Of course, the main background is neural networks and machine learning algorithms. As I have told you, my previous hands-on experience is protected by NDA. If you want to have a look at it, I can share it via a personal message. The company is CVisionLab («СиВижинЛаб»), you may call them for a confirmation.
This agreement is in the Russian language as it is concluded with the Russian company. To check it, it’ll be enough to contact them at the provided corporate details. If someone is really interested, bitcointalk has a significant Russian-speaking community, you may request assistance from some of the authority figures there.
For 8 months of work, a number of developments has been done at uKit. I’m afraid, there are very few areas where fast results are possible with machine learning. Speaking about something you can see in action and “touch”, here https://aesthetics-demo.ukit.space/
At the moment, https://aesthetics-demo.ukit.space/ looks more like a toy that makes it possible to evaluate your or another person’s website. But we meant it this way so that people had something to ‘touch’. Also, through gathering opinions of a wide variety of people and offering them to correct our ratings we are able to collect a bigger data set.
Basically, this is the scoring system for uKit AI and is a subsequently important, yet supplementary part of the product.
This system is built using machine learning methods. It should reflect an opinion of an average user about the website’s appearance. It’s neither about a professional designer’s opinion, nor about the website’s usability, nor about its functionality. Let me emphasize this point — it’s only about the APPEARANCE and only an AVERAGE user.
In this project, we are using one of the variants of the gradient boosting method (catboost library). To date, this approach is one of the most effective in solving problems of machine learning.
Besides, in the near future (this month), we are planning to introduce an additional evaluation by the neural network, which should improve the overall quality.
To train the model, we have gathered a sample of 12K websites. Despite the fact that at the moment the system gives reasonably good results (it is possible to play around with it on your own), it is still the first alpha version and has a number of issues:
- it can’t evaluate minimalistic websites such as google.com (this is a solvable issue, but we’d like to achieve this through machine learning rather than using “white lists”);
- it can’t work correctly with hieroglyphic websites (Chinese, Japanese…), since there were none of this kind in the existing training base;
- it has an insufficient ratings bank. We need more people and different opinions to improve the quality. It doesn’t make the system unrepresentative at the moment, but the data set needs significant enhancements, which will certainly influence the quality and robustness of the output results.