Interpretation of the thesis -- factorization meets the neighborhood: a multifaceted collaborative filtering model

brief introduction

This article is about Koren stay KDD08 Conference articles , Its core idea is mainly

1. CF Neighborhood model and hidden factor model can be fused , Get better results
2. Implicit feedback can be introduced into the model , Achieve better results

primary coverage

1 Collaborative filtering CF No domain knowledge is required 、 There is no need to collect large amounts of data , And discover the advantages of some complex and unexpected models , Make it very popular .

2 CF It mainly includes domain model and hidden factor model , Both have advantages and disadvantages . The domain model focuses on the nearest neighbor , So he is good at finding local relationships , But it is difficult to capture all the weak signals in the data set , The implicit factor model can effectively estimate the overall structure of most or all projects , But we can't find a small number of strong correlations between projects .

3 Include purchase history 、 Browse history 、 Search history , Even mouse movement , Can be added to the model as implicit feedback .

4 In terms of models , The author first put forward Baseline Algorithm , Then, the domain model and hidden factor model are introduced .

5 Neighborhood model , The author first introduces and Interpretation of the thesis ——Performance of Recommender Algorithms on Top-N Recommendation Tasks The same domain model in , This model is a classical expression based on domain idea , It can only be calculated directly from the data , Can't adjust , Without considering all the neighbors , And it will lose its function when the user's score is very small , So the author uses an interpolation factor ${\theta }_{ij}^u$ To solve this problem , So the model becomes
$${\hat{r}}_{ui}=b_{ui}+\sum _{j\in S^k(i;u)}{\theta }_{ij}^u(r_{uj}-b_{uj})$$, The model is detailed in .
On this model , Adding implicit feedback from users is the domain model proposed by the author in this paper , The model can be calculated by using optimization model or least square .

6 In hidden factor model , First of all, it is proposed that Asymmetric-SVD Asymmetric SVD, Asymmetry is reflected in the use of item features to represent user features ,
$${\hat{r}}_{ui}=b_{ui}+q_i^T\left(|R(u){|}^{-\frac{1}{2}}\sum _{j\in R(u)}(r_{uj}-b_{uj})x_j+|N(u){|}^{-\frac{1}{2}}\sum _{j\in N(u)}{y}_j\right)$$
common P The vector is represented by some item feature , So it is called asymmetric , The advantage of this model is that there are fewer parameters （ No training required P Matrix ）、 It does not involve user preferences, so it can recommend new users to new users 、 Interpretability （ Explain through the project ） And effective use of implicit feedback .
The common P, That's it SVD++ 了 , You can refer to another blog post Recommendation system surprise Library Tutorial .

7 The author fused the two models , It is the fusion of the last two scoring prediction formulas .