{
  "cells": [
    {
      "cell_type": "markdown",
      "metadata": {},
      "source": [
        "\n# Topic extraction with Non-negative Matrix Factorization and Latent Dirichlet Allocation\n\nThis is an example of applying :class:`~sklearn.decomposition.NMF` and\n:class:`~sklearn.decomposition.LatentDirichletAllocation` on a corpus\nof documents and extract additive models of the topic structure of the\ncorpus.  The output is a plot of topics, each represented as bar plot\nusing top few words based on weights.\n\nNon-negative Matrix Factorization is applied with two different objective\nfunctions: the Frobenius norm, and the generalized Kullback-Leibler divergence.\nThe latter is equivalent to Probabilistic Latent Semantic Indexing.\n\nThe default parameters (n_samples / n_features / n_components) should make\nthe example runnable in a couple of tens of seconds. You can try to\nincrease the dimensions of the problem, but be aware that the time\ncomplexity is polynomial in NMF. In LDA, the time complexity is\nproportional to (n_samples * iterations).\n"
      ]
    },
    {
      "cell_type": "code",
      "execution_count": null,
      "metadata": {
        "collapsed": false
      },
      "outputs": [],
      "source": [
        "# Authors: The scikit-learn developers\n# SPDX-License-Identifier: BSD-3-Clause\n\nfrom time import time\n\nimport matplotlib.pyplot as plt\n\nfrom sklearn.datasets import fetch_20newsgroups\nfrom sklearn.decomposition import NMF, LatentDirichletAllocation, MiniBatchNMF\nfrom sklearn.feature_extraction.text import CountVectorizer, TfidfVectorizer\n\nn_samples = 2000\nn_features = 1000\nn_components = 10\nn_top_words = 20\nbatch_size = 128\ninit = \"nndsvda\"\n\n\ndef plot_top_words(model, feature_names, n_top_words, title):\n    fig, axes = plt.subplots(2, 5, figsize=(30, 15), sharex=True)\n    axes = axes.flatten()\n    for topic_idx, topic in enumerate(model.components_):\n        top_features_ind = topic.argsort()[-n_top_words:]\n        top_features = feature_names[top_features_ind]\n        weights = topic[top_features_ind]\n\n        ax = axes[topic_idx]\n        ax.barh(top_features, weights, height=0.7)\n        ax.set_title(f\"Topic {topic_idx +1}\", fontdict={\"fontsize\": 30})\n        ax.tick_params(axis=\"both\", which=\"major\", labelsize=20)\n        for i in \"top right left\".split():\n            ax.spines[i].set_visible(False)\n        fig.suptitle(title, fontsize=40)\n\n    plt.subplots_adjust(top=0.90, bottom=0.05, wspace=0.90, hspace=0.3)\n    plt.show()\n\n\n# Load the 20 newsgroups dataset and vectorize it. We use a few heuristics\n# to filter out useless terms early on: the posts are stripped of headers,\n# footers and quoted replies, and common English words, words occurring in\n# only one document or in at least 95% of the documents are removed.\n\nprint(\"Loading dataset...\")\nt0 = time()\ndata, _ = fetch_20newsgroups(\n    shuffle=True,\n    random_state=1,\n    remove=(\"headers\", \"footers\", \"quotes\"),\n    return_X_y=True,\n)\ndata_samples = data[:n_samples]\nprint(\"done in %0.3fs.\" % (time() - t0))\n\n# Use tf-idf features for NMF.\nprint(\"Extracting tf-idf features for NMF...\")\ntfidf_vectorizer = TfidfVectorizer(\n    max_df=0.95, min_df=2, max_features=n_features, stop_words=\"english\"\n)\nt0 = time()\ntfidf = tfidf_vectorizer.fit_transform(data_samples)\nprint(\"done in %0.3fs.\" % (time() - t0))\n\n# Use tf (raw term count) features for LDA.\nprint(\"Extracting tf features for LDA...\")\ntf_vectorizer = CountVectorizer(\n    max_df=0.95, min_df=2, max_features=n_features, stop_words=\"english\"\n)\nt0 = time()\ntf = tf_vectorizer.fit_transform(data_samples)\nprint(\"done in %0.3fs.\" % (time() - t0))\nprint()\n\n# Fit the NMF model\nprint(\n    \"Fitting the NMF model (Frobenius norm) with tf-idf features, \"\n    \"n_samples=%d and n_features=%d...\" % (n_samples, n_features)\n)\nt0 = time()\nnmf = NMF(\n    n_components=n_components,\n    random_state=1,\n    init=init,\n    beta_loss=\"frobenius\",\n    alpha_W=0.00005,\n    alpha_H=0.00005,\n    l1_ratio=1,\n).fit(tfidf)\nprint(\"done in %0.3fs.\" % (time() - t0))\n\n\ntfidf_feature_names = tfidf_vectorizer.get_feature_names_out()\nplot_top_words(\n    nmf, tfidf_feature_names, n_top_words, \"Topics in NMF model (Frobenius norm)\"\n)\n\n# Fit the NMF model\nprint(\n    \"\\n\" * 2,\n    \"Fitting the NMF model (generalized Kullback-Leibler \"\n    \"divergence) with tf-idf features, n_samples=%d and n_features=%d...\"\n    % (n_samples, n_features),\n)\nt0 = time()\nnmf = NMF(\n    n_components=n_components,\n    random_state=1,\n    init=init,\n    beta_loss=\"kullback-leibler\",\n    solver=\"mu\",\n    max_iter=1000,\n    alpha_W=0.00005,\n    alpha_H=0.00005,\n    l1_ratio=0.5,\n).fit(tfidf)\nprint(\"done in %0.3fs.\" % (time() - t0))\n\ntfidf_feature_names = tfidf_vectorizer.get_feature_names_out()\nplot_top_words(\n    nmf,\n    tfidf_feature_names,\n    n_top_words,\n    \"Topics in NMF model (generalized Kullback-Leibler divergence)\",\n)\n\n# Fit the MiniBatchNMF model\nprint(\n    \"\\n\" * 2,\n    \"Fitting the MiniBatchNMF model (Frobenius norm) with tf-idf \"\n    \"features, n_samples=%d and n_features=%d, batch_size=%d...\"\n    % (n_samples, n_features, batch_size),\n)\nt0 = time()\nmbnmf = MiniBatchNMF(\n    n_components=n_components,\n    random_state=1,\n    batch_size=batch_size,\n    init=init,\n    beta_loss=\"frobenius\",\n    alpha_W=0.00005,\n    alpha_H=0.00005,\n    l1_ratio=0.5,\n).fit(tfidf)\nprint(\"done in %0.3fs.\" % (time() - t0))\n\n\ntfidf_feature_names = tfidf_vectorizer.get_feature_names_out()\nplot_top_words(\n    mbnmf,\n    tfidf_feature_names,\n    n_top_words,\n    \"Topics in MiniBatchNMF model (Frobenius norm)\",\n)\n\n# Fit the MiniBatchNMF model\nprint(\n    \"\\n\" * 2,\n    \"Fitting the MiniBatchNMF model (generalized Kullback-Leibler \"\n    \"divergence) with tf-idf features, n_samples=%d and n_features=%d, \"\n    \"batch_size=%d...\" % (n_samples, n_features, batch_size),\n)\nt0 = time()\nmbnmf = MiniBatchNMF(\n    n_components=n_components,\n    random_state=1,\n    batch_size=batch_size,\n    init=init,\n    beta_loss=\"kullback-leibler\",\n    alpha_W=0.00005,\n    alpha_H=0.00005,\n    l1_ratio=0.5,\n).fit(tfidf)\nprint(\"done in %0.3fs.\" % (time() - t0))\n\ntfidf_feature_names = tfidf_vectorizer.get_feature_names_out()\nplot_top_words(\n    mbnmf,\n    tfidf_feature_names,\n    n_top_words,\n    \"Topics in MiniBatchNMF model (generalized Kullback-Leibler divergence)\",\n)\n\nprint(\n    \"\\n\" * 2,\n    \"Fitting LDA models with tf features, n_samples=%d and n_features=%d...\"\n    % (n_samples, n_features),\n)\nlda = LatentDirichletAllocation(\n    n_components=n_components,\n    max_iter=5,\n    learning_method=\"online\",\n    learning_offset=50.0,\n    random_state=0,\n)\nt0 = time()\nlda.fit(tf)\nprint(\"done in %0.3fs.\" % (time() - t0))\n\ntf_feature_names = tf_vectorizer.get_feature_names_out()\nplot_top_words(lda, tf_feature_names, n_top_words, \"Topics in LDA model\")"
      ]
    }
  ],
  "metadata": {
    "kernelspec": {
      "display_name": "Python 3",
      "language": "python",
      "name": "python3"
    },
    "language_info": {
      "codemirror_mode": {
        "name": "ipython",
        "version": 3
      },
      "file_extension": ".py",
      "mimetype": "text/x-python",
      "name": "python",
      "nbconvert_exporter": "python",
      "pygments_lexer": "ipython3",
      "version": "3.9.21"
    }
  },
  "nbformat": 4,
  "nbformat_minor": 0
}