Does the 'Not good!' Button make sense?



In today's world, where there is so much content that it can't be consumed even if it takes a lifetime, the recommendation function that recommends the right content for each user is active everywhere. Especially in music streaming services such as Spotify, it is an important function that users will be separated depending on the superiority or inferiority of recommendations. There is no low rating button in such Spotify, but researchers at Cornell University have published a paper on how much the accuracy will differ between the case of 'high rating button only' and the case of 'high rating button & low rating button'. Did.

Evaluating Music Recommendations with Binary Feedback for Multiple Stakeholders by Sasha Stoikov, Hongyi Wen :: SSRN

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=3919046

'Dislike' button would improve Spotify's recommendations | Cornell Chronicle
https://news.cornell.edu/stories/2021/09/dislike-button-would-improve-spotifys-recommendations


The experiment was conducted in the form of letting users listen to music selected from 5 million songs and have them evaluate it. According to Sasha Stoykov, who led the experiment, the first 3 seconds of the song could not be rated, 3 seconds after starting to listen to the song, 'low rating' could be selected, and 6 seconds later 'high rating', 12 By making it possible to select the 'ultra-high rating' option after a second, he said that he was careful so that the high rating could be selected only when the user really liked it.

Based on the 400,000 data collected in this way, when the algorithm is trained only with the data of 'high evaluation & no evaluation' excluding the low evaluation data, and when the data without evaluation is excluded, 'high evaluation & low evaluation' We investigated the difference in accuracy when training with only the data. Then, it is said that songs recommended by the algorithm trained with 'high evaluation & low evaluation' data are 20% more likely to receive high evaluation than songs recommended by the algorithm trained with 'high evaluation & no evaluation' data. It turned out that.



In addition, when Mr. Stoykov investigated the relationship with name recognition in the experiment, it was found that lesser-known artists are less likely to receive high praise and are less likely to be recommended by many people on platforms such as Spotify. In addition to the fact that lesser-known artists are less likely to get high marks, many algorithms are trained to focus on which song they listened to rather than the user's rating, so they are already well-known and listened to by many. Stoykov says that only the artists who are known tend to be recommended.

According to Stoykov, the research team's ultimate goal is to work with record labels to 'determine how likely users are to like a song before it's delivered to a larger platform.' .. Future research will focus on other distribution platforms such as Netflix, incorporating more advanced algorithms, and modeling how the goals of stakeholders such as platforms and record companies are linked to the goals of the algorithm. States.

in Note, Posted by log1d_ts