Alibaba Innovative Research (AIR) > Research on Supply and Demand Matching Optimization for Large-scale E-commerce Platforms
Research on Interpretable Modeling of E-commerce Entities with Few/Zero Annotations

Research Themes

Research on Supply and Demand Matching Optimization for Large-scale E-commerce Platforms

Background

Supply and consumption are two sides of the retail trade. It is of great value for e-commerce businesses to better understand the characteristics of corresponding entities as well as their underlying relations. Generally, e-commerce entities include commodities, videos, articles, queries, users, etc. Characterizing these entities with informative embeddings is a typical scheme, which provides an effective way for the supply and demand matching in e-commerce scenarios. Such a solution has a wide range of practical value in real-world applications like personalized recommendation, advertising, and promotion campaign.

Recently, deep learning has made tremendous progress in modeling e-commerce entities, especially the supervised/self-supervised methods training with large-scale data. However, this issue is far from being solved. At this stage, we are facing two main challenges, i.e., the ability of learning from few annotations and the interpretability of the model and result.

1. The traffic dividend of e-commerce business has probably peaked. We believe that the increasing cost of customer acquisition would change and reshape the business. Better mining customer demands and customer trends is one of the key problems in the future. In our scenario, there are increasing label prediction tasks for e-commerce entities based on actual business demands. For example, vertical industries would continue to define new tags for videos on Taobao, which reflect some kinds of shopping interests and can be used for further commercial operations. The challenge is that the annotations are often very scarce in the early stage, but the overall labeling process is usually required to be completed within a week. Therefore, it is more and more important to explore a few/zero-shot framework or pipeline for efficient labeling of large-scale e-commerce entities.

2. The government continues to strengthen the regulation of the processing and application of data. Accordingly, nowadays it is critical to promote the capacity of privacy protection, data security and algorithms transparency. From this point of view, previous complex deep learning based solutions (that rely on large-scale training and work in a “black box” manner) would encounter bottlenecks. Modeling e-commerce entities should be adapting to current needs. Therefore, we argue that the interpretability of a model is a must in the future, regardless of the specific modeling methods (representation learning or statistical ones).

Target

  • A general-purpose framework and pipeline for interpretable label/link prediction of e-commerce entities, being capable of performing few/zero shot learning
  • Choose one or more type of e-commerce entities and apply the framework in a specific label/link prediction task(such as product attributes prediction, multi-modal content classification, interest recongnition) to evaluate its effectiveness.
  • Candidate E-commerce entities: commodity (image + title), article, video, user

Related Research Topics

  • Few-shot/Zero-shot learning
  • Weakly supervised learning
  • Interpretability Representation/Machine learning
  • Data augmentation and contrastive learning

Scan QR code
关注Ali TechnologyWechat Account