What is the labels for SVM classification when we firstly run LDA (lda->SVM)

Question

I am using LDA (Latent Dirichlet Allocation) to extract topics. I want to do topic modelling and use the topics as features to do document classification. the reason for doing classification is to evaluate my LDA model.

the same as this link lda , my question is that when I generated my topics, what should be the labels for the classification method?

I was supposed it has to be the topics generated by the LDA, would you please provide me with your idea, suppose we have six topics generated:

topi0: computer, internet, net, technology topic1: politicians, political, barack, topic2: music, hobby,... topic3:... topic4: topic5: ... topic6: ...

so what is the labels for the classification part?

sv=SVC()
labels = [???]

sv.fit(lda_x,labels)
predictclass = sv.predict(lda_x)

testLables=[???]
from sklearn import metrics
yacc=metrics.accuracy_score(testLables,predictclass)

Update

my dataset is 20news group, my aim is to use LDA to extract for example 6 topic(also each topic has its own distribution of terms) then use one classification method to evaluate that.

Update 2 code part this is the output for the matric to send to classification method. I now think there is something wrong with that,

[[ 9.29756948e-07 9.99991632e-01 9.29752853e-07 9.29718960e-07 9.29696185e-07 9.29741802e-07 9.29735986e-07 9.29749709e-07 9.29697295e-07 9.29716150e-07]]

from __future__ import print_function

import nltk
import numpy as np

from time import time

import gensim
from gensim import corpora
from sklearn.feature_extraction.text import TfidfVectorizer, CountVectorizer
from sklearn.decomposition import NMF, LatentDirichletAllocation
from sklearn.datasets import fetch_20newsgroups
from nltk.tokenize import RegexpTokenizer, sent_tokenize, word_tokenize
from nltk.stem.porter import PorterStemmer

from sklearn.naive_bayes import MultinomialNB
from sklearn.preprocessing import LabelEncoder

n_samples = 2000
n_features = 1000
n_topics = 10
n_top_words = 20


def print_top_words(model, feature_names, n_top_words):
    for topic_idx, topic in enumerate(model.components_):
        print("Topic #%d:" % topic_idx)
        print(" ".join([feature_names[i]
                        for i in topic.argsort()[:-n_top_words - 1:-1]]))
    print()


# Load the 20 newsgroups dataset and vectorize it. We use a few heuristics
# to filter out useless terms early on: the posts are stripped of headers,
# footers and quoted replies, and common English words, words occurring in
# only one document or in at least 95% of the documents are removed.

texts=[]
print("Loading dataset...")
t0 = time()
dataset = fetch_20newsgroups(shuffle=True, random_state=1,
                             remove=('headers', 'footers', 'quotes'))
data_samples = dataset.data[:n_samples]
print("done in %0.3fs." % (time() - t0))

# Use tf-idf features for NMF.
print("Extracting tf-idf features for NMF...")
tfidf_vectorizer = TfidfVectorizer(max_df=0.95, min_df=2,
                                   max_features=n_features,
                                   stop_words='english')
t0 = time()

tfidf = tfidf_vectorizer.fit_transform(data_samples)
p_stemmer = PorterStemmer()

str=''.join(data_samples)
sent = sent_tokenize(str)
str2=''.join(sent)
words = word_tokenize(str2)
filtered_sentence = [w for w in words]
stemmed_tokens = [p_stemmer.stem(i) for i in filtered_sentence]
# add tokens to list
texts.append(stemmed_tokens)

values = ' '.join(' '.join(elems) for elems in texts).lower()
print (values)

print("done in %0.3fs." % (time() - t0))

# Use tf (raw term count) features for LDA.
print("Extracting tf features for LDA...")
tf_vectorizer = CountVectorizer(max_df=3, min_df=0.02,
                                max_features=n_features,
                                stop_words='english')
t0 = time()
y=[values]
tf = tf_vectorizer.fit_transform(y)
print("done in %0.3fs." % (time() - t0))

# Fit the NMF model
print("Fitting the NMF model with tf-idf features, "
      "n_samples=%d and n_features=%d..."
      % (n_samples, n_features))
t0 = time()
nmf = NMF(n_components=n_topics, random_state=1,
          alpha=.1, l1_ratio=.5).fit(tfidf)
print("done in %0.3fs." % (time() - t0))

print("\nTopics in NMF model:")
tfidf_feature_names = tfidf_vectorizer.get_feature_names()
print_top_words(nmf, tfidf_feature_names, n_top_words)

print("Fitting LDA models with tf features, "
      "n_samples=%d and n_features=%d..."
      % (n_samples, n_features))
lda = LatentDirichletAllocation(n_topics=n_topics, max_iter=5,
                                learning_method='online',
                                learning_offset=50.,
                                random_state=0)
t0 = time()
lda_x=lda.fit_transform(tf)
from sklearn.svm import SVC
for i in lda_x:
    print (i)
print("done in %0.3fs." % (time() - t0))
print("\nTopics in LDA model:")
sv=MultinomialNB()
labels = [0]

sv.fit(lda_x,labels)
predictclass = sv.predict(lda_x)

testLables=[0]
from sklearn import metrics
yacc=metrics.accuracy_score(testLables,predictclass)
print (yacc)
n=dataset.target[:20]
print (n)

Answer 1

I assume that you refer to Section 7.2 in the original LDA paper.

In that case, as the paper suggests in the first paragraph of Section 7.2, you must have a discriminative framework; namely, a dataset with labeled documents. This is the only way to train an SVM classifier.

From you question, I infer that you use LDA in order to extract topics for a corpus with unknown topics and that you don't have any labeled data. Therefore, you simply cannot build an SVM classifier with what you have.

Answer 2

LDA topic assignments are informative as an exploratory unsupervised exercise but they are way too imperfect to use directly as source of labels for SVM. The problem is that SVM expects one or more distinct labels that align with your conceptual understanding of the subject matter while LDA provides a statistical but thoughtless mixture of topic labels based on word co-occurrences and distributions. The most you can do is use the LDA labels as a starting point to assist you in manually creating a training dataset for SVM.

Answer 3

Still not clear your question. But assume that we work on classification problem of 20newsgroups dataset, we can get:

data_samples = dataset.data[:n_samples]
labels = dataset.target[:n_samples]

data_samples contain list of document, labels contain list of category that each document in data_samples belong.

As your code, we get document-topic_distribution matrix lda_x by code:

tf_vectorizer = CountVectorizer(max_features=n_features,
                                stop_words='english')
tf = tf_vectorizer.fit_transform(data_samples)

lda = LatentDirichletAllocation(n_topics=n_topics, max_iter=5,
                                learning_method='online',
                                learning_offset=50.,
                                random_state=0)

lda_x=lda.fit_transform(tf) 

Example: lda_x[0] is a topic distribution of data_samples[0] (the probabilities document data_samples[0] belong to topics)

Now we use topic-distribution as the features to predict the category of document. It's seem that we apply any classifier to the lda_x matrix:

sv=MultinomialNB()
sv.fit(lda_x,labels)
predictclass = sv.predict(lda_x)
yacc=metrics.accuracy_score(labels,predictclass)
print (yacc)