).

---

Feature Engineering

Behavioral Spending Features

Transactions are mapped via MCC codes into nine business categories:

• Groceries
• Dining & Food
• Transport & Fuel
• Retail & Apparel
• Healthcare & Wellness
• Utilities & Bills
• Digital Subscriptions
• Entertainment & Leisure
• Travel & Lodging

For each user, the following metrics are computed:

• txn_count
• total_spend
• avg_txn_amount
• spend_cv (volatility)
• category_entropy (diversification)
• share_<category> (category spend share)

These features represent lifestyle intensity and preference structure.

---

Credit Capacity Features

From card portfolio data:

• num_cards
• total_credit_limit
• avg_credit_limit

These describe exposure and financial maturity.

---

Aggregated Risk Signals

Fraud signals are safely aggregated per user:

• avg_fraud_score
• % high-risk
• % critical-risk

Risk is treated as a behavioral trait, not a filtering condition.

---

Why FICO Was Excluded from Clustering

FICO score was intentionally excluded because:

• It is a policy variable, not behavioral
• It would dominate cluster formation
• It is reserved for eligibility gating in recommendations

---

Clustering Model

Rejected Approach: K-Means

K-Means was discarded due to:

• Instability across runs
• Sensitivity to heavy-tailed financial features
• Hard cluster boundaries
• Poor handling of compositional spending shares

---

Final Model: Gaussian Mixture Model (GMM)

Gaussian Mixture Models were selected because they:

• Provide soft (probabilistic) cluster membership
• Handle non-spherical cluster shapes
• Model covariance between features
• Represent overlapping lifestyle patterns

---

Cluster Selection

The number of clusters was selected using Bayesian Information Criterion (BIC) to balance fit and complexity.

Final choice:

k = 7 clusters

---

Resulting Personas

Seven interpretable behavioral segments were identified, including:

• Retail-Centric Lifestyle Spenders
• Dining & Convenience Seekers
• Fuel & Commute Heavy Users
• Digital & Subscription Natives
• Travel & Experience Spenders
• Budget-Conscious Essentials Users
• High-Capacity Risk-Edge Users

Each persona reflects a combination of:

• Spending intensity
• Lifestyle preference
• Credit capacity
• Risk exposure

---

Segmentation Output Artifact

user_level_segmentation_features.csv

Characteristics:

• One row per user
• All engineered behavioral features
• Cluster and persona labels
• FICO included for downstream eligibility logic

This file serves as the single source of truth for recommendation modeling.

---

3. Credit Card Recommendation System

Overview

The Recommendation System is the final stage of the pipeline.

It recommends credit cards based on:

• Behavioral spending patterns
• Credit capacity and FICO score
• Aggregated fraud risk
• Cluster persona alignment
• Card reward structure and eligibility

The system uses a Hybrid Neural Collaborative Filtering (Hybrid NCF) architecture.

---

Key Challenges

• No historical user–card interaction labels
• High-dimensional behavioral features
• Strict eligibility constraints
• Need for interpretability
• Cold-start stability

---

Data Inputs

1. User-Level Features

user_recommendation.csv

Contains:

• Spending aggregates
• Category shares
• Spending entropy
• Credit capacity features
• Fraud aggregates
• FICO score
• Cluster ID

---

2. Card Master Table

card_master_table.csv

Includes:

• Card ID
• Brand and tier
• Annual fee and APR
• Reward structure
• FICO eligibility range
• Cluster alignment metadata

---

3. Pairwise Training Data

pairwise_training_data.csv

Contains:

• user_id
• pos_card_id
• neg_card_id

Generated using compatibility scoring.

---

Weak Supervision Framework

Since no historical labels exist:

1. Compatibility scores are generated using:

   • Reward alignment
   • Category match
   • Risk penalty
   • Cluster alignment
   • FICO eligibility

2. Scores are converted into pairwise ranking tuples: (user, preferred_card, less_preferred_card)

This enables learning relative preferences.

---

Model Architecture: Hybrid Neural Collaborative Filtering

The architecture combines:

Latent Embeddings

• User embedding
• Card embedding

Explicit Feature Encoders

• MLP for user features
• MLP for card features

Interaction Network

Concatenated representations are passed through an MLP to produce:

score(user, card)

---

Training Strategy

The model is trained using:

• Pairwise ranking objective
• Bayesian Personalized Ranking (BPR) loss

Loss:

L = -log(sigmoid(score_pos - score_neg))

Users are split at the user level to prevent data leakage.

---

Evaluation Metrics

• NDCG@K
• Ranking consistency across validation users

Loss alone is not used to judge model quality.

---

Inference Pipeline

For each user:

1. Filter eligible cards using FICO and policy rules
2. Score all eligible cards
3. Apply risk-aware penalty
4. Optionally enforce diversity constraint
5. Return Top-10 ranked cards

Final output includes:

• user_id
• cluster
• persona
• rank
• card_id
• card_name
• brand
• card_tier
• annual_fee
• apr
• final_score

---

Explainability

Recommendations are explainable through:

• Persona alignment
• Reward-category matching
• Risk transparency
• Cluster-based reasoning

---

Final Summary

The Credit Card Based Product Recommendation System:

• Detects fraud using behavioral modeling
• Segments customers using probabilistic clustering
• Recommends products using Hybrid Neural Collaborative Filtering
• Enforces eligibility and risk constraints
• Maintains interpretability and production realism

This project simulates a real-world banking intelligence pipeline from raw transaction data to deployable recommendation output.