Internship Experiences

Tier Criteria Management System

• Designed and built a full-stack rule management system from scratch, streamlining management of 40+ evolving rules
• Introduced version control and dry-run evaluation using 700+ models, enhancing update auditability and reliability
• Backfilled error data and launched the system to production, enabling real-time downstream consumption of updated rules
• Authored comprehensive design documentation, developer guides and user manuals to ensure long-term maintainability
• Implemented multiple compliance-related tickets to collect AI model information required under current regulations

Click Farming Detection based on the Similarity of Comment Pictures

• Mastered the implementation and applications of various existing image similarity algorithms, evaluated their performance on real comment image datasets, and studied the feasibility of identifying suspected click farming according to image similarity
• Put forward a special image similarity algorithm based on measuring the abnormal local identity of images with statistic methods and the feature matching algorithm, which yielded an average accuracy of over 95% on experimental synthetic datasets and over 98% on real datasets
• Independently developed a click farming detection method with the proposed image similarity algorithm and the comment image dataset, which could detect almost 100% of the tested anomalous image groups
• Completed the optimization and environment configuration of the algorithm and deployed it as a scheduled task in the production environment successfully, processing up to 300,000 images per day

Click Farming Detection based on Comment Content and Language Model

• Learned key technical means of click farming detection, and acquired necessary knowledge concerning the field of big data including Hadoop, Hive, Spark and Oozie by participating in practical projects
• Mastered the principle and application of click farming detection models based on comment content and BERT language model and on customer and seller behavior on the big data platform provided by JD
• Fitted the existing pretrained click farming detection model to a comment dataset of another business scope, mainly by modifying the data processing procedure and tuning configuration parameters