LHUC is good for sorting at the last step of a recommendation system. The network was originally from research in speech recognition.
Consider the case of speech recognition, taking the voice signals as an input vector, and the feature of the speaker another input. Let the two inputs pass two separate networks (could be fully connected networks). And perform Hadamard product of the two outcomes. (Obviously the output dimensions need to be the same). Wire combined result into another fully connected layer.
Take the feature of the speaker as input, with another fully connected layer. Perform Hadamard product at the result. This process can repeat many times (by adding fully connected layer at the main path and Hadmard product the input).
Note that the network that transforms the feature of the speaker input can be a multiple fully connected layer with a sigmoid function and multiplies 2. The scale 2 allows features to be exaggerated to adapt to distinct feature of a speaker.
In case of apply LHUC network to a recommendation system, the voice signal is replaced with product features, while the speaker feature is replaced with the user features.
Reference: https://www.youtube.com/watch?v=TxIedW94hu0
No comments:
Post a Comment