As the Netflix competition has proved, the matrix decomposition model is superior to the traditional nearest neighbor technology in generating product recommendations, allowing the combination of other information, such as implicit feedback, time effect and confidence level. This paper mainly summarizes the existing matrix decomposition technology in recommendation system.
1. This paper introduces the principle of content-based recommendation system, and gives an example of music website Pandora Com.
2. This paper introduces collaborative filtering algorithms and divides them into neighbor methods and late factor models. The principle of matrix decomposition is introduced.
3. One advantage of matrix decomposition is that it allows the merging of additional information. When explicit feedback cannot be obtained, the recommendation system can infer user preferences using implicit feedback, which indirectly reflects opinions by observing user behavior (including purchase history, browsing history, search mode and even mouse movement). Implicit feedback usually represents the presence or absence of events, so it is usually represented by a dense matrix.
4. Basic MF, bias MF, MF embedded with additional information, timing dynamic MF, MF considering implicit feedback
5. Netflix recommendation contest
Matrix decomposition technology has become the main method of collaborative filtering recommendation. The experience of data sets such as Netflix shows that they provide better accuracy than the classical nearest neighbor technology. At the same time, they provide a compact memory efficiency model, and the system can learn relatively easily. What makes these technologies more convenient is that the model can naturally integrate many key aspects of data, such as various forms of feedback, time dynamics and confidence level.
This work adoptsCC agreement, reprint must indicate the author and the link to this article