WWW’24 | Enhance click through rate estimation with similar users and items

Follow us and learn together

Title: Recall Augmented Ranking: Enhancing Click Through Rate Prediction Accuracy with Cross Stage Data

Address: https://arxiv.org/pdf/2404.09578

School, Company: Jiaotong University, Huawei

Conference: WWW 2024

1. Introduction

alt text

2.1 Cross stage user/item selection module

The function of this module is to select the most similar users and related items. Taking the selection of items as an example:

• Calculate the similarity between the target item and each recalled item using the similarity function f()
• Select the top k relevant recall items based on similarity scores

A simple and direct way is to calculate the inner product between EMBs and then select the top k ones. But this will involve a large number of multiplication operations, which are too computationally intensive. In the experiment, the author used SimHash function (a locally sensitive hash algorithm, friends who are not familiar with it can search for it).

• Simply explain: Here, the emb is divided into N groups, and a set of random…