improving "Probable Score" predictions

Ideas to improve Criticker and new feature requests, as well as announcements about new enhancements.
andr
Posts: 36
1785 Ratings
Your TCI: na
Joined: Fri Apr 22, 2016 6:35 am

improving "Probable Score" predictions

Post by andr »

I have ranked 471 movies so far; and still "Probable Score" predictions often are not accurate for me. When I rated the best of the same movies on imdb, the imdb recommended me some great films I saw long ago, however Criticker never recommended me those films, though I did not rank them previously on Criticker.

I believe the current algorithm does not always work. Say I liked many action, comedy and romance movies and ranked only a few horror movies, because I do not like horror movies. When calculating Probable Score of a new horror movie it will use my ratings of action, comedy and romance (because there are many of them) and might give a high Probable Score, because many other users who liked action, comedy and romance did not mind horror too. But it would not be the rating I would give.

I wounder if Criticker's algorithm might be improved... My thinking is that instead of the Probable Score of a movie being built on ratings of users, having rated all other movies similarly to me, what if the Probable Score was built on the ratings of users, having similarly rated the movies, SIMILAR TO THE MOVIE BEING RATED. The similarity of movies to each other could be calculated for example by the movie data: genres, film cast and crew, country. Or it can be calculated by the similarity of distributions of ratings across all the users who rated them. Several factors of similarity could be weighted and used together. For the movies on which there was not enough data to apply this algorithm, the current algorithm could be used instead. Or even both the current and the suggested algorithm could be always used together and the resulting score would be a (weighted) average of the both methods.

I believe it is possible to compare the effectiveness of the suggested method with the current one or any other. One would only need to apply the formula of Probable Score to all the existing ratings in the database and calculate the total error of the predictions as the standard deviation of the difference between predicted and actual scores. The algorithm with the least error would be the best.

Post Reply