Artscorer 1.0

General information about the Artscorer

Q: What is the Artscorer?

A: The Artscorer is an algorithm. It was designed by Michael Dalvean. It measures the extent to which traditional design, colour selection and compositional features are present in western paintings from the last 500 years.

Q: What paintings were used to calibrate the algorithm?

A: One thousand paintings were used to train the algorithm. The study group (n = 500) was selected from artist’s lists of paintings from the last 500 years judged to be of high technical quality (canonical paintings). The control group ( n = 500) was selected from paintings that are not mentioned in such lists (non-canonical paintings). The algorithm is trained to recognise differences in design, colour selection and compositional elements as between canonical and non-canonical paintings.

Q: How accurate is the algorithm?

A: On the corpus of 1000 paintings, the Artscorer is able to correctly classify 86.5% of paintings. It is able to correctly classify the canonical paintings with an accuracy rate of 89% and non-canonical paintings with an accuracy rate of 84%.

[Technical note: two types of classifier were used in the construction of the algorithm - linear regression and random forest. The score given by the app is the average probability scores for the regression and random forest classifiers less 0.5. and multiplied by 10 to give a score between -5 and +5]

Q: How do I interpret the score?

A: The higherst possible score is 5. The lowest possible score is -5. The average score (the score halfway between the scores for canonical and the non-canonical paintings) is 0. A score above 0 means that the scored painting is more like the canonical paintings than the non-canonical paintings and conversely for scores below 0. )

Q: Will the algorithm score images other than those of paintings?

A: The algorithm was calibrated on paintings. As such, the score generated is valid only for paintings. Thus, if a jpeg of a photo that is not of a painting is scored, the score will not be valid. As an analogy, consider a spellchecker. A spellchecker is designed for a given dialect of a given language. Therefore, if we spellcheck a document written is Scots English using a spellchecker for Modern Standard American English, we would generate a score that suggests that there are many spelling errors when if fact the document might be written in perfect Scots English. Thus, the spellcheck result is simply not valid. Similarly, a score generated by a photo of anything other than a painting is not valid.