An F-shape Click Model for Information Retrieval on Multi-block Pages

Introduction

This is the tensorflow implementation of FSCM proposed in the paper: An F-shape Click Model for Information Retrieval on Multi-block Pages. WSDM 2023.

Requirements

• python 3.6
• tensorflow 1.14
• numpy
• matplotlib
• seaborn
• pandas Note that seaborn and matplotlib only be used in eye-tracking codes.

Input Data Formats

We collect AppStore dataset from a mainstream commercial App Store with F-shape pages, from September 17, 2021 to November 14, 2021. The first $54$ days are training set, while the last $5$ days are randomly split into valid set and test set. The user behavioral history is collected in real-time. We discard queries that have no positive interactions (clicks) with the retrieval system. The dataset consists of $394046$ unique queries and $1646$ items (Apps). More details of the dataset could be found in the supplementary material. Each item contains $7$ fields of features and each query contains $26$ fields of features. Each app belongs to one of 23 different categories, so that the item converge $\tau_v^j$ is a one-hot vector for the App Store dataset. The dataset statistics are shown as follows:

	training	validating	testing
#sessions	637959	40792	40792
avg. block per session	3.6654	3.7392	3.7374
avg. click per session	0.6510	0.7248	0.7313

The form of train/valid/test input file is .pkl, and the format is as follows:

• Vertical: [length(1), session_id(1), request_id(1), row_pos(4), item(PER_VER_LENGTH*(FEAT_NUM+2))]

• Horizontal:[row_id, item(PER_HOR_LENGTH*(FEAT_NUM+2))]

• Item:[click, feat(FEAT_NUM)]

Quick Start

We provide quick start command in run.sh or run following code in command line.

python run.py --train --data_dir data

Citation

If you find the resources in this repo useful, please cite our work.