In Sanskrit, the avatar (Avatar) Refers to the “incarnation of human form”. In Roblox, few things reflect the user’s identity more directly than their avatar. As we will discover, there is no “standard” Roblox user, the fantasy aesthetic diversity in our user profile directly reflects the diversity of the user base itself.
Representation incarnation (methodology)
If we are interested in aesthetic diversity, we need to start with the representational incarnation aesthetics.The most natural place is the 2D avatar thumbnail, which often represents the user. For aesthetic analysis, we need to convert this thumbnail into a semantically meaningful digital representation. There are many ways to reduce the dimensionality, but we can try the following methods.
- The easiest way: apply directly Principal component analysis To the flattened thumbnail image. To evaluate the “quality” of the reduction, we visualized thumbnails of the extreme cases of the principal component (PC). We can see that although the first PC distinguishes between interpretable avatar types, the twelfth is too broad to make sense.
PC 1 (explains 14.3% of the difference):
PC 12 (explains the difference of 1.5%):
2. It’s almost as simple: we can apply the last hidden layer of the ready-made pre-trained image classification network (Resnet 18) and evaluate the embedding quality by clustering them. Observe how Resnet captures color information very effectively (see all blue shoes in the second cluster), but sometimes fails to encode shape information (see first cluster).
Examples of thumbnails from 2 clusters are as follows:
3. In order to understand cohesion intuitively, we can apply UMAP to reduce the image classification embedding to 2 dimensions. Although the dose seems to be discernible clusters, the large spots in the lower right corner look suspicious. It is true: the samples from this giant cluster are visually disjointed.
Two-dimensional embedded image:
Samples from giant clusters in 2D embedding space:
4. Train a small custom variational autoencoder (VAE) directly on the thumbnail data. Ideally, this can better capture the unique aesthetic changes in Roblox’s avatars compared to general image classifiers. (By the way: K-means is particularly suitable for clustering these embeddings, because its normal prior matches the latent variable posterior of VAE)
Although there are some indicators to try to quantify the benefits of different methods, the actual use cases of unsupervised learning usually boil down to subjective judgments. Interestingly, we found that #4 was the most successful.
Using VAE, we can convert thumbnails into compact 64-dimensional vectors for clustering. Here are some examples of VAE + K-means clustering from 20-way clustering:
Some very customized avatars in a cluster:
The tall and thin avatar, we call “Rthro” in another cluster:
In this cluster we call the massive incarnation of “lumps”:
The default avatar here:
In this one, there is a slight customization between the Rthro and Blocky body types:
Roblox’s dark angel
“Look over there!”
I believe I Can Fly
The consistency of multiple runs, random initialization, and k-selected clusters suggests that avatars naturally belong to different (albeit ambiguous) categories. At the extreme of the silhouette, we have the old-fashioned, square “Blocky” character, as opposed to the tall, thin, and more realistic “Rthro” avatar. We also found some default avatars, which have not been edited by users since joining Roblox (cluster 4 above). In between, everything from “goth ninja” to “go clubbing”.
Recognition by avatar
How are these aesthetic clusters related to our users themselves?
The simplest starting point is user behavior on the platform.When plotting the last month’s avatar edit, account age (in weeks), total seconds of game time, and one-month retention rate (engagement indicator) by cluster, we will see four charts. These charts Shows dramatic Differences between clusters. Users with highly customized avatars tend to have the highest engagement and retention rates. And avatars that have not undergone a lot of customization are often not involved.
There are two opposite causal explanations for this. One is that users who edit avatars therefore prefer Roblox. Another possibility is that over time, users who have already invested in Roblox tend to devote more energy to their avatars. The rest of Roblox did a lot of great work Decide which explanation to believe.
Regardless of the causality, we see that the two aspects of platform identity—aesthetic performance and participation—are closely intertwined. But what about off-platform identities? How does the real identity of our users-age, geography, gender, etc.-intersect with their Roblox identity? Check out Part 2 of this blog post to find out!
Nameer Hirschkind is a data science engineer at Roblox. He works in the Avatar store to ensure its economy is healthy and prosperous. Neither Roblox Corporation nor this blog endorse or endorse any company or service. In addition, no guarantee or promise is made for the accuracy, reliability or completeness of the information contained in this blog.
©2021 Roblox Corporation. Roblox, the Roblox logo, and Powering Imagination are our registered and unregistered trademarks in the United States and other countries.