GCN thinking - word2vec directly calculates text classification

Optimize one ：

reference GCN Thought , In the process of training, there is an edge between words （ For example, the interval between the same sentence is no more than 5 Two words of ）, So in word2vec in , Words are used to predict words , such as CBOW in 4 A word predicts an intermediate word , Then the same model can be used in training CBOW When predicting, predict an article at the same time .

At this moment huffman code Is by word 、 Article categories Made up of .

 

Optimization II ：

Conventional text-gcn It's semi supervised learning , Every doc and word It's all one node, The adjacent 5 individual word There's always edge, doc And all that it contains word There is also a edge, In fact, it is the clustering work , That is, unknown kinds doc Also as a node Put in the picture , This cannot be used to estimate the new samples , For example, in the online prediction scenario , It can't be used . 

Then can you doc Don't put it in the figure , And just word-word Create a graph , This is equivalent to GCN+Fasttext or GCN + CNN To predict .