Machine learning-based rarity assigning for NFT is giving artists more freedom in their creations and owners' choice with which values they identify.
Rarity assigning using machine learning estimated likability
The standard procedure in assigning rarity for the NFTs comes from creating the trait table and manually assigning the percentages for spawning. The scientific language says that people have different tastes. Choosing the rarest defined traits is not correlated with the fact that generated NFT will also be perceived as likeable. This and the freedom of choosing a variety of traits independently on the basis of manually assigned rarity tables lead us to address this issue using machine learning. Our NFTs will not only have the likeness which is an appealing combination of the individual traits but will also have the corresponding traits added. The distribution will be calculated from the randomly assigned seed for selection of the NFTs combination within the calculated distribution bins when the minting process will finish.
To put it into cat-human language your golden Rolex watches are precious, but they will not always go well with your favourite hoodie. This is why we rate this combination less likeable than Casio G-shock with Gundam hoodie.
The random generation of NFT and the likability distribution
So how are we estimating the likability? After all, we will need some value to assign to NFTs and all of us have different tastes, as the saying goes, beauty is in the eye of the beholder. First, we have generated three times our NFT collection with machine learning-based grading (likability) and compared the results. The features (count) selection were as follows: background (13), skin (25), face (9), body style (71), mouth (21), ears (15), head (29), eyes (24) and the frames representing the likability score. The following layers could be blank (no visual feature added): face, ears, body style and head.
The normal distribution of likability in all three sets was evaluated using the Shapiro-Wilk normality test where generated set 1 and 2 passed and we are confident that distribution is normal, while the set 3 likability is not following a normal distribution.
The individual contribution to the likability divided by the individual features represented by the notched box plots are displayed below:
There is a unity in the most positive and negative features trends for 4 out of 7: background, skin, face and head across all three sets. Out of 199 individual features, 113 are within the significant effect unison.
Our NFT is composed of multiple variables contributing to the overall likability as an outcome variable. Therefore as our data are composed of multiple factors as categorical data with continuous likability variables, we have started exploring whether there is an occurrence of clustering. Clustering would suggest that there is a combination of factors that have similar characteristics. The method of choice for us was a well-established clustering analysis derived from the K-means method, the K-prototypes from the kmodes package. We doubted there would be any observable clusters, and we wanted to confirm it with the algorithm. We have not observed any clustering. This means that contribution of the combination of the individual factors does not have any distinguishable rule.
Grading accuracy across the social spectrum
We have randomly selected 10 generated NFTs from the bottom 5% and top 5% of the likability distribution. Then we let people sort them according to their personal likeliness preference. Since the taste is personal and we have not had time to conduct extensive questionnaires for thousands of people, we simplified our method for comparison between the groups by the accuracy of assigning the NFTs to the upper and lower halves. Which actually failed, since we have had a sample size that was not even close to the normal distribution. But what we wanted to confirm with our experiment was to see whether generation and sex have an influence on likability. The results proved that there is a significant difference using analysis of variance between the groups' males and females of millennials and generation Z with 95% significance level (p-value = 0.27, N=40) The more accurate results were obtained when comparing the Millenials generation likability to the machine learning estimated likability. The major influence of the difference between the millennials and generation z is that the model was trained using data from social networks, where we had not accounted for the generation on the input. For any likability algorithms, this has to be accounted for, to train the models using also accurate socioeconomic data.
Conclusion
Based on our results, we would love our NFTs to be appealing based on individual taste and not be driven by scarcity/likability. The best saying which would apply for our collection is “One man’s trash is another man’s treasure”. Our likability approach is not yet perfect and we would love to improve it with more accurate data extended with the age groups and other demographics / socioeconomic factors. We will dedicate portions of our ETH treasury to the project to describe and develop a better likability ML model which we will opensource for other projects to use.
