machine-learning - RBM for collaborative filtering

Question

My algorithm RBM for collaborative filtering will not converge... The idea of what I think RBM　for collaborative filtering is

1. initial w , b , c and random at [0,1]

2. For By User clamp data -> visible (softmax)

   Hidden = sigmoid(b+W*V)

   Run Gibbs on Hidden -> Hidden_gibbs

   Positive = Hidden*Visible

   Hidden -> reconstruct -> reconstruct_visible

   Run Gibbs on reconstruct_visible -> reconstruct_visible_gibbs

   negative = Hidden_gibbs*reconstruct_visible_gibbs

   End for

3. Update

   w = w + (positive-negative)/Number_User

   b = b + (visible - reconstruct_visible_gibbs)/Number_User

   c = c + (Hidden - Hidden_gibbs)/Number_User

I have seen lots of paper or lecture, and have no idea where is wrong

Answer 1

这不是一个简单的问题！您对学习过程的描述看起来不错。但是，从描述到实际代码都有很大的错误空间。此外，对于 CF，“香草”RBM 将不起作用。

• 你是如何实现可见的“softmax”单元的？

• 您是否按照原始工作 [1] 中的建议使用“单用户”数据集训练您的 RBM？

还有 2 个关于权重更新和预测过程的细节与 vanilla 的 RBM 略有不同

[1] Salakhutdinov http://www.machinelearning.org/proceedings/icml2007/papers/407.pdf