wais5psl
Iron
- Joined
- Mar 10, 2026
- Posts
- 14
- Reputation
- 17
I created the website LOOKSIQ because all existing systems are inherently flawed and viciously circular. Let’s take harmony as an example. Even in studies where ratios are combined with machine learning algorithms they explain only about 60% of the total variance and that’s using a cleaner and more scientific approach than what we typically see here where the scoring system is largely made up (for example 70% harmony= chadlite with no validation).
The modus operandi should not be simply attributing certain ratios to percentiles. Instead one can use ML algorithms like XGBoost to find the relationship between ratios and attractiveness examining both nonlinear and linear relationships. This is done by minimizing a loss function mostly multiclass classification loss by adjusting the weights of each feature in the model. The conclusion is that even near perfectly weighting and combining ratios does not produce an adequate measure of attractiveness.
The alternative I propose is Item Response Theory (IRT) which maps the relationship between the latent variable attractiveness and the likelihood that a person is perceived as more attractive than another person. If enough people vote individual biases cancel out under certain conditions according to the central limit theorem leaving only the true variance being the person’s “true” attractiveness.
Currently I am planning to add features so that users can give advice to others. This advice would then be standardized so that only advice aligned with the consensus of the general population remains ( meaning that most people would recommend you to do X and not only a small biased group of people)
**Check out the site and tell me about features you guys want to see**
**ITS 100% FREE AND IM DOING THIS FOR THE LOVE OF THE GAME**
The modus operandi should not be simply attributing certain ratios to percentiles. Instead one can use ML algorithms like XGBoost to find the relationship between ratios and attractiveness examining both nonlinear and linear relationships. This is done by minimizing a loss function mostly multiclass classification loss by adjusting the weights of each feature in the model. The conclusion is that even near perfectly weighting and combining ratios does not produce an adequate measure of attractiveness.
The alternative I propose is Item Response Theory (IRT) which maps the relationship between the latent variable attractiveness and the likelihood that a person is perceived as more attractive than another person. If enough people vote individual biases cancel out under certain conditions according to the central limit theorem leaving only the true variance being the person’s “true” attractiveness.
Currently I am planning to add features so that users can give advice to others. This advice would then be standardized so that only advice aligned with the consensus of the general population remains ( meaning that most people would recommend you to do X and not only a small biased group of people)
**Check out the site and tell me about features you guys want to see**
**ITS 100% FREE AND IM DOING THIS FOR THE LOVE OF THE GAME**
