add preproc notebooks

preprocessing/MKB-generate-triplets.ipynb

@@ -0,0 +1,363 @@
+{
+ "cells": [
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from topicnet.cooking_machine import Dataset\n",
+    "import pandas as pd\n",
+    "\n",
+    "dataset = Dataset('MKB10.csv', batch_vectorizer_path=\"./mkb_batches2\")\n",
+    "\n",
+    "data_orig = pd.read_pickle(\"MKB10.pkl\")\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 19,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "data_orig.title = data_orig.title.str.replace(\" \", \"_\")\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "data": {
+      "text/plain": [
+       "{'@letter', '@ngram', '@text'}"
+      ]
+     },
+     "execution_count": 5,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "dataset.get_possible_modalities()"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 6,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "\n",
+    "import numpy as np\n",
+    "\n",
+    "# all articles are from category \"Симптомы по алфавиту\", so threshold is 1\n",
+    "THRESHOLD = 1\n",
+    "\n",
+    "def randomly_select(data_orig, point_a, should_intersect):\n",
+    "    attempts_left = 100\n",
+    "    while attempts_left > 0:\n",
+    "        rnd_idx_b = np.random.choice(data_orig.shape[0])\n",
+    "        point_b = data_orig.iloc[rnd_idx_b]\n",
+    "        intersection = set(point_b.categories) & set(point_a.categories)\n",
+    "        is_big = (len(intersection) > THRESHOLD)\n",
+    "        if should_intersect == is_big:\n",
+    "            # if is_big:\n",
+    "            #    print(point_a.title, point_b.title, intersection)\n",
+    "            return point_b.title, rnd_idx_b, intersection\n",
+    "        attempts_left -= 1\n",
+    "    return None, None, None\n",
+    "\n",
+    "def generate_triplet(data_orig):\n",
+    "\n",
+    "    rnd_idx_a = np.random.choice(data_orig.shape[0])\n",
+    "    point_a = data_orig.iloc[rnd_idx_a]\n",
+    "    point_b, rnd_idx_b, intersection = randomly_select(data_orig, point_a, should_intersect=True)\n",
+    "    point_c, rnd_idx_c, empty_intersection = randomly_select(data_orig, point_a, should_intersect=False)\n",
+    "    # return [point_a, point_b, point_c]\n",
+    "    return [point_a.title, point_b, point_c, intersection]\n",
+    "    # return [rnd_idx_a, rnd_idx_b, rnd_idx_c]\n",
+    "\n",
+    "    \n",
+    "    "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 21,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stderr",
+     "output_type": "stream",
+     "text": [
+      "100%|██████████| 10000/10000 [00:57<00:00, 173.50it/s]\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "7221"
+      ]
+     },
+     "execution_count": 21,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "from tqdm import tqdm\n",
+    "\n",
+    "triplets = []\n",
+    "\n",
+    "for i in tqdm(range(10000)):\n",
+    "    a, b, c, explanation = generate_triplet(data_orig)\n",
+    "    if b is not None and c is not None:\n",
+    "        triplets.append( (a, b, c, explanation) )\n",
+    "\n",
+    "len(triplets)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 22,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import pickle\n",
+    "\n",
+    "with open(\"triplets_mkb.p\", \"wb\") as f:\n",
+    "    pickle.dump(triplets, f)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  },
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "## Scoring model by ranking quality"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 23,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from topicnet.cooking_machine.models import BaseScore as BaseTopicNetScore, TopicModel\n",
+    "\n",
+    "\n",
+    "class ValidationRankingQuality(BaseTopicNetScore):\n",
+    "    def __init__(self, validation_dataset, triplets):\n",
+    "        super().__init__()\n",
+    "\n",
+    "        self.validation_dataset = validation_dataset\n",
+    "        self.triplets = triplets\n",
+    "\n",
+    "    def call(self, model: TopicModel):\n",
+    "        theta = model.get_theta(dataset=self.validation_dataset)\n",
+    "        \n",
+    "        correct_rankings = 0\n",
+    "\n",
+    "        for (a, b, c, _) in self.triplets:\n",
+    "            # L1 distance, just for example\n",
+    "            similar_dist = sum(abs(theta[a] - theta[b]))\n",
+    "            diffrnt_dist = sum(abs(theta[a] - theta[c]))\n",
+    "\n",
+    "            correct_rankings += (similar_dist < diffrnt_dist)\n",
+    "\n",
+    "        return correct_rankings / len(self.triplets)\n",
+    "\n",
+    "        "
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 24,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "import artm\n",
+    "\n",
+    "artm_model = artm.ARTM(\n",
+    "    num_topics=20, \n",
+    "    dictionary=dataset.get_dictionary(),\n",
+    "    class_ids={'@lemmatized': 1, '@ngram': 50},  # absolute values, just for example\n",
+    "    theta_columns_naming=\"title\"\n",
+    ")\n",
+    "\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 25,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "tm = TopicModel(artm_model, custom_scores={\"ranking\": ValidationRankingQuality(dataset, triplets)})"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 26,
+   "metadata": {
+    "scrolled": true
+   },
+   "outputs": [],
+   "source": [
+    "tm._fit(dataset.get_batch_vectorizer(), 10)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 27,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "0.641600886303836\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "[<matplotlib.lines.Line2D at 0x7f133d31e8d0>]"
+      ]
+     },
+     "execution_count": 27,
+     "metadata": {},
+     "output_type": "execute_result"
+    },
+    {
+     "data": {
+      "image/png": "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\n",
+      "text/plain": [
+       "<Figure size 432x288 with 1 Axes>"
+      ]
+     },
+     "metadata": {
+      "needs_background": "light"
+     },
+     "output_type": "display_data"
+    }
+   ],
+   "source": [
+    "import matplotlib.pyplot as plt\n",
+    "%matplotlib inline\n",
+    "\n",
+    "print(tm.scores['ranking'][-1])\n",
+    "plt.plot(tm.scores['ranking'])"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 28,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "theta = artm_model.transform(batch_vectorizer=dataset.get_batch_vectorizer())"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 30,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "Асимптоматический_хронический_простатит Некротический_фасциит Болезнь_Галлервордена_—_Шпатца\n",
+      "{'Синдромы по алфавиту', 'Бактериальные инфекции'}\n",
+      "Листериоз Трихофития Дермацентороз\n",
+      "{'Инфекционные заболевания', 'Синдромы по алфавиту'}\n",
+      "Гестоз Маловодие Психогенное_переедание\n",
+      "{'Патология беременности', 'Синдромы по алфавиту'}\n"
+     ]
+    }
+   ],
+   "source": [
+    "for (a, b, c, explanation) in triplets[:10]:\n",
+    "    # L1 distance, just for example\n",
+    "    similar_dist = sum(abs(theta[a] - theta[b])) \n",
+    "    diffrnt_dist = sum(abs(theta[a] - theta[c]))\n",
+    "\n",
+    "    if (similar_dist > diffrnt_dist):\n",
+    "        print(a, b, c)\n",
+    "        print(explanation)\n"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "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": "3.8.11"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}