On June 29, 19:00-20:00, Big Taobao Technology and DataFun jointly planned this event, and invited Mr. Wang Yuan, an algorithm expert of Alibaba Big Taobao Technology, to discuss the application of the rearrangement model in Taobao streaming recommendation scenarios. In-depth sharing and communication, welcome to watch the live broadcast on time~
Wang YuanAlibaba Big Taobao Technology Algorithm Expert
Personal introduction: Ph.D., New York University, USA. Joined Alibaba in 2018; used to be responsible for mobile Taobao private domain product search recommendation; currently responsible for mobile Taobao homepage "follow" information flow recommendation algorithm.
Speech title: Application of a rearrangement model that integrates complex targets and supports real-time regulation in Taobao streaming recommendation scenarios
Speech Outline:
In addition to traditional correlation, complex information flow recommendation scenarios also need to take into account various requirements, including dispersal (diversity), traffic control, multi-display form/multi-channel supply integration, etc. The traditional recommendation system adopts the form of pipeline to process the above requirements step by step, which has poor scalability and lack of overall planning and optimization, which limits the effect of scene recommendation. We propose a brand-new rearrangement model based on the Generator-Evaluator (GE) architecture. It can not only break through the traditional correlation-greedy sorting paradigm, generate sequences with the overall effect of the sequence as the goal, but also break through the pipeline recommendation paradigm. In a model Organically integrate complex business rules to give an end2end joint optimal solution. We verified the effectiveness of the proposal in the Taobao information flow scene and launched it in full.
Weights are usually used to express the relative importance between rearranging multiple objectives. Existing schemes usually rely on static modeling, that is, the preference weights are determined during offline training and cannot be changed afterwards. Adjusting the weights requires retraining the model, which is time-consuming and resource-intensive. But in practical applications, real-time weight adjustment can bring us many advantages, such as the response to traffic status during the big promotion. To this end, we propose a controllable weight rearrangement model based on a hypernetwork, which can support the real-time assignment of fusion weights for each test sample without training the model, and generate the optimal sequence that meets the weight settings.
Audience benefits:
1. Challenges of complex streaming scenarios and unique advantages of rearrangement models
2. Summary of rearrangement modeling paradigms
3. How to integrate complex objectives into a unified model and give e2e optimal solutions
4. How to achieve the relationship between objectives flexible adjustment
