BJTU Web Search & Recommendation 2023 (M502075B/C402005B)

Demo code for news click-through rate prediction competition.

Base Model

1. Framework

The framework of base model for news recommendation, it consists of several parts: (1) a news encoder that uses rich textual content to learn news embeddings, and all news encoder parameters are shared; (2) a cheap sequential behavior moduel learns user embeddings from news click history; (3) a prediction moduel calculates predictive scores using the dot product of user embeddings with candidate news embeddings.

2. News Encoder

Note: only the title information is used in this demo.

Requirements

torch==1.12.1
tqdm==4.65.0
pandas==2.0.3
numpy==1.23.5
scikit-learn==1.3.0
nltk==3.8.1

The code works fine on Python 3.8

Usage

We recommend developing this repository on Linux.

1. Clone this repository

   git clone https://github.com/nipirennipi/BJTU-M502075B-2023
   cd BJTU-M502075B-2023

2. Download dataset
   • Link1: https://www.aliyundrive.com/s/y9jegdc2nVG
   • Link2: https://www.aliyundrive.com/s/4ixMCMzc5k6
   • Move them to the data folder

     └── data
         ├── train_behaviors.txt
         ├── test_behaviors_1.txt
         ├── test_behaviors_2.txt
         ├── news.txt
         └── ...
     

3. Download pre-trained word vectors

   cd data
   bash download.sh

4. Training and prediction

   cd ..
   bash run.sh

Hints

• More Powerful News Encoder
• Sequential Behavior Modeling
• Training Strategy
• Data Imbalance

Reading List

Some papers that might be helpful.

• Survey

  • Wu et al., Personalized News Recommendation: Methods and Challenges, TOIS 2023.
  • Zhang et al., Deep Learning for Click-Through Rate Estimation, IJCAI 2021.

• Sequential Behavior Modeling

  • Zhou et al., Deep Interest Network for Click-Through Rate Prediction, KDD 2018.

• Prompt Learning

  • Zhang et al., Prompt Learning for News Recommendation, SIGIR 2023.
  • Li et al., PBNR: Prompt-based News Recommender System, arXiv 2023.