Analysis of Badcase in personalized recommendation system

For the beta user feedback, because the previous day hit a few animations, resulting in the next days recommended animation screen accounted for a relatively high, so began to analyze this badcase.

First, the user's history view record is analyzed, due to system upgrade, log defect problem, caused the user between 10.15-11.3 log is not collected, but this is the previous hypothesis is not established phenomenon, because no new data collected, the user portrait is not possible to reflect the day before the animation, animation and Disney These have the emphasis on the portrait of the word.

First, the problem is analyzed, the problem may occur at the time interval, because we quoted the time interval as the user's preference for a movie preferences weight calculation, resulting in this problem may have two reasons:

1. Over time, some of the words on the user's portrait appear small fluctuations, but these small fluctuations can lead to unpredictable consequences, such as: The portrait of the user before the problem occurs may be:

Person Ain Belly: 2.694, Ivan Adal: 1.992, Sophie Marceau: 1.897 label Sophie Marceau: 0.451, Jean Reno: 0.361, Getaway: 0.271, you shut up: 0.271,sophiemarceau : 0.226,jeanreno:0.226, ambush on eight sides: 0.203, France: 0.169, French movie: 0.155, Suspense: 0.130

But after the passage of time some weights begin to appear as follows:

Person Alain Ritchison: 3.979, Rai Hongyu: 3.979, Guo Duojing: 3.678 label Disney: 1.186, Animation: 1.129, Romance: 1.119, Classic: 1.028,3d:1.010, Fantasy: 0.940, Country diffuse: 0.900 , Effect good: 0.800, Jurassic Park: 0.800, Spielberg: 0.775

Analysis found that the person in the three actors belong to the unpopular actor, so the film will not be too much, label label Disney, Animation, the two have a great representation, they cover the animation is very much, resulting in the recommendation is the animation inside, so the effect is not good, later can do some processing, but the problem came, Is it the time that causes the situation? Or did the user really look at the behavior of many animated movies caused by the movie? Continue to find the reason through the log.

Analysis of Badcase in personalized recommendation system