Make AI Explainable with User-Friendly Product Design

Table of Contents

1. Introduction
2. Understanding the Challenges
3. Targeting Different User Personas
   1. Marketers
   2. Analysts
   3. Data Wranglers
4. Assessing User Understanding and Proficiency
5. Identifying Design Issues
   1. Confusion with Model Performance Grade
   2. Misinterpretation of Churn Likelihood Chart
6. Redesigning the Dashboard
   1. Improving Chart Hierarchy and Labeling
   2. Enhancing Model Performance Visualization
   3. Implementing Progressive Disclosure
7. Evaluating the Impact of Redesigns
   1. Increased Confidence and Trust
   2. Quicker Response Time and Accurate Recommendations
   3. Request for Additional Data and Actionable Insights
8. Design Implications for Explainable AI
   1. Hover Annotations and Plain Language Descriptions
   2. Numerical Score with Explicit Scale for Model Performance
   3. Progressive Disclosure with Hierarchical Information
9. Conclusion

Surfacing AI Explainability in Enterprise Product Visual Design to Address User Tech Proficiency Differences

In this case study, we explore the challenges faced in surfacing AI explainability in enterprise product visual design to address user tech proficiency differences. The goal is to empower business end users with varying levels of AI and statistical expertise to make effective findings and decisions using enterprise products. We targeted three user personas - marketers, analysts, and data wranglers - to understand their specific needs and assess their understanding of AI explanations. Through the study, we identified design issues and subsequently redesigned the dashboard to improve user comprehension and confidence. The impact of the redesigns was evaluated, and we highlight the design implications for explainable AI in an applied setting.

1. Introduction

In today's data-driven world, enterprise products play a crucial role in providing insights and recommendations to businesses. However, not all end users have expertise in AI and statistics, making it challenging for them to interpret and trust the outputs generated by these products. This case study aims to address this challenge by focusing on the visual design of enterprise products to improve AI explainability and cater to users with varying levels of tech proficiency.

2. Understanding the Challenges

Before delving into the solution, it is essential to understand the challenges faced by users with different levels of AI and statistical expertise. While business end users may be experts in their respective domains, they often lack the necessary knowledge to interpret AI models effectively. Additionally, limited access to data experts further complicates the process of deriving Meaningful insights from enterprise products. To tackle these challenges, we conducted a study to gain insights into user understanding and proficiency.

3. Targeting Different User Personas

To ensure a comprehensive understanding of user needs, we targeted three distinct user personas: marketers, analysts, and data wranglers. Marketers are responsible for customer campaigns and content editing, whereas analysts access Relevant data and summarize it for decision-making. Data wranglers have expertise in statistics and are involved in data management tasks. Understanding the specific needs and challenges faced by these personas helped inform our design decisions.

3.1 Marketers

Marketers play a crucial role in creating customer campaigns and strategy recommendations. They possess domain expertise but may lack proficiency in AI and statistics. Understanding their perspective and information requirements was essential to design a user-friendly interface that empowers them to make effective decisions.

3.2 Analysts

Analysts are responsible for accessing relevant data and converting it into meaningful summaries. While they may have some familiarity with AI and machine learning, their understanding may vary. Identifying their specific needs and preferences allowed us to tailor the visual design to enhance their comprehension.

3.3 Data Wranglers

Data wranglers are statistical experts who handle data management tasks. They possess higher levels of AI and statistical expertise compared to marketers and analysts. Recognizing their proficiency level and providing them with the necessary tools and insights helped ensure optimal utilization of the enterprise product.

4. Assessing User Understanding and Proficiency

To gain a deeper understanding of user understanding and proficiency, we conducted a study involving participants with diverse AI and machine learning familiarity. Our aim was to identify potential differences in explanations and design needs between users with varying levels of tech proficiency. The study involved presenting participants with a visual system that analyzed customer behavior and product information to generate a predictive AI model.

5. Identifying Design Issues

Through the study, we identified several design issues that hindered user comprehension and trust. Participants struggled with interpreting the model performance grade, which lacked Clarity and had sociocultural ambiguities. The chart indicating churn likelihood also posed challenges due to inadequate labeling, leading to misinterpretations. Recognizing these issues was crucial for improving the visual design and providing clearer explanations.

5.1 Confusion with Model Performance Grade

Participants found it challenging to understand the model performance grade, which lacked a clear scale and was open to interpretation. Some participants assumed no error in the model, while others found it challenging to comprehend its meaning. The cultural connotations of the grading system also posed difficulties for non-American participants.

5.2 Misinterpretation of Churn Likelihood Chart

The chart indicating churn likelihood caused confusion among participants due to a lack of labeling. Many participants targeted a group with a smaller risk to churn simply because it had a higher number of customers. Improving the labeling and visual representation of the churn likelihood chart was crucial for better user comprehension.

6. Redesigning the Dashboard

Based on the identified design issues, we conducted workshops with experts within our business to redesign the dashboard. The goal was to make improvements that catered to a broad range of users, especially those with lower tech proficiency. We focused on creating a human-machine-explanation Scenario where users receive clear and concise explanations alongside the machine learning output.

6.1 Improving Chart Hierarchy and Labeling

To address the confusion surrounding the model performance grade, we redesigned the dashboard to prioritize chart hierarchy and labeling. The churn chart was moved to the left and provided increased labeling to enhance user understanding. This allowed users to focus on the essential factors influencing churn and make informed recommendations.

6.2 Enhancing Model Performance Visualization

To avoid over-reliance on the model performance grade, we replaced it with a numerical score accompanied by an explicit scale. This helped users interpret the model performance more accurately and avoid misinterpretations. The new visualization provided a clear metric for assessing model accuracy.

6.3 Implementing Progressive Disclosure

To provide users with contextual information without distracting from the visual context, we implemented progressive disclosure. This approach allowed users to access additional information through inline entry points and side panels. Plain language descriptions and hyperlinks provided users with a deeper understanding of the model's performance and history.

7. Evaluating the Impact of Redesigns

Once the redesigned dashboard was implemented, we conducted further assessments to evaluate its impact on user understanding, trust, and decision-making. Participants went through the same methodology as the original assessment, including understanding comprehension questions and making recommendations based on the given scenario.

7.1 Increased Confidence and Trust

The redesigned model performance chart instilled more confidence in participants, particularly those with lower tech proficiency. Participants felt the data was trustworthy, and they were more knowledgeable about its accuracy. This increased confidence empowered them to share the data's accuracy with their superiors and non-statistical experts.

7.2 Quicker Response Time and Accurate Recommendations

The improved labeling and on-hover annotations led to quicker response times and accurate recommendations by participants. The color scale used in the churn likelihood chart allowed for Instant conclusions and a better understanding of the likelihood to churn amongst all customers.

7.3 Request for Additional Data and Actionable Insights

Participants expressed a desire for further data and actionable insights rather than solely focusing on understanding the model itself. The redesigns provided resources that enabled users to understand how the data was being generated and make informed recommendations based on the available information.

8. Design Implications for Explainable AI

Based on our findings, we present several design implications for explaining AI in an enterprise product visual design scenario. These implications aim to enhance user understanding and confidence in data interpretation.

8.1 Hover Annotations and Plain Language Descriptions

Implementing hover annotations with plain language descriptions empowers users to make clear and concise conclusions while minimizing visual clutter. By providing information only when desired, users can focus on the essential aspects of the visual design.

8.2 Numerical Score with Explicit Scale for Model Performance

Displaying model performance metrics as a numerical score with an explicit scale avoids over-reliance on ambiguous grading systems. Providing a clear metric helps users accurately assess the model's performance and avoid misinterpretations.

8.3 Progressive Disclosure with Hierarchical Information

Progressive disclosure allows users to access additional information without navigating away from the visual Context. By providing hierarchical information and jargon-free descriptions, users can Deepen their understanding of the data interpretation process while maintaining focus on the primary visual design.

9. Conclusion

In conclusion, surfacing AI explainability in enterprise product visual design to address user tech proficiency differences is crucial for empowering users with varying levels of AI and statistical expertise. Through targeted user personas and iterative design processes, we identified design issues, implemented improvements, and evaluated the impact of the redesigns. The findings provide valuable insights and design implications for future projects aiming to make AI more explainable and accessible to a broader audience.

Highlights

• Designing enterprise products for users with varying levels of AI and statistical expertise is crucial for effective data interpretation and decision-making.
• Confusion with model performance grades and misinterpretation of churn likelihood charts hinder user comprehension and trust in enterprise products.
• Redesigning the dashboard by improving chart hierarchy, providing clearer model performance metrics, and implementing progressive disclosure enhances user understanding and confidence.
• Redesigned dashboards prompt increased confidence and trust, quicker response times, and accurate recommendations from users with varied tech proficiency levels.
• Design implications include hover annotations with plain language descriptions, numerical scores with explicit scales, and progressive disclosure with hierarchical information to improve AI explainability.

FAQs

Q: How does the redesigned dashboard improve user comprehension and trust? A: The redesigned dashboard improves user comprehension and trust by providing clearer explanations, enhancing visual hierarchy and labeling, and presenting model performance metrics with explicit scales. These design changes empower users to make accurate and confident recommendations based on the given data.

Q: What were the specific challenges faced by marketers, analysts, and data wranglers? A: Marketers struggled with AI and statistical concepts but were confident in the data they were seeing. Analysts had varying levels of AI familiarity and required relevant summaries from the data. Data wranglers possessed higher tech proficiency but understood the limitations of machine learning models and desired additional information.

Q: How does progressive disclosure benefit user understanding? A: Progressive disclosure helps users access contextual information without distracting from the visual context. By providing hierarchical information and jargon-free descriptions, users can deepen their understanding of the model's performance and history. This approach enhances user comprehension without overwhelming them with excessive information.

Q: What are the design implications for explainable AI in an applied setting? A: The design implications include using hover annotations and plain language descriptions to provide clear and concise explanations, displaying model performance metrics as numerical scores with explicit scales to avoid misinterpretations, and implementing progressive disclosure with hierarchical information to enhance user understanding and confidence in data interpretation. These design considerations make AI more explainable and accessible to users with varying tech proficiency levels.