Qualitative Analysis of Structured Generative Interviews

Objective:

Understand user perceptions, concerns, interests, and desired features for an ML-powered food classification app. Session will be audio-recorded.

Open-ended Questions

1. Do you currently or have you ever tracked or logged your food?
2. What methods or tools do you currently use to track your food?
3. What were your primary goals in tracking your food, and what metrics do you prefer to track?
4. What challenges do you face when tracking or logging your food?
5. What are the best features of food logging apps you have previously used, and why?
6. How familiar are you with AI and machine learning?
7. How would you feel about an app that automatically recognizes and logs your food using AI?
8. What concerns would you have about using an AI-powered food logging app? (e.g., accuracy, privacy, data usage)
9. What, if anything, could be done to address those concerns?
10. What features would you find most beneficial in such an app?
11. Are there any specific foods or cuisines you’d want the app to recognize with high accuracy?
12. Would you be interested in having the app log servings or nutritional information based on the recognized food? Why or why not?
13. How important is the app’s speed and responsiveness in recognizing and logging food to you?
14. What would make you trust the app’s food recognition more?
15. Would you prefer a more hands-on approach where you can correct the app’s mistakes, or would you prefer the app to work more autonomously?
16. Are there any features or functionalities you would definitely NOT want in such an app?
17. How do you feel about the app learning from your corrections or inputs, and creating a customized food logging profile for you to increase logging accuracy?
18. What would encourage you to use this app regularly?
19. What social or gamelike features would improve your consistency in food logging?

Interview 1

Theme 1: Usability and Time/Equipment Constraints

• The interviewee expresses a preference for an app that is not time-consuming and does not add significant overhead to the user’s daily routine.
• The interviewee highlighted the difficulty of trying to log individual ingredients while cooking, and would prefer an app that does not require so much effort to log home-cooked meals.
• Quote: “If I cook I add so many things… especially Chinese style [foods], you have to add so many vegetables… and then how can I take a picture one by one [to] calculate how many calories I take in?”
• The interviewee complained that there is usually no barcode for fresh foods, making it difficult to log these foods.
• The interviewee also stated they did not have access to a full kitchen, only a hot pot at home, and they ate out on-campus and at restaurants much of the time. If a recipe feature were added, it might be useful to filter recipes by cooking methods / equipment available.

Theme 2: Accuracy and Recognition of Diverse Foods

• The user’s top concern with using an AI model to calculate nutrition information was accuracy of the app’s food recognition and portion estimation capabilities.
• It was very important for the AI model to be able to recognize diverse foods, especially Asian foods which are less common on the US mainland.
• Quote: “The food in Hawaii can be very diversified… like Filipino food and Korean food.”

Theme 3: Privacy

• The user had no concerns / did not care about privacy.

Theme 4: Emotional and Psychological Impact

• The interviewee is concerned about the potential for the app to induce feelings of guilt or anxiety regarding food choices.
• Quote: “For example, like French Fries, I don’t want to know [how many calories it is], I just want to enjoy my food and not feel guilty.”
• This theme suggests the importance of positive reinforcement and motivational messaging in the app’s design.
• Quote: “I don’t want the app to make me feel stupid or make me feel like, okay, I shouldn’t eat this or I shouldn’t eat that and so on… If maybe someone is thinking, like, ‘oh, I’m logging these french fries, I know it’s not the healthiest thing’, you make it so it’s not triggering these bad feelings in the person. Just give them feedback in a positive way, for example, don’t say ‘oh you already reached your calorie intake limit for today’ but instead say ‘oh you know by the way you ate some french fries, and I hope you enjoyed eating those; would you perhaps consider some other way to improve your habits today, like saving other high-calorie foods for another day?’ And then yeah, something like that, just a nice way to talk to people, just giving them constructive feedback.”

Theme 5: Personalization and Gamification

• The interviewee likes the idea of creating a personified character to represent the AI and make the process of correcting the AI’s mistakes more enjoyable.
• They suggested rewards for the user when they correct the mistakes of the AI.
• Quote: “Then they can work with that character and then [it feels like] they are talking to a real person or they can have a relationship or some emotional bonds between the character, the user, and the app.”

Theme 6: Nutritional Awareness and App Feedback

• The interviewee shows a keen interest in gaining feedback on how their habits are improving over time.
• There is an expressed desire for the app to provide educational insights into healthier eating habits.
• Quote: “And then also the result… I want to see that. I want to see that by using this [app] I’m eating healthier and then I’m getting better, in a way; I’m changing. I stop eating the food that makes me feel guilty, or I don’t feel guilty and just accept that I love junk food, but then I naturally switch to healthier eating habits. So if people feel that by using this app there are some behavioral changes then I think that the user will really appreciate that.”

Theme 7: Easing the “Learning Curve” for the ML Model

• The participant emphasized that the app should give users a heads that the model will get more accurate over time, as the user trains it.
• Quote: “[Just give the user] a little bit of heads up like ‘oh, you know, when you start to use the app you will need to work with it to help it learn and then after one week or one month, once the app learns more about your information, you can expect such-and-such accuracy from the model” because people have their own, what we call, imagined affordances, where they imagine what the app should do in their head, and we want to make it clear what to expect in the beginning in order to close the gap between what the designer thinks the user should do and what the user thinks the app should do.”

Conclusion and Recommendations:

• Simplicity and Speed: Make the app as quick to log food as possible; minimize disruption to the user’s cooking and eating routines.
• Accuracy and Learning: For a commercially viable product, the app must be very accurate at recognizing foods and calculating portion sizes, while simultaneously being effective for users who may consume various cuisines in different parts of the world.
• Privacy: Privacy is of low concern relative to other concerns about the app.
• Psychology: Avoid negative messaging or causing the user anxiety, as they may already be anxious about their food choices. Use positive reinforcement to make sustainable changes over time.
• Gamification: A gamified “AI-bot” or pet to train could help the user become more invested in training the AI model.
• App Feedback: Provide unique insights into behavior trends so users can see themselves improving over time.
• Expectation management: Ensure the user understands up front the app’s initial capabilities and the approximate time required to personalize the AI model for maximum accuracy.

Interview 2

Theme 1: Accuracy and Reliability

• The participant expressed frustration with searching for foods in the app database and trying to correctly guess portion sizes.
• The participant enjoyed barcode scanning when it worked but often got error messages stating that the barcode was not present in the database.
• Quote: “I sometimes couldn’t find the food I was eating and if I was eating even like hamburger meat, it was tough to track it because I didn’t know what the… portions of food were.”

Theme 2: Ease of Use

• The participant was not currently tracking calories because of frustrations with ease of use, but was open to trying it again if a user-friendly app came along.
• Manually searching the database felt complicated and the participant often wasn’t successful in finding desired food items.
• When the barcode scanner did work, it was very convenient.

Theme 3: Positive Feelings Toward AI

• The user had generally positive feelings about AI and mainly was concerned about whether it would be effective at recognizing foods.
• The user showed interest in an app that adapts and learns from their inputs, suggesting a desire for personalized experience.

Theme 4: Quick Logging

• The interviewee wanted an accessible button to log food that takes the user straight to the camera, then shows the results and calculated nutritional information.

Theme 5: Data Privacy and Security

• While the user personally did not express concern about privacy, they acknowledged that others might have reservations about an app that collects data on their eating habits.
• The user believed that a main privacy concern for some might be worries that their data would be sold to advertisers.

Theme 6: Social and Gaming Elements

• The user expressed interest in social features, such as the ability to share progress with friends or engage in friendly competition. This suggests that incorporating social or gamified elements could enhance user engagement and consistency in using the app.

Theme 7: Nutritional Guidance and Meal Planning

• The user liked the idea of the app offering nutritional guidance and personalized meal planning based on their eating habits.
• Participant expressed an interest in AI-powered meal planning that waited a month or so to learn the user’s habits, then constructed a meal plan for them that included favorite foods but adjusted amounts to achieve a better nutritional profile.

Theme 8: Extra Features Should be Unobtrusive

• Participant liked the idea of many extra features such as workout programs, adding friends, social groups and social media-like features, and meal plans, but felt that these extra features should avoid distracting the user from the main functionality of the app.

Conclusion and Recommendations:

• Simplicity and Accuracy: Make the logging process as straightforward as possible for the user; focus on developing robust AI capabilities for precise food recognition and portion size estimation.
• Quick-access food logging: Make the button for logging food available on all pages and have it go straight to the camera.
• Privacy: Provide a clear data privacy policy that lets users know that their data will not be secure and will not be sold to advertisers.
• Social Interaction and Gamification: Integrate social features that allow users to connect and compete with friends, enhancing motivation.
• Meal Planning: Offer customized meal plans and dietary suggestions based on user’s preferred foods to make it easy to achieve goals with fewer trips out of a user’s “food comfort zone.”
• Minimize Distractions: Offer extra features to users but make sure they are not distracting or in the user’s face; perhaps allow the user to turn on or off certain features when they set up the app, then adjust later in the settings menu.

Interview 3

Theme 1: Diverse Functionalities and User-Friendly Interface

• The participant has used three apps simultaneously to track and manage their food intake (MyFitnessPal, Ate, Yummly), each with different functionalities.
• Participant wishes there were a single app that could fulfill all three functions.

Theme 2: Visual versus Numerical Representation

• A preference for visual representations over numerical data is evident. The participant liked the visual journaling aspect of Ate and suggested the inclusion of visual elements like graphs or trends to track progress.

Theme 3: Privacy and Data Security

• Privacy concerns were a significant factor. The participant emphasized the need for clear communication about data usage, storage, and security. Options like end-to-end encryption and storing data on the user’s device were viewed positively.

Theme 4: Cost and Subscription Models

• The participant had previously paid for subscriptions but eventually decided to stop paying for certain features and switched to the free tier of an app. There’s a preference for free or trial versions before committing to payment.

Theme 5: Social and Motivational Features

• Social features, like connecting with friends and setting up challenges, were seen as beneficial for motivation and consistent app usage.

Theme 6: Recognition of Diverse Foods

• The user noted that many apps fail to recognize diverse, ethnic foods, particularly those specific to Hawaiian cuisine. An AI-powered app that can accurately identify a wide range of foods would be more appealing.

Theme 7: Feelings Toward and Expectations of AI

• Participant feels that AI for food logging is a good idea because of its ability to learn user preferences over time.
• The user likes the idea of an AI feature that can automatically recognize foods and calculate portion sizes and nutritional information.

Recommendations:

• Unified App Experience with Options to Show or Hide Features: Develop an app that combines the best features of different apps, offering a range of functionalities within a single, user-friendly platform.
• Visual Data Representation: Incorporate visual elements for tracking progress and data representation.
• Strong Privacy and Security Measures: Clearly communicate privacy policies and incorporate robust data security measures.
• Cost-Effective Model: Consider offering an extended free or trial version before introducing a paid subscription.
• Social and Gamified Engagement Features: Integrate gamified social functionalities to foster community engagement and motivation.
• Diverse Food Recognition: Focus on the AI’s ability to recognize a wide range of cuisines, catering to cultural diversity in dietary habits.
• Expectations of AI: A commercial-level product would be expected to have an 80% top-1 accuracy to be desirable by this participant.