Sampling-Bias-Corrected Neural Modeling for Large Corpus Item Recommendations

根据他们的实验表明：
batch softmax is subject to sampling bias and could severely restrict the model performance without any correction, therefore, they correct sampling bias of batch softmax using estimated item frequency.
所以他们提出了一个新的算法去使用梯度下降估计物品的频率。
论文的主要贡献：
流式频率估计

继续填坑


update:


Nearest neighbor search:
maximum inner product search, a temperature is added to each logit to sharpen the predictions, namely
model