V2

[kushaank]

• Use auto encoder
• Hards part to represent: words, characters, tags (variable in number)
• Way to encode these into a fixed length vector and then you're done. Fixed length representation -> then use auto encoder
• Have a super sparse vector: with words/chars and tags
• Huge length vector for user tag, user words, user characters, question tag, question words, queastion charaacters
• u1 and q1 will have ID, word IDs, char IDs, tags etc (super sparse vector)
• put u1 and q1 into auto encoder
• Turn the vector into a smaller dimension representation and then reconstruct it
• User auto encoder and question auto encoder
• then train user question vectors (encoded) through third neural net
• Don't use hand tailored features for now
• Tune the neural network well: momentum, restart etc
• minimize diff between input and output of autoencoder and decoder simultaneously
• TODO: exclude ID field of the auto encoder, in the end append it to the smaller vector we get