Loading field map
Data visualization in computer graphics is built on a single, persistent challenge: how to turn abstract numbers into images that the human visual system can interpret quickly and accurately. The difficulty is that the same data can be rendered in countless ways—as a scatterplot, a 3D volume, a treemap, or an interactive dashboard—and each choice shapes what a viewer sees and misses. The history of the subfield is a story of four frameworks that each offered a different answer to that challenge, and the relationships among them—absorption, competition, synthesis, and ongoing disagreement—define the field today.
Before computer graphics could visualize data, there had to be a theory of how graphical marks encode quantitative information. That theory emerged from statistics, not computer science. In the 1960s and 1970s, Jacques Bertin’s Semiology of Graphics (1967) laid out a systematic vocabulary of visual variables—position, size, shape, color, texture—and rules for how they map to different kinds of data. Bertin’s framework treated a chart as a sign system: every mark on a page carried a meaning that could be read if the encoding was consistent. Around the same time, John Tukey’s Exploratory Data Analysis (1977) championed the use of simple, fast-to-draw graphics—stem-and-leaf plots, box plots, scatterplot matrices—as tools for discovery rather than mere presentation. Tukey’s work shifted the purpose of visualization from confirming what the analyst already knew to revealing what the data might hide.
Statistical Graphics and Charting was a narrow framework by design. It assumed static, two-dimensional displays, small datasets that fit on a page, and a human analyst who did all the interpretive work. Its principles were precise but limited. The framework did not address interactive exploration, three-dimensional data, or the massive datasets that would soon arrive from scientific simulation and digital sensors. Those limitations set the stage for the next framework, which would borrow Bertin’s encoding rules but embed them in a radically different technological environment.
In 1987, a landmark report from the National Science Foundation, Visualization in Scientific Computing, argued that supercomputer simulations were generating data too vast and too complex for traditional charts. The report called for a new field—scientific visualization—that would use the full power of computer graphics to render three-dimensional, time-varying phenomena. Where Statistical Graphics had worked with flat marks on paper, Scientific Visualization embraced depth, lighting, motion, and interactivity. Its signature techniques—isosurfacing, volume rendering, flow visualization—were designed to represent physical quantities that already had a spatial structure: temperature inside a combustion chamber, blood flow through an artery, pressure across an airplane wing.
Scientific Visualization absorbed the encoding principles of Statistical Graphics but transformed them. Bertin’s visual variables still applied, but now they operated in a 3D scene with perspective, occlusion, and dynamic viewpoint control. The framework’s central commitment was to spatial fidelity: the visualization should look as much like the physical phenomenon as possible, because the goal was to let scientists see what the simulation had produced. This commitment created a sharp boundary. Scientific Visualization had little to say about data that lacked a natural spatial embedding—stock prices, social networks, database tables. Those kinds of data would require a different framework, one that rejected the assumption that spatial realism was always the best path to understanding.
By the early 1990s, researchers in human-computer interaction and computer graphics began asking a disruptive question: what if the data has no inherent shape? Information Visualization, crystallized in Stuart Card, Jock Mackinlay, and Ben Shneiderman’s Readings in Information Visualization: Using Vision to Think (1999), argued that the designer’s job was not to reproduce physical reality but to invent a visual structure that made abstract relationships legible. A file system is not a tree in nature, but a treemap or a cone tree can make its hierarchy visible. A social network is not a force-directed graph, but a node-link diagram can reveal clusters and bridges.
This framework directly competed with Scientific Visualization on methodological grounds. Scientific Visualization started with the data’s spatial coordinates and rendered them faithfully; Information Visualization started with the user’s task and invented a spatial mapping from scratch. Where Scientific Visualization prized realism, Information Visualization prized perceptual efficiency—the ability to encode as much information as possible in a display that the human eye could scan in milliseconds. The techniques were correspondingly different: scatterplot matrices and parallel coordinates for multidimensional data, treemaps for hierarchies, adjacency matrices for networks. Information Visualization did not replace Scientific Visualization; it carved out a domain that Scientific Visualization could not address. The two frameworks coexisted, each dominant in its own territory, but their assumptions remained in tension. Should a visualization look like the world, or should it look like a diagram of the data? The field has never fully settled that question.
By the mid-2000s, both Scientific Visualization and Information Visualization faced a common limit: they assumed that a human analyst would drive the exploration. But datasets had grown too large for manual browsing. A single genome-wide association study or a day’s worth of network traffic logs could contain millions of points—far more than any interactive scatterplot could display without overwhelming the user. Visual Analytics, defined in the 2005 report Illuminating the Path: The Research and Development Agenda for Visual Analytics, proposed a synthesis: combine interactive visualization with automated data mining and machine learning. The computer would find patterns algorithmically; the human would interpret, refine, and decide.
Visual Analytics did not reject its predecessors. It absorbed them. A visual analytics system might use a Scientific Visualization-style 3D render for geospatial data, an Information Visualization-style parallel coordinates plot for high-dimensional variables, and a statistical model that highlighted outliers or clusters in real time. The framework’s distinctive contribution was the tight coupling between computation and perception. The algorithm did not just pre-process the data; it responded to the user’s interactions, updating its model as the analyst zoomed, filtered, or selected. This created a new debate that earlier frameworks had not faced: how much control should the algorithm have? If the machine highlights what it considers important, does it bias the analyst’s exploration? Visual Analytics introduced the tension between automation and human judgment, a tension that remains unresolved and productive.
Today, all four frameworks remain active, but they have settled into a division of labor. Statistical Graphics and Charting survives as the foundation of every plotting library and business dashboard; its rules for visual encoding are so thoroughly absorbed that most practitioners use them without knowing Bertin’s name. Scientific Visualization dominates in engineering, medicine, and the physical sciences, where data has a natural spatial home. Information Visualization leads in business intelligence, network analysis, and digital humanities, where the data is abstract and the design challenge is to invent a useful visual metaphor. Visual Analytics is the framework of choice for big-data platforms, cybersecurity monitoring, and any domain where the volume of data exceeds what a human can scan unaided.
The frameworks agree on one fundamental principle: visualization is a tool for thinking, not just for presenting. They disagree on what kind of thinking matters most. Scientific Visualization assumes that seeing the shape of a physical process is the primary cognitive act. Information Visualization assumes that comparing abstract quantities and detecting patterns in high-dimensional space is the primary act. Visual Analytics assumes that the primary act is a dialogue between human pattern recognition and machine computation. These are not contradictions that will be resolved by a single unified framework. They are different answers to the same question—how to map data to vision—and the field’s vitality comes from keeping all four in play, each pushing against the others’ limits.