Introduction
Accurately recommending sizes on apparel sites plays an important role in improving customer satisfaction. Here at our company, Virutsize, we provide a virtual try-on service with the same name- ``Virtusize.'' It recommends the most suitable size based on product size data and physical information provided by the customer. In order to recommend an accurate size, our in-house data science team verifies results on a daily basis, regularly preforming brush-ups to improve accuracy.
In this article, we will show examples of how we improve recommendation accuracy, and reveal the series of steps to improve accuracy for our users with a high BMI!
Model size comparison method
One of the types of recommendations we provide uses a model for size comparisons (the model-size comparison method). This is a method of making size recommendations to customers, using the body measurement of a client-provided model, along with which size fits that model best.
The data is assembled based on a model's physical information as well as the product information on what the model is wearing. This is especially effective for making suggestions for a-typical types of clothing such as: flowy, low-cut, and loose fitting items.
Finding the issue and forming a hypothesis
When comparing measurements using the model- size comparison method, it was found that the *size match rate of Virtusize users with a high BMI, was less accurate than those with a more generic body type. So we decided to review the logic behind the recommendations.
*A look back at changes over the last year and a half, in BMI value and size matching rates, when using the model-size comparison method.
![](https://cdn.prod.website-files.com/5e75b27434e76c386cb00fc3/6690eb892417580740199915_BMI_SMR_chart_en.png)
Looking at this graph, the *size match rate decreases, as the BMI increases to 28 or higher. This makes it evident that the model size comparison method, is not very suitable for Virtusize users with a BMI of 28 or higher.
If you take a closer look at "model information" on the product page, used for the model size comparison method, you will find that most of the models are listed as having a standard body shape. This means wearing items such as "S", "M", and "L". However, there isn't a lot of information on models who actually wear sizes larger than that. For example, 2XL or 3XL, which are worn by users with high BMI. But, it is true that finding models to cover all sizes larger than 2XL could be costly and unrealistic for companies.
Because of the limited range of model information that we can obtain, Virtusize has come to the conclusion that internally, we need to apply a different logic to high BMI, users than the model size comparison method. However, as we said at the beginning, the model size comparison method is still a very good recommendation logic for users with general body types! Plus- changing the mechanisms of the entire site at once, ultimately, won't lead to improving the accuracy of the entire site.
Verifying the improvements and the end results
The in-house data science team conducted a study to apply the size comparison method to users with a typical body type, and to apply a different method to users with a high BMI. In the end, we were able to develop a mechanism that applies a different logic to those users instead of applying the model size comparison method.
After applying this new method, a report showed that the *size match rate of users with high BMI, (specifically with a BMI of 28 or higher) improved by up to 20% for each brand.
And even looking at the average of all brands, we can see an overall improvement in accuracy of 5.81%.
Conclusion
Even with this new method, we are continuously working to improve our product with the aim of making, "more accurate size recommendations." This time, we introduced a series of steps to improve accuracy for our users with high BMI, but this is still only one example of our work. We will continue to focus on different issues, and carefully optimize each item, one-by-one, in order to deliver a safe purchasing experience to more people than just users with more common body types.
Virtusize is continuously working to grow and improve. We look forward to showing you what we can do.
*Size match rate: This is a value that serves as our index for measuring the accuracy of the recommendation logic at our company. Of the products purchased after Virtusize was recommended…this is the percentage of products that were purchased in the size that Virtusize recommended first.
Introduction
Accurately recommending sizes on apparel sites plays an important role in improving customer satisfaction. Here at our company, Virutsize, we provide a virtual try-on service with the same name- ``Virtusize.'' It recommends the most suitable size based on product size data and physical information provided by the customer. In order to recommend an accurate size, our in-house data science team verifies results on a daily basis, regularly preforming brush-ups to improve accuracy.
In this article, we will show examples of how we improve recommendation accuracy, and reveal the series of steps to improve accuracy for our users with a high BMI!
Model size comparison method
One of the types of recommendations we provide uses a model for size comparisons (the model-size comparison method). This is a method of making size recommendations to customers, using the body measurement of a client-provided model, along with which size fits that model best.
The data is assembled based on a model's physical information as well as the product information on what the model is wearing. This is especially effective for making suggestions for a-typical types of clothing such as: flowy, low-cut, and loose fitting items.
Finding the issue and forming a hypothesis
When comparing measurements using the model- size comparison method, it was found that the *size match rate of Virtusize users with a high BMI, was less accurate than those with a more generic body type. So we decided to review the logic behind the recommendations.
*A look back at changes over the last year and a half, in BMI value and size matching rates, when using the model-size comparison method.
![](https://cdn.prod.website-files.com/5e75b27434e76c386cb00fc3/6690eb892417580740199915_BMI_SMR_chart_en.png)
Looking at this graph, the *size match rate decreases, as the BMI increases to 28 or higher. This makes it evident that the model size comparison method, is not very suitable for Virtusize users with a BMI of 28 or higher.
If you take a closer look at "model information" on the product page, used for the model size comparison method, you will find that most of the models are listed as having a standard body shape. This means wearing items such as "S", "M", and "L". However, there isn't a lot of information on models who actually wear sizes larger than that. For example, 2XL or 3XL, which are worn by users with high BMI. But, it is true that finding models to cover all sizes larger than 2XL could be costly and unrealistic for companies.
Because of the limited range of model information that we can obtain, Virtusize has come to the conclusion that internally, we need to apply a different logic to high BMI, users than the model size comparison method. However, as we said at the beginning, the model size comparison method is still a very good recommendation logic for users with general body types! Plus- changing the mechanisms of the entire site at once, ultimately, won't lead to improving the accuracy of the entire site.
Verifying the improvements and the end results
The in-house data science team conducted a study to apply the size comparison method to users with a typical body type, and to apply a different method to users with a high BMI. In the end, we were able to develop a mechanism that applies a different logic to those users instead of applying the model size comparison method.
After applying this new method, a report showed that the *size match rate of users with high BMI, (specifically with a BMI of 28 or higher) improved by up to 20% for each brand.
And even looking at the average of all brands, we can see an overall improvement in accuracy of 5.81%.
Conclusion
Even with this new method, we are continuously working to improve our product with the aim of making, "more accurate size recommendations." This time, we introduced a series of steps to improve accuracy for our users with high BMI, but this is still only one example of our work. We will continue to focus on different issues, and carefully optimize each item, one-by-one, in order to deliver a safe purchasing experience to more people than just users with more common body types.
Virtusize is continuously working to grow and improve. We look forward to showing you what we can do.
*Size match rate: This is a value that serves as our index for measuring the accuracy of the recommendation logic at our company. Of the products purchased after Virtusize was recommended…this is the percentage of products that were purchased in the size that Virtusize recommended first.
Introduction
Accurately recommending sizes on apparel sites plays an important role in improving customer satisfaction. Here at our company, Virutsize, we provide a virtual try-on service with the same name- ``Virtusize.'' It recommends the most suitable size based on product size data and physical information provided by the customer. In order to recommend an accurate size, our in-house data science team verifies results on a daily basis, regularly preforming brush-ups to improve accuracy.
In this article, we will show examples of how we improve recommendation accuracy, and reveal the series of steps to improve accuracy for our users with a high BMI!
Model size comparison method
One of the types of recommendations we provide uses a model for size comparisons (the model-size comparison method). This is a method of making size recommendations to customers, using the body measurement of a client-provided model, along with which size fits that model best.
The data is assembled based on a model's physical information as well as the product information on what the model is wearing. This is especially effective for making suggestions for a-typical types of clothing such as: flowy, low-cut, and loose fitting items.
Finding the issue and forming a hypothesis
When comparing measurements using the model- size comparison method, it was found that the *size match rate of Virtusize users with a high BMI, was less accurate than those with a more generic body type. So we decided to review the logic behind the recommendations.
*A look back at changes over the last year and a half, in BMI value and size matching rates, when using the model-size comparison method.
![](https://cdn.prod.website-files.com/5e75b27434e76c386cb00fc3/6690eb892417580740199915_BMI_SMR_chart_en.png)
Looking at this graph, the *size match rate decreases, as the BMI increases to 28 or higher. This makes it evident that the model size comparison method, is not very suitable for Virtusize users with a BMI of 28 or higher.
If you take a closer look at "model information" on the product page, used for the model size comparison method, you will find that most of the models are listed as having a standard body shape. This means wearing items such as "S", "M", and "L". However, there isn't a lot of information on models who actually wear sizes larger than that. For example, 2XL or 3XL, which are worn by users with high BMI. But, it is true that finding models to cover all sizes larger than 2XL could be costly and unrealistic for companies.
Because of the limited range of model information that we can obtain, Virtusize has come to the conclusion that internally, we need to apply a different logic to high BMI, users than the model size comparison method. However, as we said at the beginning, the model size comparison method is still a very good recommendation logic for users with general body types! Plus- changing the mechanisms of the entire site at once, ultimately, won't lead to improving the accuracy of the entire site.
Verifying the improvements and the end results
The in-house data science team conducted a study to apply the size comparison method to users with a typical body type, and to apply a different method to users with a high BMI. In the end, we were able to develop a mechanism that applies a different logic to those users instead of applying the model size comparison method.
After applying this new method, a report showed that the *size match rate of users with high BMI, (specifically with a BMI of 28 or higher) improved by up to 20% for each brand.
And even looking at the average of all brands, we can see an overall improvement in accuracy of 5.81%.
Conclusion
Even with this new method, we are continuously working to improve our product with the aim of making, "more accurate size recommendations." This time, we introduced a series of steps to improve accuracy for our users with high BMI, but this is still only one example of our work. We will continue to focus on different issues, and carefully optimize each item, one-by-one, in order to deliver a safe purchasing experience to more people than just users with more common body types.
Virtusize is continuously working to grow and improve. We look forward to showing you what we can do.
*Size match rate: This is a value that serves as our index for measuring the accuracy of the recommendation logic at our company. Of the products purchased after Virtusize was recommended…this is the percentage of products that were purchased in the size that Virtusize recommended first.