HIVE: Hierarchical Information Visualization for Explainability

Yi-Ning Juan, Yi-Shyuan Chiang, Shang-Chuan Liu, Ming-Feng Tsai, Chuan-Ju WangIJCAI 2021 · DOI

TL;DR

This paper introduces HIVE, an interactive web tool that visualizes hierarchical attention networks (HANs) for financial statement analysis. HIVE aims to make complex financial documents more understandable and comparable by highlighting key information and recommending similar reports, bridging the gap between deep learning models and financial professionals.

What Problem Does This Paper Address?

Financial professionals face significant challenges when analyzing company financial statements. These documents are often lengthy and dense, making it difficult to quickly grasp crucial information. Comparing multiple statements across companies is a tedious process, and linking textual information within reports to actual financial risks (like stock volatility) is not straightforward. Existing deep learning models, while powerful, often lack interpretability, making it hard for users to understand why a model makes a certain prediction. This paper addresses the need for a tool that not only processes financial statements using advanced AI but also provides clear, actionable insights to aid decision-making.

Methods

The core of HIVE is a HAN-inspired ranking model trained on 10-K financial reports. This model utilizes a hierarchical structure, processing information at both the word and sentence levels. Crucially, it employs a multilevel attention mechanism to identify and weigh the importance of different text segments in predicting financial risk. This attention mechanism provides explainability by highlighting influential words and sentences.

For risk assessment, HIVE uses a pairwise ranking loss based on a siamese network with two parallel HAN structures. This allows it to rank financial reports by their financial risk levels, using post-event return volatility as a proxy for financial risk.

To enhance usability, HIVE also incorporates a recommendation algorithm. This algorithm extracts document embeddings from the HAN model and uses a k-d tree structure to efficiently find similar financial reports based on cosine similarity, enabling low-latency recommendations.

Datasets

The primary dataset used is a large collection of 10-K financial reports from 1996 to 2013, obtained from the SEC. These reports are used to train the HAN-based ranking model and extract document embeddings for the recommendation system.

Key Insights and Ideas

  • Bridging the AI-Finance Divide: The paper highlights the significant challenge of making advanced NLP models (like HANs) accessible and useful to financial professionals who may not be AI experts. HIVE's interactive visualization is key to this.
  • Hierarchical Attention for Financial Nuance: Financial documents have inherent hierarchical structures (words form sentences, sentences form paragraphs, etc.). Using hierarchical attention networks is a natural fit for capturing these complex relationships and identifying key drivers of financial risk.
  • Explainability as a Core Feature: Simply predicting risk is not enough. HIVE's focus on visualizing attention weights allows users to understand why a report is flagged as high-risk, building trust and enabling deeper analysis.
  • Integrated Workflow: HIVE combines risk ranking, explainability visualization, and report recommendation into a single, streamlined interface, addressing multiple pain points in the financial analysis workflow.

Key Results

The paper demonstrates HIVE's functionality through case studies. For instance, it shows how the customizable highlighter can effectively pinpoint sentences with high attention weights, such as those detailing revenue increases and new product lines for TESLA MOTORS INC. The statement recommendation system is shown to successfully suggest relevant competitor reports, like recommending LENNOX INTERNATIONAL Inc. for AAON Inc., based on shared business strategies and similar financial performance impacts (e.g., losses due to economic environment). The system is designed for real-time performance, with recommendations generated efficiently.

Limitations and Future Work

The paper acknowledges that while the attention mechanism provides some level of explainability, it is not a complete solution for understanding complex model decisions. The authors suggest that HIVE is a prototype that successfully addresses many workflow challenges for financial professionals, allowing them to focus on critical information and make informed decisions. Future work could involve further enhancing the explainability features and expanding the range of financial documents and analysis tasks supported by the system.

Summary updated

BibTeX
@inproceedings{juan2021hive,
  author = {Yi-Ning Juan and Yi-Shyuan Chiang and Shang-Chuan Liu and Ming-Feng Tsai and Chuan-Ju Wang},
  title = {{HIVE: Hierarchical Information Visualization for Explainability}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2021},
  doi = {10.24963/ijcai.2021/709},
}
Figure