The Art and Science of Presenting Data: A Guide for Data Scientists

As artificial intelligence and machine learning revolutionize industries, the role of the data scientist has never been more critical. Data scientists are at the forefront of extracting insights from the ever-increasing amounts of data being collected. However, uncovering game-changing insights is only half the battle. Equally important is the ability to effectively communicate and present those findings in a way that drives impact.

This is where data visualization comes in. The ability to present complex data in a clear, engaging, and visually compelling way is now a core skill for data scientists. A survey by Tableau and Harris Poll found that 92% of senior executives believe their organization is missing out on opportunities because staff lack the skills to effectively present data visually.^1 Well-designed visualizations significantly increase the speed and accuracy of data-driven decision making.

In this guide, we‘ll dive deep into data visualization theory and practice from the perspective of a data scientist working on AI/ML projects. We‘ll cover essential concepts like choosing the right visualizations, data storytelling, designing interactive dashboards, and the latest tools and technologies. Whether you‘re a beginner or a seasoned pro, this guide will level up your ability to create data visuals that inspire aha moments and drive business value.

Why Data Visualization is a Critical Skill for Data Scientists

While often overlooked in favor of technical abilities, data visualization is actually one of the most high-impact skills a data scientist can develop. Here‘s why it‘s so crucial:

1. Bridge the gap between insights and action. Having a groundbreaking insight is great, but it doesn‘t matter if you can‘t get decision makers to understand and act on it. Research by Deloitte found that companies who excel at data visualization are 2.3 times more likely to be performing significantly ahead of peers.^2 Clear, compelling visuals help translate insights into real business outcomes.

2. Make the invisible visible. We‘re collecting data today on pretty much everything, but most of it is invisible to the human eye in the form of massive databases of numbers. Data visualization makes the patterns, trends, and outliers hidden in the data immediately apparent. As Fernanda Viegas and Martin Wattenberg, legendary data viz pioneers now at Google, put it: "Visualization is a method of computing that leverages the human visual system to help unearth insights."^3

3. Facilitate visual discovery. Interacting with data visually allows analysts and stakeholders to quickly test out theories and spot unexpected patterns. Visualization is an essential part of exploratory data analysis (EDA).

4. Tell persuasive data stories. As author and Tableau Chief Evangelist Ben Jones notes: "Data storytelling weaves data and visualizations into a narrative tailored for a particular audience in order to convey credibility in the analytical approach, confidence in the results, and a compelling set of insights that is actionable to the audience."^4 Charts are powerful tools for making a data-driven case.

5. Collaborate across teams. Visuals provide a common language that allows business stakeholders, designers, product managers, and engineers to get on the same page about the data. Visualization democratizes data science insights.

According to Martin Chalfie, Nobel Laureate in Chemistry: "What you spend years building in the lab, may only exist for a few minutes during your talk. Make your data beautiful and accessible."^5 The same holds true for data scientists. Visualization breathes life into your work and gives it a lasting impact.

Designing Effective Visualizations for AI/ML Projects

Now let‘s get into some specifics on creating great data visuals, with a focus on common AI and machine learning use cases. Follow these tips to take your visualizations to the next level:

Choose the Right Chart Type

One of the keys to effective visualization is matching the chart type to the data and message. Different types of charts are suited for presenting different types of data relationships:

• Line charts show trends and changes over time, like error rates during model training
• Bar charts compare discrete categories, like feature importance scores
• Scatter plots show the relationship between two continuous variables, like clusters in a dataset
• Heatmaps reveal patterns in tables of numbers, like a confusion matrix
• Box plots illustrate the distribution of a dataset, useful for comparing performance across models
• Tree charts display hierarchical data, like decision trees

Carefully consider what point you want to make with the data and choose the chart type that does so most clearly. Check out data-to-viz.com^6 for a comprehensive flowchart of which chart types to use for different data stories.

Simplify and Highlight Key Points

One of the biggest pitfalls in data visualization is trying to show too much at once. Resist the urge to cram in every data point. Aim for each visual to convey one main message.

Guide the viewer‘s attention to the key points you want them to take away using visual cues like color, size, and positioning. Use text callouts and annotations to spell out key insights.

Legendary data viz pioneer Edward Tufte introduced the concept of the "data-ink ratio," advocating that "a large share of ink on a graphic should present data-information."^7 Cut out extraneous visual elements that don‘t add informational value.

Tell a Data Story

The most engaging and memorable visualizations don‘t just present data, they tell a story. Craft a clear narrative arc and use a series of connected visuals to gradually reveal the story. Provide the necessary context up front, then emphasize key events and wrap up with a clear conclusion or call-to-action.

Airbnb is known for its compelling data stories. For example, their project "Snow White and the Seven Dwarfs" told the story of how seven key factors like price and availability impact bookings. A series of creative visuals walked through how Airbnb optimizes each factor.^8

As Cole Nussbaumer Knaflic advises in her book "Storytelling with Data," think of your data visualizations as a form of narrative: "Constructing a good data visualization is like writing a good novel: you need to have some overarching big idea and then ruthlessly edit out everything that doesn‘t serve that idea."^9

Embrace Interactivity

Today‘s leading data viz tools make it easy to create interactive dashboards that let users filter, drill-down, and explore the data themselves. This is especially important for AI/ML projects where users need to be able to dive into model performance, hyperparameters, datasets, etc.

Some specific examples of interactive visualizations for AI/ML:

• Confusion matrices where you can click on each cell to see example misclassifications
• Model performance comparison tools to see error rates, training time, inference latency, etc. side-by-side
• Dashboards for monitoring drift in production models over time
• Interactive tools for exploring and visualizing training datasets

Google‘s People + AI Guidebook notes that interactive explanations of AI systems foster user trust, understanding, and control.^10

Follow Data Viz Design Best Practices

To take your visualizations from good to great, incorporate these key design principles:

• Choose an intentional color palette. Use a simple color scheme with 2-6 colors. Employ colors strategically to highlight key data points. Avoid using default color schemes like Excel‘s. Ensure colors are easily distinguishable, even for those with color blindness. Tools like ColorBrewer and Adobe Color are useful for selecting palettes.^11

• Make text legible. Ensure all text is readable. Use clear fonts and large enough type. Don‘t make viewers squint to read labels.

• Include essential components. Every chart should have a descriptive title, clearly labeled x and y axes, a legend (if needed), and data source.

• Keep scales consistent. When comparing multiple charts, keep the axis ranges consistent between them so viewers can accurately compare.

• Avoid chartjunk. Cut out any unnecessary visual elements that distract from the data, like garish backgrounds or 3D effects.

• Make it accessible. Don‘t rely solely on color to convey meaning. Use patterns/textures as well to differentiate. Provide alt text for images.

Visualize Throughout the Data Science Workflow

Data visualization isn‘t just for final results. It should be incorporated throughout the data science process:

• Data Exploration: Visualize datasets using histograms, scatter plots, etc. to spot initial trends and outliers. Helps validate data quality.

• Experimentation: Visually compare results across different models and parameters. See error rates over time.

• Optimization: Chart how performance improves with different optimization techniques.

• Communication: Craft visual stories to share insights with stakeholders.

• Monitoring: Build dashboards to track model performance in production over time.

At Netflix, data scientists visualize a metric called the "efficiency frontier" at each stage of a project to determine if the tradeoffs between compute power and model performance are worthwhile.^12

Choosing the Right Tools

There are a range of powerful data visualization tools available, from drag-and-drop business intelligence software to code-based libraries. Here‘s an overview of some of the most popular:

Tool	Best For	Example	Learning Curve
Tableau	Business intelligence & interactive dashboards	^13	Low
PowerBI	Microsoft-based business intelligence and data viz		Low
Matplotlib	Creating basic charts in Python	^14	Medium
Seaborn	Statistical data visualization in Python	^15	Medium
Plotly	Interactive web-based visualization in Python or R	^16	Medium
D3.js	Complex, highly customizable web-based data viz	^17	High

For AI/ML projects, it‘s often best to use a code-based tool that allows you to integrate visualizations directly with the rest of your workflow. As a general principle, aim to automate as much of the visualization process as possible so charts can be easily updated as the data changes.

It‘s also ideal if the tool connects to a web-based interface for easy sharing, like Plotly‘s Chart Studio,^18 or Tensorflow‘s TensorBoard.^19 This allows data viz to be a collaborative, ongoing process.

Keeping Up with Advances in Data Visualization

The field of data visualization is rapidly evolving. To get and stay ahead of the curve, keep tabs on these advancements:

• Virtual and augmented reality for data visualization. Imagine walking through a 3D visualization of data^20
• Natural language interfaces for data visualization, like technology previewed by Tableau^21
• AIOps tools that automatically surface insights and generate visualizations^22

Leading data visualization practitioners are also always sharing new ideas:

• Semiotic A.I. is doing intriguing work on explaining AI systems through visual interfaces^23
• Nadieh Bremer‘s Visual Cinnamon blog showcases stunning data art projects^24
• The Data Visualisation Society‘s Nightingale publication is a great resource^25

Make a habit of studying the data visualizations in leading publications and reflecting on what makes them effective. The New York Times, The Economist, Bloomberg, National Geographic, and The Pudding publish some of the most cutting-edge data stories. Inspiration is everywhere.

Conclusion: Continued Growth as a Data Viz Craftsperson

As a data scientist, you have a powerful voice in your organization. Through compelling data visualization, you can enlighten business strategy, influence product development, and shape decision making. You can take findings from abstract to actionable.

Treat learning data visualization as a continual process rather than a one-time effort. The more you practice and study the craft, the better you‘ll become at making an impact with data. Challenge yourself to find creative ways to visualize results at each stage of your AI/ML projects. Show, don‘t just tell.

Ultimately, the goal of data science is to drive positive change in the world. Doing that requires getting buy-in and inspiring action from those outside the data science team. Hone your ability to create clear, eye-catching, story-driven visualizations and you‘ll be well on your way to achieving that goal. You‘ll help make data science an integral part of organizational culture.

As Florence Nightingale, one of the pioneers of data visualization, astutely noted: "The most important practical lesson that can be given to nurses is to teach them what to observe – how to observe – what symptoms indicate improvement – which are of none – which are the evidence of neglect – and what kind of neglect."^26 The same could be said of data scientists today. Our visualizations illuminate reality and neglect alike – and inspire progress.