Vbpr (visual Bayesian personalized ranking) paper summary

One 、 lead

         This paper presents an extensible factorization model , This model can integrate the visual signal into the predictor . This paper uses deep network （CNN） Extract visual features from product images , And then tap people's feedback . This approach can not only make the model More accurate , And it can Alleviate cold start problems .

Two 、 There are models ——MF（ Matrix decomposition ）

1. The following is the basic prediction formula ：

 2. Symbol description ：

3. The problem is ：

        Due to the real-world dataset sparsity , They still face Cold start The problem of .

4. Solutions ：

         Divide the rating dimension into Visual factors and potential （ Non visual ） factors . Use explicit features to alleviate this problem .

3、 ... and 、VBPR： Visual Bayesian personalized ranking

1. Picture features pass CNN To extract , Combine image features with potential features for recommendation . As shown in the figure below ：

 

2. Preference predictor model ：

① Model ：

② Symbol description ：

 

3. Model implementation ：

Learn one Embedded kernel , Linear transformation of high-dimensional features into a lower dimensional feature （ such as 20 about ）“ Visual rating space ”

① Embedded kernel formula ：

② Symbol description ：

 

③ evaluation ：

This embedding is effective , Because all projects share the same embedded matrix , This significantly reduces the number of parameters to learn

5.VBPR Model formula ：

① The formula ：

 ② explain ：

β'： Visual deviation , And fi The inner product of is equivalent to the user's overall view of the visual appearance of a given item .

Four 、 Use BPR Model learning

Bayesian personalized ordering (BPR) It's a kind of adoption Random gradient rise As part of the training process Pairwise sorting optimization framework .

1. Personalized sorting （BPR-OPT）

2.  When using matrix factorization as a preference predictor （BPR-MF）, Defined as

 3. Study BPR-MF( Update parameters ) Methods

 

η： Is the rate of learning

5、 ... and 、 Method of updating parameters

1. Non visual parameters ：

In order to BPR-MF Update in the same form

2. Visual parameters

5. experiment ：

1. Data sets ：Amazon.com、Tradesy.com

Data set statistics after preprocessing  

2. The baseline ：

Random (RAND)	This baseline will be applied to all users' projects Stochastic ranking .
Most Popular (MP)	This baseline is based on the Popularity Sort the projects , And it is not personalized
MM-MF	in pairs MF Model , The model is right xuij Hinge on Ranking loss is optimized , And like in BPR-MF Use... In the same way SGA Conduct Training .
BPR-MF	Implicit feedback Data sets Personalized ranking
Image-based Recommendation (IBR)（ Image based recommendation services ）	It learned a Visual space , And retrieve and query images Projects with similar styles . Then in the learned visual space through Nearest neighbor search To predict .

3. Experimental results and conclusions

①

Average of different data sets AUC result （ share 20 A factor ）

Conclusion ：

VBPR Better than all baselines in most cases .

stay BPR-MF On the basis of ,VBPR For all projects BPR-MF The average increase is more than 12%, Improve cold start by more than 28%. take CNN function Inclusion in our ranking task has significant benefits .

② Sensitivity

Under different dimensions AUC

  Conclusion ：

         As the number of factors increases ,MM-MF、BPR-MF and VBPR Perform better , This shows the ability of pairwise method to avoid over fitting .

③ Training efficiency

Increasing the number of training iterations AUC（ On the test set ）

Conclusion ：

         Our proposed model MM-MF and BPR-MF Need a longer convergence time , Although still only about 3.5 Hours of training in our largest data set （ Women's wear ） Upper convergence .