[WIP] Sparse Composite Document Vectors

scdv.py

import numpy as np
from sklearn.feature_extraction.text import TfidfVectorizer
from sklearn.mixture import GaussianMixture

class SCDV(object):
    def __init__(self, word_embedding_model):
        self._word_embedding_model = word_embedding_model
        self._tfidf = TfidfVectorizer(use_idf=True)
        self._gmm = GaussianMixture()
        
    def fit_transform(self, X):
        # D: number of documents
        # W: number of vocablary
        # V: vector size
        # C: number of clusters
        idf = self._tfidf.fit(X).idf_  # shape: D x W

        vocabs = self._tfidf.get_feature_names()  # shape: W x V
        vecs = np.vstack([word_embedding_model.get(v) for v in vocabs])

        self._gmm.fit(vecs)
        proba = self._gmm.predict_proba(vecs)  # shape: W x C
        
        wcv = np.einsum('ij,ik->ijk', proba vocabs)  # shape: W x C x V

        # shape: {matrix that shape is (D x W)} \dots {matrix that shape is (W x CV)} = D x CV
        docvecs = idf @ np.stack(wcv, axis=1)
        return docvecs

ref: https://qiita.com/fufufukakaka/items/a7316273908a7c400868