CTR Prediction with Intent-Aware Recommendation System

1. Project Overview

This project focuses on improving Click-Through Rate (CTR) prediction in sparse and privacy-constrained environments, where traditional user tracking methods such as cookies are limited.

Conventional recommendation systems rely heavily on user interaction history, which leads to performance degradation in cold-start and sparse data scenarios.

To address this problem, we propose an intent-aware recommendation framework that combines:

• Aggregated user behavior features
• Semantic user intent embeddings generated from behavior summaries
• DeepFM-based CTR prediction model

The goal is to enhance recommendation performance by incorporating user intent, especially when behavioral data is limited.

---

2. Pipeline

The overall pipeline consists of the following steps:

1. Behavior Aggregation
   User interaction logs are aggregated to create features such as:

   • total views, cart, purchases
   • buy rate, cart rate
   • category distribution and activity level

2. Ad Embedding
   Ad features are converted into text descriptions and encoded using a Sentence Transformer.
   The embeddings are then reduced using PCA.

3. Intent Feature Engineering
   User behavior is transformed into high-level funnel features and converted into natural language templates describing user intent.

4. Intent Embedding
   The generated intent text is encoded into dense vectors using a Sentence Transformer.

5. Data Preparation
   All features (user, ad, behavior) are merged and normalized.
   The dataset is split into training and test sets based on timestamp.

6. Model Training

   • Model A: DeepFM using structured features only
   • Model B: DeepFM with additional intent and ad embeddings

7. Evaluation
   Performance is evaluated using:

   • AUC
   • Log Loss
   • ECE (Expected Calibration Error)

##  Dataset Setup

This project uses the **Taobao Ad Click Dataset**.

Download the dataset from:
https://www.kaggle.com/datasets/pavansanagapati/ad-displayclick-data-on-taobaocom

After downloading, place the following files inside the `data/` directory:

- raw_sample.csv  
- user_profile.csv  
- ad_feature.csv  
- behavior_log.csv  

Example structure:

data/
├── raw_sample.csv  
├── user_profile.csv  
├── ad_feature.csv  
└── behavior_log.csv  

⚠️ Note:
- The dataset is not included in this repository due to size limitations.
- Make sure file names match exactly, or the pipeline will fail.

---

3. Output (Evaluation)

Model C (Upper Bound)

• BST (Behavior Sequence Transformer)
• Uses sequential user behavior data
• Implemented using FuxiCTR
• Provides upper-bound performance benchmark

Model A (Baseline: DeepFM)

[Sparse & Dense Features]
        │
        ▼
[Embedding Layer]
(Sparse → Dense Vector)
        │
        ├───────────────┐
        ▼               ▼
[FM Component]     [Deep Component]
(2nd-order         (MLP)
 interactions)     nonlinear learning
        │               │
        └──────┬────────┘
               ▼
        [Output Layer]
        CTR Prediction

• AUC: 0.6424
• LogLoss: 0.1943
• ECE: 0.0122

Model B (Enhanced)

### Model B (Enhanced: Intent-aware DeepFM)

[User Behavior Data]
        │
        ▼
[Feature Engineering]
(pv_to_cart, buy_rate, etc.)
        │
        ▼
[Text Template]
"User is a high conversion buyer..."
        │
        ▼
[Sentence Transformer]
        │
        ▼
[PCA]
(Intent Embedding)
        │
        ▼
────────────────────────────────
[Structured Features + Intent Embedding]
        │
        ▼
[Embedding Layer]
        │
        ├───────────────┐
        ▼               ▼
[FM Component]     [Deep Component]
        │               │
        └──────┬────────┘
               ▼
        [Output Layer]
        CTR Prediction

Model B (Enhanced)

• AUC: 0.6592
• LogLoss: 0.1925
• ECE: 0.0076

Key observations:

• Model B outperforms Model A in overall performance
• Improvement is especially significant for sparse users
• Intent embeddings help recover performance loss in limited data settings

---

4. Repository Structure

ctr_project/
- data/                  (dataset not included)
- output/                (generated results, ignored)
- scripts/
  - run_model_a.py       (baseline model execution)
  - run_model_b.py       (intent-enhanced model execution)
  - run_model_c.py       (placeholder for upper-bound model)
- src/
  - config.py            (configuration settings)
  - pipeline.py          (data processing pipeline)
  - dataset.py           (PyTorch dataset)
  - models.py            (DeepFM model)
  - train_eval.py        (training and evaluation logic)
- requirements.txt
- README.md

---

5. References

• Guo et al. (2017), DeepFM: A Factorization-Machine based Neural Network for CTR Prediction
• Chen et al. (2019), Behavior Sequence Transformer (BST)
• Reimers & Gurevych (2019), Sentence-BERT
• Taobao Ad Click Dataset (Alibaba)