Leading Design and Development of the Advertising Recommender System at Tencent: An Interview with Xiangting Kong
Introduction: Sharing Insights and Best Practices
Working closely with the broader recommender community, including innovative global leaders like Tencent, enables NVIDIA to incorporate best practices, insights, and learnings back into our recommender framework. At NVIDIA, we are committed to streamlining the building, deploying, and optimizing of recommender systems. Our engineering teams participate in industry challenges, host community events, and work closely with early adopters like Tencent. Xiangting Kong, Expert Engineer, leads the design and development of Tencent’s Advertising and Deep Learning Platform. While Kong presented at GTC Spring, we asked him to share some additional insights on building recommender systems in this interview.
Interview with Xiangting Kong, Expert Engineer, Tencent
Question: What is your role at Tencent?
Xiangting Kong: I am an expert engineer at Tencent and am responsible for the design and development of the advertising recommendation system. I am also the lead of the Tencent Advertising and Deep Learning Platform. Our platform supports the machine learning model optimization, training, and inference in a variety of business scenarios from advertising and fintech to networking data mining.
Questions: What does your team at Tencent work on?
Xiangting Kong: Our team mainly develops machine learning platforms and we are responsible for feature engineering, model training and online inference. We are working on implementing a new generation of high-performance distributed training system for advertising recommendation based on GPU from 0 to 1.
Question: How does your work and your team’s work on recommenders relate to Tencent’s overall business?
Xiangting Kong: Our advertising recommendation training platform covers the entire Tencent traffic business. Tencent advertising recommendations are widely used in services such as WeChat, Moments, QQ, Tencent Games, Tencent Video, Tencent News and so on. Tencent advertising revenue is in the hundreds of millions. The accuracy of our advertising recommendation helps increase advertising revenue.
Question: Is your team a relatively new team? Why did Tencent decide to invest in recommenders?
Xiangting Kong: Our team has been established for years. The advertising business is a relatively important business inside Tencent and the recommendation system is used to increase the overall advertising revenue.
Question: What kind of recommender systems does your team focus on?
Xiangting Kong: The main focus of our team is the advertising recommendation system, responsible for the optimization of the advertising training platform. Tencent advertising recommendations system consists of parts including offline feature engineering, training platform, online inference system, online feature engineering and play platform. Advertising recommendations is a process of gradual filtering. Sorting stages include recall, pre-ranking, and ranking. Each stage has different requirements. The rapid investigation and iteration of the model puts forth higher requirements for training performance.
Question: How does your team conduct training?
Xiangting Kong: We organize some technology sharing every one or two weeks.
Question: How does your team evaluate your recommender systems? fine tune?
Xiangting Kong: Through our recommendation system, we optimize the algorithm strategy, add more samples and features, and assess whether it can drive the increase of income. The accuracy of the advertisement recommendation can be improved by training more sample data, by adding more sample features. But this leads to longer training time and effects the update frequency of the model. In order to ensure that the model updates will not be derailed, the training performance of the model needs to be continuously improved. After the training model performance is improved, more data can be trained to improve the accuracy of the model, thereby increasing the advertising revenue.
Question: How do you optimize your recommender systems? For example, it is our understanding that Tencent uses HugeCTR for embeddings optimization. How has this helped you optimize your workflow?
Xiangting Kong: HugeCTR, as a recommendation training framework, is integrated into the advertising recommendation training system to make the update frequency of model training faster, and more samples can be trained to improve online effects.
Question: How do you choose the appropriate technique, package, method, or frameworks to support your work?
Xiangting Kong: The technology or framework we choose must be compatible with the community ecosystem, so that we can do better follow-up upgrades.
Question: How do you address scaling your models?
Xiangting Kong: Using a larger model is conducive to learning more features, thereby improving the accuracy of the model.
Question: What is a recent success for the team?
Xiangting Kong: In our training framework, a data-parallel distributed solution has been developed.
Question: Have you recently integrated specific methods into your recommender workflow?
Xiangting Kong: We recently integrated the CSR [Compressed Sparse Row] pipeline into our ad recommendation training system. CSR type training data is generated so that the data can be directly read on the GPU for training. Through our optimization of the data processing pipeline, the CPU load is greatly reduced and the GPU utilization is greatly improved.
Question: If a team lead was just starting out and currently evaluating building, deploying, and optimizing recommenders for their company….what advice would you relay to help them accelerate or streamline their recommender workflows?
Xiangting Kong: Choose a mature technical framework and be compatible with the community ecosystem to facilitate subsequent system upgrades.
Additional Community Resources to Consider
As NVIDIA is committed to streamlining recommender workflows, we incorporate best practices, insights, and learnings from the broader industry, including innovative early adopters like Tencent. Additional resources and events to consider if you are looking for additional best practices to help accelerate recommender workflows: