A Pytorch Implementation of Neural Click Model (NCM)

Introduction

This is an non-official implementation of Neural Click Model proposed in the paper: A Neural Click Model for Web Search. WWW 2016.

NOTE: There are several modification in my implementation compared to the original paper:

1. The original paper use one-hot encoding for query/doc embeddings. This makes the embedding dimensionality up to 10240, which is too large for training on GPU efficiently. So here I use nn.embedding provided by PyTorch to constrain the dimensionality.
2. Vertical type is added as side information. Vertical type means the representation style of a displayed document (e.g., organic vertical, the illustrated vertical, the encyclopediavertical. To the best of my knowledge, up to now, only TianGong-ST dataset provides this side information.
3. I input the query information at each RNN time step, instead of only using query for RNN state initialization.

Requirements

• python 3.7
• pytorch 1.6.0+cu101
• torchvision 0.7.0+cu101
• tensorboard 2.1

Input Formats & Data Preparation

After data pre-processing, we can put all the generated files into ./data/dataset/ folder as input files for NCM. Demo input files are available under the ./data/demo/ directory.

The format of train & dev & test & label input files is as follows:

• Each line: <session id><tab><query id><tab>[<document ids>]<tab>[<vtype ids>]<tab>[<clicks infos>]<tab>[<relevance>]

Recommended Hyperparameter

I provided some recommended hyperparameters below. The example training shell can be found at ./example_run.sh

• optim: adam
• eval_freq: 100
• check_point: 100
• learning_rate: 0.001
• lr_decay: 0.5
• weight_decay: 1e-5
• dropout_rate: 0.5
• num_steps: 40000
• embed_size: 64
• hidden_size: 64
• batch_size: 128
• patience: 5

Important Tips

For brevity and legibility of the code, I remove many non-core code:

1. Data-preprocessing for TianGong-ST dataset
2. Sequence-independent ranking for relevance estimation
3. Synthetic dataset generation

However, you can checkout these codes at the commit, (but the code might lack comments):

commit 5b114cbdab0da4de1c43ebc3d666a7cac1818a23
Author: LinJianghao <[email protected]>
Date:   Thu Nov 19 10:31:32 2020 +0800