使用图表的 3D 旋转和其他 3D 效果来追求视觉效果,而不是表示 3D 数据,很可能会误导用户。以 3D 对象替代条形图中的条形即可。如果数据只按长度进行编码(就像标准条形图一样),那么读者可能会将按比例较大的对象解读为体积较大,因此值大于适当值。4如果设计师使用 2D 数据表示形式(例如气泡),并按半径或直径(而不是面积)对数据进行编码,则在比较每个细分(例如面积)时,也会造成难以比较的比例。5饼图还表示,所有细分加起来就是一个整体,不一定是这样。
[null,null,["最后更新时间 (UTC):2025-07-27。"],[[["\u003cp\u003eCharts and graphs can be misleading if not created and interpreted carefully, considering context, audience, and purpose.\u003c/p\u003e\n"],["\u003cp\u003eData visualizations consist of scaffolding (titles, axes, labels) and content (visual encoding of data like length, position, color).\u003c/p\u003e\n"],["\u003cp\u003eMisleading visualizations can arise from manipulated baselines, aspect ratios, 3D effects, and inappropriate color choices.\u003c/p\u003e\n"],["\u003cp\u003eStrive for clarity and honesty in data visualization, providing sufficient information without overwhelming the viewer.\u003c/p\u003e\n"],["\u003cp\u003eConsider data literacy of the audience and potential misinterpretations due to visual design choices when creating or reviewing visualizations.\u003c/p\u003e\n"]]],[],null,["# Visualization traps\n\n\u003cbr /\u003e\n\nCharts, graphs, and maps are compelling and persuasive tools for communicating\ninsight and information. They are also, when badly or maliciously deployed,\nsources of confusion, misinformation, and untruth.\n\nCharts as art rather than science\n---------------------------------\n\nML practitioners often visualize potential training datasets to understand their\nusefulness for models, as well as model outputs to understand performance.\n\nAlways ask about the intended **context, audience** and **purpose** of a data\nvisualization, whether you are creating or reading one. These three factors are\nkey to graphical communication. The same chart can be useful and insightful, or\nmisleading and exaggerated, in different contexts.^[1](#fn1)^ The intended viewer, and\nthe viewer's level of graph and data literacy, will vary. Design can help or\nhinder. For example, breathtakingly beautiful charts can be too convoluted to\nclearly communicate information.\n\nThere are no hard and fast rules for how to make a perfect chart, only\nguidelines and best practices. Visualizing data is as much an art as a\nscience. But when visualizing data, strive, above all, for clarity and honesty.\nProvide enough information to communicate clearly and accurately, and not so\nmuch information that it overwhelms the viewer.\n\nScaffolding, content, and misleading moves\n------------------------------------------\n\nAlberto Cairo, in *How Charts Lie* , splits data visualizations into two parts:\n**scaffolding** and **content**.\n\nA chart's scaffolding includes titles, axes, legends, labels, and the source of\nthe data, if given.\n\nContent includes the visual encoding of the data and any short textual\nannotations. Methods for visually encoding data commonly include:\n\n- length, as in bar charts\n- position, as in scatterplots\n- proportional angles, areas, and arcs in pie charts\n- color and hue\n- more rarely, width and thickness^[2](#fn2)^\n\nAll of these elements can be used to mislead. Starting a bar chart at a nonzero\nbaseline, or truncating the longest bars, can create inaccurate perceptions,\neven if the intent was to save space. See Sarah Leo's\n[essay](https://medium.economist.com/mistakes-weve-drawn-a-few-8cdd8a42d368)\non data visualization mistakes in the *Economist* for some examples.\n| **Important:** In certain contexts, a nonzero baseline does makes sense: in charts of average temperature scales, zero is not a special number, and in charts of life expectancy, zero is not a likely value. Context is crucial.\n\nAn inappropriate aspect ratio can make a small change seem very large,\nor a large change seem very small. Cairo suggests picking an\naspect ratio that matches the proportional change being depicted, say 3:1 for a\n30% change, but also advises close attention to context, because there are\nmany important exceptions to the rule. Relatively tiny fluctuations in average\nglobal temperature, say a rise of 2C out of 100C, for example, are\nhighly significant and will be understated in a chart with a 50:1 aspect\nratio.^[3](#fn3)^\n\nUsing 3D rotations of charts and other 3D effects for visual impact, rather\nthan for representing 3D data, is very likely to mislead. So are\ndepictions of 3D objects that substitute for bars in bar charts. If the\ndata is encoded by length only, as with a standard bar chart, the reader may\ninterpret the proportionally larger object as having a greater volume,\nand therefore a higher value, than appropriate.^[4](#fn4)^ Designers who use 2D\nrepresentations of data, like bubbles, and encode data by radius or diameter\nrather than by area, will also create misleading proportions.^[5](#fn5)^ 2D\nrepresentations like pie charts can make it difficult to compare segments\nagainst each other. Pie charts also imply that all segments add up to a whole,\nwhich may or may not be the case.\nExample of difficult-to-read data visualizations.\n\nColor is its own subject. In general:\n\n- Use 6 or fewer color divisions, since that is the limit of what most people can handle without confusion.\n- Avoid a wide selection of spectral hues, because different people order them differently.^[6](#fn6)^\n- If possible, select shades of a single hue, which is more distinguishable in grayscale.\n- Be aware of different types of [color blindness](https://developers.google.com/tech-writing/accessibility/self-study/sufficient-contrast).\n\nReferences\n----------\n\nCairo, Alberto. *How Charts Lie: Getting Smarter about Visual Information.* NY:\nW.W. Norton, 2019.\n\nHuff, Darrell. *How to Lie with Statistics.* NY: W.W. Norton, 1954.\n\nMonmonier, Mark. *How to Lie with Maps,* 3rd ed. Chicago: U of Chicago P, 2018.\n\nImage references\n----------------\n\n\"Chart of an example of Throughput Accounting structure.\" TAUser, 2008.\n[GNU FDL.](https://commons.wikimedia.org/wiki/Commons:GNU_Free_Documentation_License,_version_1.2)\n[Source](https://commons.wikimedia.org/wiki/File:ThroughputStructure.jpg)\n\n\"GWP (MTCO2E) for all vehicle types over life cycle.\" B2.Team.Leader, 2006.\n[Source](https://commons.wikimedia.org/wiki/File:B2.EIO-LCA_Summary.jpg) \n\n*** ** * ** ***\n\n1. Cairo 72-73, 79. [↩](#fnref1)\n\n2. Cairo 24-26, 36-38. [↩](#fnref2)\n\n3. Cairo 69-70. [↩](#fnref3)\n\n4. Huff 21-25. [↩](#fnref4)\n\n5. Cairo 34, 58-59. [↩](#fnref5)\n\n6. Monmonier 65-66. [↩](#fnref6)"]]