Autoencoder
It turns out that a well-defined Autoencoder can compete with the best Collaborative Fitering algorithms.
Paper: Strub, F., Mary, J., Gaudel R., (2016). Hybrid Collaborative Filterings and Neural Networks.
自动编码器的设计初衷是通过编码解码得到鲁棒的特征，它的输出目标是重建输入，但是这里，稀疏的向量是为了评分预测，而不是还原输入。
When no rating exist, the error is set to zero as it cannot be computed.
The bottleneck of the autoencoder is often greater than the number of ratings in the input vector! If we change the toys example by a bit more realistic autoencoder. It is not straightfroward that the network will be able to learn a useful representation from a very sparse representation.
The autoencoder is acctually diverted from its original purpose. It does not aim to reconstruct the initial input anymore, it aims at predicting the missing data.
为了防止丢失，使用两种方式上传的图片。
误差 = (5-Error) = (1-Input) - (4-Output)
圈里面有“\”符号的为corrupted