NQTM.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "id": "c4296f4d-1674-4b7b-9b27-83b7e762d4a2",
   "metadata": {},
   "outputs": [],
   "source": [
    "import numpy as np\n",
    "import pandas as pd\n",
    "from sklearn.feature_extraction.text import CountVectorizer\n",
    "from sklearn.preprocessing import normalize"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "id": "7589fa7f-5e46-49a1-92bd-e09ce05f751b",
   "metadata": {},
   "outputs": [],
   "source": [
    "class NQTM:\n",
    "    def __init__(self, n_topics, n_samples, max_iter=100, tol=1e-4):\n",
    "        self.n_topics = n_topics\n",
    "        self.n_samples = n_samples\n",
    "        self.max_iter = max_iter\n",
    "        self.tol = tol\n",
    "        self.Q = None\n",
    "\n",
    "    def _negative_sampling(self, P):\n",
    "        neg_P = np.zeros_like(P)\n",
    "        for t in range(self.n_topics):\n",
    "            sorted_indices = np.argsort(-P[:, t])\n",
    "            neg_indices = sorted_indices[self.n_samples:]\n",
    "            neg_P[neg_indices, t] = P[neg_indices, t]\n",
    "        return neg_P\n",
    "\n",
    "    def _quantization(self, P):\n",
    "        return (P > 0).astype(float)\n",
    "\n",
    "    def fit_transform(self, X):\n",
    "        n_docs, n_words = X.shape\n",
    "        \n",
    "        # Inicializar a matriz P e a matriz Q.\n",
    "        P = np.random.rand(n_docs, self.n_topics)\n",
    "        P /= P.sum(axis=1, keepdims=True)\n",
    "\n",
    "        self.Q = np.random.rand(self.n_topics, n_words)\n",
    "        self.Q /= self.Q.sum(axis=1, keepdims=True)\n",
    "\n",
    "\n",
    "        for _ in range(self.max_iter):\n",
    "            P_old = P.copy()\n",
    "\n",
    "            # E-step\n",
    "            P = X @ self.Q.T\n",
    "            P = normalize(P, axis=1, norm='l1')\n",
    "\n",
    "            # M-step\n",
    "            neg_P = self._negative_sampling(P)\n",
    "            self.Q = normalize((X.T @ (P - neg_P)).T, axis=1, norm='l1')\n",
    "            self.Q = self._quantization(self.Q)\n",
    "\n",
    "            if np.linalg.norm(P - P_old) < self.tol:\n",
    "                break\n",
    "\n",
    "        return P"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "83de4e4b-6391-46f3-bb3b-8116d620b816",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Leitura dos dados do dataframe\n",
    "#df = pd.read_csv('datasets/(processado-final)textos_tuitesPt_2020.csv.gz', names=['texto'])\n",
    "df = pd.read_csv('datasets/(processado)textos_tuitesPt_2020_0.csv', names=['texto'])\n",
    "\n",
    "# Elimina um valor flutuante que aparece no dataframe (por razões misteriosas)\n",
    "# o algoritmo não aceita o valor flutuante, que precisa ser filtrado\n",
    "df = df[df['texto'].apply(lambda x: isinstance(x, str))]\n",
    "df['texto'].apply(type).value_counts()\n",
    "\n",
    "df"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "id": "825b8a45-09e1-4f17-b8ca-50362e08e9ec",
   "metadata": {},
   "outputs": [],
   "source": [
    "documents = df[\"texto\"].tolist()"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 5,
   "id": "6d0216ba-0286-4a6b-bd3b-677589320a42",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Vetorizar os documentos\n",
    "vectorizer = CountVectorizer(max_df=0.95, min_df=2, stop_words=None)\n",
    "X = vectorizer.fit_transform(documents)"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "ba34bc99-1b73-41e5-b8ec-dc720e3f68d1",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Treinar o modelo NQTM\n",
    "n_topics = 10\n",
    "n_samples = 10\n",
    "nqtm = NQTM(n_topics, n_samples)\n",
    "topic_matrix = nqtm.fit_transform(X.toarray())"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "8bfbbf5f-4f2a-4eb7-b959-68858084894d",
   "metadata": {},
   "outputs": [],
   "source": [
    "# Exibir os tópicos\n",
    "topic_word_matrix = nqtm.Q\n",
    "feature_names = vectorizer.get_feature_names_out()\n",
    "for topic_idx, topic in enumerate(topic_word_matrix):\n",
    "    print(f\"Topic #{topic_idx + 1}:\")\n",
    "    print(\" \".join([feature_names[i] for i in topic.argsort()[:-10 - 1:-1]]))"
   ]
  }
 ],
 "metadata": {
  "kernelspec": {
   "display_name": "Python 3 (ipykernel)",
   "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": "3.9.7"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}