https://github.com/aister2020/KDDCUP_2020_Debiasing_1st_Place

https://github.com/aister2020/KDDCUP_2020_Debiasing_1st_Place

Last synced: 3 months ago
JSON representation

• Host: GitHub
• URL: https://github.com/aister2020/KDDCUP_2020_Debiasing_1st_Place
• Owner: aister2020
• License: apache-2.0
• Created: 2020-07-15T06:29:28.000Z (over 4 years ago)
• Default Branch: master
• Last Pushed: 2020-07-20T12:48:38.000Z (over 4 years ago)
• Last Synced: 2024-06-24T05:53:59.709Z (5 months ago)
• Language: Python
• Size: 60.5 KB
• Stars: 171
• Watchers: 5
• Forks: 56
• Open Issues: 3
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

• StarryDivineSky - aister2020/KDDCUP_2020_Debiasing_1st_Place

README

        
# KDD CUP 2020: Debiasing

### Team: aister

***

+ Members: Jianqiang Huang, Ke Hu, Mingjian Chen, Bohang Zheng, Xingyuan Tang, Tan Qu, Yi Qi, Jun Lei

+ Team Introduction: Most of our members come from the Search Ads Algorithm Team of the Meituan Dianping Advertising Platform Department. We participated in three of the five competitions held by KDD CUP 2020 and achieved promising results. We won first place in Debiasing(1/1895), first place in AutoGraph(1/149), and third place in Multimodalities Recall(3/1433).

+ Based on the business scenario of Meituan and Dianping App, the Search Ads Algorithm Team of Meituan Dianping has rich expertise in innovation and algorithm optimization in the field of cutting-edge technology, including but not limited to, conducting algorithm research and application in the fileds of Debiasing, Graph Learning and Multimodalities.

+ If you are interested in our team or would like to communicate with our team(b.t.w., we are hiring), you can email to [email protected].

### Introduction

***

+ This track focuses on the fariness of exposure, i.e., how to recommend items that are rarely exposed in the past, to combat the Matthew effect frequently encountered in a recommender system. In particular, performing bias reduction when training on the click data is crucial for the success of this task. Just like there is a gap between the logged click data and the actual online environment in a modern recommender system, there will be a gap between the training data and the test data, mostly with respect to the trends and the items' popularity. Please refer to the competition official website for more details: https://tianchi.aliyun.com/competition/entrance/231785/information