HPC-Vis: A Visual Analytic System for Interactive Exploration of Historical Painter Cohorts

TL;DR

HPC-Vis tackles the challenge of exploring vast, heterogeneous data on Chinese historical painters by automatically reconstructing inheritance networks into an interpretable forest and by modeling artistic styles with a unified, three-level label system derived from large language models. The visual analytics suite combines an Inheriting Mountain View, a foldable Artistic Label View, and multi-view geographic and identity panels to support a human-in-the-loop cohort discovery workflow, including a painter recommendation engine that fuses style, time, geography, identity, and lineage signals. The paper validates the approach with two case studies and a user study, showing improved efficiency and accuracy in identifying and validating painter cohorts and their stylistic lineages. Overall, HPC-Vis offers a practical prototype that blends automated inference with domain knowledge, enabling historians to systematically uncover, define, and compare painter cohorts and potentially extend to other historical domains.

Abstract

More than ten thousand Chinese historical painters are recorded in the literature; their cohort analysis has always been a key area of research on Chinese painting history for both professional historians and amateur enthusiasts. However, these painters have very diverse artistic styles and an extremely complex network of inheritance relationships (e.g., master-apprentice or style imitation relationships); traditional cohort analysis methods not only heavily rely on field experience, but also cost a lot of time and effort with numerous but scattered historical documents. In this paper, we propose HPC-Vis, a visual analytical system for interactive exploration of historical painter cohorts. Firstly, a three-stage reconstruction algorithm for inheritance relationships of painters is proposed, which automatically converts the complex relationship graph of historical painters into a forest structure that contains multiple trees with clear inheriting chains, and we visually encoded this forest as a mountain map to intuitively show potential cohorts of historical painters. Secondly, a unified artistic style label system with three levels (i.e., subjects, techniques, and emotions) is established by using large language models, and it is further visually encoded as a new foldable nested doughnut chart. Finally, a visually guided human-computer collaborative interactive exploration mechanism is constructed, in which a painter cohort recommendation model is designed by integrating style, identity, time, space, and relationships. Two case studies and a user study demonstrate the advantage of HPC-Vis on assisting historians in discovering, defining, and validating cohorts of historical painters.

Figures (5)

• Figure 1: A graph visualization of inheritance relationships of Chinese historical painters from a book "a data spectrum of inheritance relationships of Chinese painters"wang2022, it is rendered using a classic open source software for graph and network analysis called GephiGephi. However, it is very challenging to analyze inheritance relationships and cohorts of painters with such a graph visualization, even for professional historians.
• Figure 2: Our reconstruction algorithm of the relationship network automatically converts a complex graph to a forest structure that contains multiple trees that reveal potential cohorts. It consists of three major steps (➀ $\rightarrow$➃) and a visual encoding (➄). ➀ $\rightarrow$➁: painters sharing the same set of masters are first grouped to Logic Units; ➁ $\rightarrow$➂: Logic Units are clustered in an iterative top-to-down manner; ➂ $\rightarrow$➃: The topology is further optimized to construct the tree structure for each cluster; ➃ $\rightarrow$➄: A mountain map is designed to visualize the forest, each hill represents a tree, which reveals the inheritance relationships intuitively.
• Figure 3: Our system has an export-in-the-loop exploration Workflow. According to the visual sights provided by our visualization, users can make judgments and selections through the interactive interface; then the system will recommend a potential cohort. Users can validate this potential by a cross validation, and confirm it if they are satisfied. If they are not satisfied yet, users are encouraged to observe more related painters recommended by the system, and collaborate with the system again in the next loop.
• Figure 4: Artistic Label View. (a) Overall structure of the Artistic Label View; (b)Bidirectional highlight effect via hovering; (c)Drill-down view after clicking a label combination.
• Figure 5: Statistics of user study. Left: average accuracies of volunteers' answers, note that the accuracies of Q3 & G4 of Group A are much higher than Group B, which demonstrates the advantage of HPC-Vis. Right: View rating, whose overall average is 4.425.