Wait, Should I Buy That?
During my course "Data Visualization", we needed to find a question that I wish to answer relating to any topic of our own choice.
I have always been curious about the Interaction between the consumer and their grocery shopping.
I investigate the topic and find solutions to the questions formulated.
Timeline
4 weeks
Platform
Mobile
Context
As consumers shopping for food, we have more things to choose from more than ever. Most of the items we purchase at the grocery store fall under the category ‘low involvement and small differences between the choices - ‘habitual behavior’ in the buyer behavior matrix.
Take for example the box of cereals you would pick from the grocery store as a habit. How often does one really compare one to the other? Our exploratory data analysis suggests that we probably should! This is especially true if you have an underlying health condition like diabetes.
Making informed decisions becomes harder still because of these seemingly ‘small differences’. The nutrition information at the back of the label lists the nutrition facts but for a layperson, this information is not actionable. “How much sugar is too much sugar?” “Are all kinds of cheese bad if I have high cholesterol?”
The product opportunity gap that we identified solves exactly this - provide actionable insights to the consumer who is conscious about their health issues but needs help.
This product would help shoppers make informed decisions while purchasing food either online or at the grocery store.
Exploratory Data Analysis
I used a sample dataset that included details on the nutritional components of 7400 food products. The food products are categorized into larger categories such as cheese, oils, and nuts. To start exploring the data I decided to complete a correlation matrix (figure 3) to see how a subset of nutritional categories correlate with each other across the full dataset. Something I noticed below is that sugar is positively correlated with a few nutritional categories including fiber. This was interesting because fiber is often considered a healthy nutritional component and the matrix suggests that some products with high fiber also have higher sugar. This does not mean a food product with higher fiber and higher sugar is necessarily unhealthy but does demonstrate that deeper analysis is needed to understand what elements of a food product are actually unhealthy.
Who benefits from the problem being solved?
According to WHO, Noncommunicable diseases (NCDs) kill 41 million people each year, equivalent to 71% of all deaths globally. Cardiovascular diseases account for most NCD deaths, or 17.9 million people annually, followed by cancers (9.3 million), respiratory diseases (4.1 million), and diabetes (1.5 million). Health behaviors like tobacco use, lack of physical activities, unhealthy diets, and harmful use of alcohol all increase the risk of NCD but are modifiable. The other risk factors according to WHO is metabolic, including elevated blood pressure (19% of global deaths), obesity, high levels of glucose in the blood, and high levels of fat in the blood. Given the magnitude of the problem, enabling consumers to eat better needs innovation!
Is this the right solution, right now?
The number of health and fitness apps designed for smartphones has grown dramatically in recent years. In 2017, there were 325,000 available health apps to download from major app stores. In 2019 there were 68.7 million smartphone owners who used at least one health or fitness app on a monthly basis in the US. In 2020, the number of users increased to 87.4 million monthly health and fitness app users in the US, which shows the growth of the market, competition between the various apps, and high demand for fitness and health smartphone apps. The majority of the app is designed to help users reach specific, and individualized, diet goals (such as consuming fewer calories per day) and use notifications to alert users of their progress or lack of progress. Researchers have begun exploring the effectiveness of these apps in improving health outcomes for users and the results vary based on the apps used. However, one research study concluded that the 7 most popular diet-tracking apps in the US appear “to be useful for a wide range of dietary intervention studies.”
One of the most popular diet-tracking apps in Europe, Yuka, currently has 25 million users and launched in the US in 2020. This particular app is relevant to our project because they are trying to alert users to food scores associated with different products. A user of their app scans the barcode of an item and receives information relating to the item’s healthiness in the form of a food score. However, the app lacks a personalized health assessment for specific individuals based on their personal health indicators and instead provides general information, which might be helpful to a user. The app has grown quickly in popularity across Europe since 2017 and has succeeded in attracting users in the US since it launched here in 2020, which signals the need and interest in this type of application. An interesting aspect of the app’s value proposition is that it is independent and not funded by any food brands or food manufacturers so users feel like they can trust the recommendations more. I believe this is an important aspect of building a trusting consumer base and would follow a similar model.
How does our product solve the problem?
User Data
The user must download the application from an app store onto their smartphone and complete their user profile. The user profile includes data on personal health indicators. This includes traditional data included in most diet-tracking apps such as a person’s weight, height, BMI, gender, and age. In addition, the user must input health indicators that are specific to them. This includes the specifics of how much a user can consume of specific nutritional categories in order to maintain a healthy diet. Some examples of nutritional categories that users will be asked to enter daily consumption limits are sugar, carbohydrates, saturated fat, protein, and calories.
For example, a user might have a consumption limit of 5 units of cholesterol per day that they must follow in order to maintain a healthy diet. The user would enter this data point in their app profile. Now the app “knows” this specific user should not consume more than 5 units of cholesterol per day and this data can be compared to the ingredients of the food items the user is interested in purchasing. The app’s accuracy is dependent on the accuracy of the data they enter. In the event that a person is shopping for another person, then they would need to use the other person’s health indicators to ensure that the items selected match the end consumer’s dietary restrictions.
Categorization and rankings
When the user scans a barcode of a product the app will pull the ingredients and nutritional data about the specific item from the dataset of food products. The dataset of food products is organized into large categories such as cheese, oil, chicken, and nuts. The purpose of the categories is to compare the nutritional data of products within the same category so that alternatives can be offered. In the future, we hope to offer additional subcategories a user can select based on the type of meal they are creating. One example is that if a user is buying cheese for sandwiches, then the app would pull from a subcategory of cheeses specific for sandwiches. The dataset we are using has 7400 items listed with their nutritional data. In reality, this dataset will need to include millions of rows that capture a much wider range of products in the marketplace, so there are some limitations in the sample dataset we have been using. Once the user scans the item, the app will compare the ingredients and nutritional data of the item with the user’s profile. The backend of the app will be modeled after a simple decision tree, such as the one below, that shows how the app will provide an answer to the user on if they should purchase the item for their consumption. In this example, a user who has entered in their profile that they have high cholesterol and cannot consume more than 2 units of saturated fat per day would receive a recommendation to not purchase the item if it has more than 2 units of saturated fat.
To help prototype this process, I built some simple code in R Studio to show how the app would compare user data to a full dataset and provide a recommendation to the user about whether they should purchase an item for their consumption. For this demonstration, I selected the food category “cheese” and created a mock user profile that requires the user to consume no more than 5 units of cholesterol per day. The code below shows that if the selected cheese has more than 5 units of cholesterol, then the user should avoid the item. If it has 5 units or less of cholesterol, then the user can enjoy the item.
The chart below shows the results of the matching. In this particular case, one of the few kinds of cheese this user can “enjoy” is low-fat cottage cheese.
In the event that a user is directed to avoid a specific product, they will be provided with an alternative option to purchase that does match the requirements they listed in their user profile. The app will recommend that the user purchase the next healthiest item (healthy as defined by the user profile data they enter in their profile) within the food category that fits the user’s health parameters. Ultimately, we want users to select healthier items that they might actually purchase and consume. It would be easy to direct all consumers to the absolute most healthy option for them in each category, but research shows that humans are less likely to maintain a radically different diet from what they are used to. By following a process modeled after harm reduction theory, users are more likely to adopt incrementally healthy consumption habits. Harm reduction theory traditionally applies to behavioral health challenges that encourage less damaging habits, rather than moving immediately to abstinence.
Users will report in the app if they purchased the recommended alternative item or not. If the user decides to not purchase the recommended item, the app will learn to not suggest the same item to a user with a similar profile. For example, if a large number of users with a limit on how much cholesterol they can consume daily decide to not purchase the first recommendation they receive, then the app will stop recommending that item to similar users. As we have more users, we will have more data that can help us refine the recommendation process. This idea is modeled after Netflix’s recommendation system that has been incredibly successful at keeping users engaged with the Netflix platform. As the app is developed, we hope to include an option to purchase items directly through the app which will require users to share payment history information. This will provide clear data, and verification, of what users actually buy rather than relying on self-reports. However, we believe that users motivated to improve their health outcomes will be incentivized to accurately report their purchases.
References
https://www.who.int/news-room/fact-sheets/detail/noncommunicable-diseases
https://www.statista.com/statistics/1154994/number-us-fitness-health-app-users/
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6543803/
https://yuka.io/en/ https://www.similarweb.com/app/google-play/io.yuka.android/statistics/
http://motivatedandfit.com/wp-content/uploads/2010/03/Diets_dont_work.pdf