Summary of Data Label Research under Smart Finance

Chao Yu, Zhiqiang Gao, Xiaolong Cui, Shaozhe Guo2023 · DOI

Abstract

In the field of smart finance, understanding and managing users effectively relies heavily on the use of data labels. This paper provides a comprehensive survey of data labeling research, expounding on generation methods, summarizing foundational datasets, and reviewing open-source and commercial annotation tools. To address the critical security and privacy risks associated with crowdsourcing sensitive financial information, the authors propose a novel distributed data annotation system.

Introduction: The Role of Labeled Data in Finance

In the era of "AI + Finance," data is the fuel, but labeled data is the engine. Financial institutions require precise categorization of users, credit risk prediction, and automated customer service. However, they face a "triple threat" of challenges:

  1. Complexity: Financial data is multi-modal, encompassing text, voice, images (e.g., identity documents), and transaction logs, making manual labeling labor-intensive.
  2. Efficiency: As deep learning models evolve, the demand for high-quality labels often outpaces human capacity.
  3. Security: Traditional crowdsourcing risks leaking sensitive business secrets or personally identifiable information (PII), which is the most significant hurdle for financial institutions.

Taxonomy of Data Annotation

The authors conduct a broad survey of existing methodologies, classifying labeling by dimension (Image, Voice, Text) and composition (Structured vs. Unstructured).

Table 1: Classification of data annotation methods by dimension and concept.

As illustrated in Table 1, different data dimensions present unique challenges. For instance, voice annotation requires manual transcription into text before algorithmic identification, while image annotation has been pivotal in developing face recognition and autonomous driving technologies.

The taxonomy further explores the composition forms of labels and the types of annotators involved.

Table 2: Classification by composition form and annotator type.

Table 2 highlights the trade-offs between manual and machine annotation. While manual processes offer high quality, they are often costly and inefficient. Conversely, machine annotation is fast and cost-effective but currently struggles with high-level semantics.

Algorithmic Frameworks for Labeling

The paper reviews how labels are utilized in downstream tasks, focusing on three primary algorithmic approaches:

  • Label Propagation (LPA): Utilizing known labels to "spread" and categorize unlabeled data within a network.
  • Multi-label Classification: Handling instances where a single data point, such as a user profile, belongs to multiple categories simultaneously.
  • Label Clustering: Using techniques like K-Means to identify latent structures in data before formal labels are applied.

Proposed Solution: Distributed Data Annotation System

To address the security concerns of financial institutions, the authors propose a distributed data annotation system. This architecture segments sensitive data into "low-value" fragments, which are then distributed to different departments for isolated annotation. This ensures that no single annotator sees the full, sensitive context. Once labeled, the fragments are "stitched" back together in a secure data center, maintaining both data utility and privacy.

Key Insights and Quality Control

A significant finding is that labeling is often subjective. The paper notes that class-related label bias is primarily affected by deviations in subjective cognition. To mitigate this, the authors suggest using complex quality control algorithms like Expectation-Maximization (EM) to find the "ground truth" among conflicting labels. Furthermore, "label smoothing" is emphasized as a technique to improve model generalization on noisy financial data.

Conclusion and Future Directions

The paper identifies eight core scenarios for smart finance, including risk control, insurance, and identity verification. Despite progress, the "semantic gap" remains a challenge, as machine recognition of complex financial intent still lags behind human performance. Future work should focus on "time-index attenuation," ensuring that user portraits remain accurate as data ages and behavior patterns shift.

Summary updated

Abstract

How financial product managers better understand and manage users involves the use of data tags. From the perspective of label, this paper expounds the generation method of data labels, summarizes the classic common data set, open source data annotation tools and commercial crowdsourcing platform, e
BibTeX
@misc{yu2023summary,
  author = {Chao Yu and Zhiqiang Gao and Xiaolong Cui and Shaozhe Guo},
  title = {{Summary of Data Label Research under Smart Finance}},
  year = {2023},
  doi = {10.4108/eai.28-10-2022.2328436},
}
Figure