Looking back at the development history of recommendation systems, starting from the collaborative filtering algorithm more than 30 years ago , experiencing the wave of deep learning , to today's hot large models , recommendation systems have radiated new vitality time and time again. With the arrival of large models, the recommendation system is on the eve of change. What breakthroughs will the original system modules, such as core architecture, training inference, cold start, etc., usher in? What new explorations will model innovation and intersection with large models bring?
From August 26 to 27, "DataFunSummit 2023: Recommendation System Online Summit" arrived as scheduled. This summit covers 8 forums including core architecture, model innovation, engineering architecture and training inference, best practices, large models and recommendations, graphs and recommendations, recommended cold start, and multi-scenario multi-tasking.
Xiaohongshu Vice President of Technology Fengdi and Head of Search Engineering Miao Tai serve as forum producers of "Recommendation Cold Start" and "Engineering Architecture and Training Inference" , and join hands with many senior technical experts in the industry to jointly present the best technical practices of recommendation systems. .
At the same time, Ban Chao , the person in charge of the recommendation engineering architecture , Jin Lun , the person in charge of the machine learning engine architecture, and Li Mu , the person in charge of the graphic information flow algorithm, will also be guest speakers, respectively bringing "Xiaohongshu's graphical business architecture practice in complex search and recommendation scenarios" ", "Integrated Architecture of Model Training and Promotion in the Search and Advertising Scenario", and "Xiaohongshu Recommended Users and Content Cold Start Practice" are wonderful sharings.
From 09:00 to 12:35 on August 26th and 27th (Saturday and Sunday), the entire summit will be broadcast live online. We look forward to your attention!
Share one:
"Xiaohongshu's graphical business architecture practice in complex search and recommendation scenarios"
08.26 10:15-10:50 "Core Architecture" Forum
The person in charge of Xiaohongshu’s recommendation engineering architecture, @banchao, will start with Xiaohongshu’s recommendation engineering architecture, introduce the business characteristics, business pain points, overall architecture, recommendation process, etc., and elaborate on how Xiaohongshu integrates the graph business architecture into search and push. implementation in the scene, and how to improve algorithm iteration efficiency through hot deployment mechanism.
Share 2:
"Integrated Architecture of Model Training and Promotion in the Search and Promotion Scenario"
08.26 11:25-12:00 "Core Architecture" Forum
Xiaohongshu’s machine learning engine architecture manager @Jinglun will lead everyone into Xiaohongshu’s overall machine learning engine architecture, large-scale distributed model training framework, model prediction platform, model optimization tool chain, inference engine, and data features. platform.
Share three:
"Xiaohongshu Recommended Users and Content Cold Start Practice"
08.27 09:10-09:50 "Recommended Cold Start" Forum
Xiaohongshu’s head of image and text information flow algorithm @李木 will share in detail Xiaohongshu’s practice in recommending users and cold-starting content, focusing on how to solve the problem of crowd breaking and cold-starting content when low-active user behavior is sparse. Break the circle.
"Engineering Architecture and Training Reasoning" Forum
Producer: Miao Tai, the person in charge of the search and promotion project of Xiaohongshu
08.26 09:00-12:35
Personalized technology has become an industry standard in fields such as promotional search. Faced with the rapid iteration of data magnitude, model complexity, and heterogeneous hardware, the development of AI training and inference engineering frameworks has also shown a multi-dimensional and refined trend; In particular, the optimization of computing efficiency, high-timeliness training, and cost-effective deployment under heterogeneous hardware have become hot topics that major companies continue to explore.
This forum specially invited technical experts from the fields of hardware computing, e-commerce, content, advertising, etc. to describe in detail the acceleration framework and scenario implementation practices under GPU, real-time training solutions, integrated training and promotion optimization, etc. Provide the audience with a package selection experience of hardware-architecture-optimization.
"Recommended cold start" forum
Producer: Xiaohongshu Technical Vice President Bagpipe
08.27 09:00-12:00
Cold start is one of the important challenges of recommender systems. Faced with new content, new users, and new systems, it’s easy to get stuck with limited information. The cold start problem is thorny and serious. It runs through the entire life cycle of the product and affects user experience and retention. In different business scenarios, cold start will face various challenges and difficulties.
How to design an effective cold start plan? How to measure the effectiveness of a strategy? How to integrate with multimodality? This forum invited many experienced technical experts in the industry to share the best practices for cold start problems based on real recommendation scenarios, including but not limited to user cold start, content cold start and other directions.
Xiaohongshu Technology Sharing looks forward to your attention. Click the link below to register for the conference and experience with us the cutting-edge exploration and latest implementation practices in the field of recommendation system technology.
Click to register to participate in the "Recommendation System Online Summit"