VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback

Preference Predictor

预测模型如下：

图片特征通过CNN进行提取，将图片特征与潜在的特征联合在一起进行推荐

###Model Learning Using BPR

The following optimization criterion is used for personalized ranking (BPR-OPT):

When using Matrix Factorization as the preference predictor (i.e., BPR-MF), $\hat{x}_{uij}$x^uij​is defined as:

使用SGA（stochastic gradient ascent）进行参数更新：