NeuroIA/intro_a_ml.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "-"
    }
   },
   "source": [
    "<a href=\"https://colab.research.google.com/github/institutohumai/cursos-python/blob/neuro/NeuroIA/intro_a_ml.ipynb\"> <img src='https://colab.research.google.com/assets/colab-badge.svg' /> </a>\n",
    "    <div align=\"center\"> Recordá abrir en una nueva pestaña </div>\n",
    "    "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "# Introducción a Machine Learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "- ¿Qué es Machine Learning?\n",
    "    - Definiciones: ML, DL, aprendizaje, atributos, observaciones, target o label...\n",
    "- El proceso de ML\n",
    "  - Feature engineering, modelos supervisados y no supervisados, train/test split, cross validation, evaluación de modelos\n",
    "- Etapas de un proyecto de ML\n",
    "    - Framing del problema de negocio, integración de datos, modelado, puesta en producción y monitoreo.\n",
    "- Casos de Aplicación\n",
    "    - Fraude, churn, upselling, cross-selling, segmentación, clustering geográfico."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "# ¿Qué es Machine Learning?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "¿Cómo haríamos una serie de instrucciones para clasificar imagenes de gatos y perros? ¿O qué número aparece?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center>\n",
    "<table><tr>\n",
    "    <td><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image4.png width=400></td>\n",
    "    <td><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image5.png width=400></td>\n",
    "    </tr>\n",
    "</table>\n",
    "</center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Machine Learning, Inteligencia Artificial y Data Science"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "La “Inteligencia Artificial Simbólica” dominó el paradigma de IA desde 1950 hasta la parte de los ‘80. Su pico de popularidad: boom de los “sistemas expertos”. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Programación clásica (el paradigma de la IA simbólica): \n",
    "Los humanos ingresan reglas (un programa), datos a ser procesados según esas reglas, y de este proceso resultan las respuestas esperadas. \n",
    "\n",
    "Machine Learning: los humanos ingresan datos como input además de las respuestas esperadas, y las reglas surgen como output. Estas reglas pueden luego ser aplicadas a nuevos datos para producir respuestas originales."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image7.png width=600></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image6.png width=700></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "*Decimos que un programa “aprende” si para cierta tarea, su performance mejora cuando crece la experiencia.* T. Mitchell\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "El aprendizaje entra en juego cuando le damos a estos modelos parámetros que pueden modificarse y que se adaptan a los datos observados de manera automática; de esta forma, puede considerarse que el algoritmo aprende de los datos. \n",
    "Una vez que estos modelos han sido ajustados a datos observados previamente, pueden ser usados para predecir y entender aspectos de datos nuevos. \n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image9.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "¿Cuáles son algunos ejemplos de IA que no son ML?\n",
    "\n",
    "- Sistemas basados en reglas\n",
    "- Demostración automática (proof solvers)\n",
    "- Optimización convexa en investigación de Operaciones\n",
    "- Computer vision tradicional\n",
    "\n",
    "¿Y DS que no es ML?\n",
    "\n",
    "- Visualización\n",
    "- EDA\n",
    "- Inferencia Estadística \n",
    "- Análisis de audio \n",
    "- Métodos econométricos\n",
    "- Biometría\n",
    "- …?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Factores influyentes"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "- Enorme cantidad de datos generados por minuto\n",
    "- Posibilidad de trabajar sobre datos desestructurados\n",
    "- Capacidad computacional\n",
    "\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=\"https://mloapwwtsct7.i.optimole.com/A6Igeow.l1xK~43710/w:auto/h:auto/q:75/https://www.bondhighplus.com/wp-content/uploads/2020/05/Internet-Minute-2021.jpg\" width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image11.png width=600></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### No es todo nuevo"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "- Regla de Bayes (Bayes, 1763) de probabilidad\n",
    "- Regresión por Mínimos Cuadrados (Gauss, 1795) de astronomía\n",
    "- Máxima verosimilitud (Fisher, 1922) de estadística\n",
    "- Redes neuronales artificiales (McCulloch/Pitts, 1943) de neurociencias\n",
    "- Juegos Minimax (von Neumann, 1944) de economía\n",
    "- Descenso de gradiente estocástico (Robbins/Monro, 1951) de optimización\n",
    "- Búsqueda de costo uniforme (Dijkstra, 1956) de algoritmos\n",
    "- Iteración de valores (Bellman, 1957) de teoría de control"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Machine Learning"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Un modelo de ML es una función matemática. El **aprendizaje** consiste en ajustar los parámetros del modelo de modo que el output se acerce al *target* correspondiente de cada input. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image12.gif width=700></center>\n",
    "\n",
    "https://www.cs.ryerson.ca/~aharley/vis/conv/"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "\n",
    "Existen ya unos cuantos hitos acumulados de IA en los últimos años. En el programa Jeopardy! (IBM Watson, 2011), el juego Go (DeepMind, AlphaGo, 2016), el multijugador online masivo Dota 2 (OpenAI, 2019), Poker (CMU and Facebook, 2019).\n",
    "\n",
    "También contamos con sistemas que rinden de manera excepcional en lectura y comprensión de lenguaje natural, reconocimiento del discurso, reconocimiento facial, o sistemas de imagen médicos. "
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Algunos ejemplos sci-fi"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Instance segmentation\n",
    "\n",
    "<center><img src=https://miro.medium.com/max/2472/1*QKfIOzB-N1GrhULIvVxkQg.gif width=800></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Robótica\n",
    "\n",
    "<!--<img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/boston.png width=400>\n",
    "<img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image16.gif width=400>-->\n",
    "\n",
    "<center><img src=https://bgr.com/wp-content/uploads/2019/10/gif.gif width=800></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Nuevas posibilidades...\n",
    "\n",
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image17.png width=800></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Transferencia de estilo\n",
    "\n",
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image18.png width=600></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Speech2face\n",
    "\n",
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image20.png width=800></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "GPT-3 \n",
    "\n",
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image21.gif width=800></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Go \n",
    "\n",
    "<center><img src=https://c.files.bbci.co.uk/6DFD/production/_88775182_b810d6ed99.jpg width=800></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Poker\n",
    "\n",
    "<center><img src=https://i.ytimg.com/vi/wInvN096he8/maxresdefault.jpg width=800></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Volviendo a lo pragmático..."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Conceptos Básicos"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "En nuestro caso la experiencia es un conjunto de observaciones con atributos. En algunos casos, estas observaciones tienen una variable target que queremos predecir. El target es optativo.\n",
    "Para que los modelos de Machine Learning aprendan es necesario que la información esté organizada de manera matricial como a continuación:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/basics.png width=900></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "## Tipos de tareas de ML"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/tipos_modelos.png width=500></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "### Modelos supervisados\n",
    "\n",
    "- Regresión lineal\n",
    "- Regresión logística\n",
    "- KNN (K nearest neighbours)\n",
    "- Y muchos más..."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "![image.png](data:image/png;base64,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)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Modelos basados en árboles: en general los mejores predictores.\n",
    "\n",
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image33.png width=600></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Los modelos de Machine Learning más performantes, se basan en “ensamblar” muchos árboles simples. Sirven tanto para clasificación como para regresión. \n",
    "\n",
    "- Random Forest\n",
    "- XGBoost\n",
    "- LightGBM\n",
    "- Tree Gradient Boosting"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "Los modelos de ML sólo son capaces de aprender de representaciones numéricas.\n",
    "¿Cómo podremos entonces trabajar con texto, categorías o imágenes?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image30.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Categorías: One Hot Encoding o variables dummies"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image28.png width=800></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Texto: Bag of Word"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image29.png width=900></center>"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 2,
   "metadata": {},
   "outputs": [],
   "source": [
    "from sklearn.feature_extraction.text import CountVectorizer\n",
    "\n",
    "textos = ['Cello de madera', 'Dentífrico de Datos', 'Historia de la Mesopotamia', 'Animense a preguntar']"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "<4x10 sparse matrix of type '<class 'numpy.int64'>'\n",
       "\twith 12 stored elements in Compressed Sparse Row format>"
      ]
     },
     "execution_count": 3,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "vec = CountVectorizer()\n",
    "X = vec.fit_transform(textos)\n",
    "X"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 4,
   "metadata": {},
   "outputs": [
    {
     "data": {
      "text/plain": [
       "matrix([[0, 1, 0, 1, 0, 0, 0, 1, 0, 0],\n",
       "        [0, 0, 1, 1, 1, 0, 0, 0, 0, 0],\n",
       "        [0, 0, 0, 1, 0, 1, 1, 0, 1, 0],\n",
       "        [1, 0, 0, 0, 0, 0, 0, 0, 0, 1]])"
      ]
     },
     "execution_count": 4,
     "metadata": {},
     "output_type": "execute_result"
    }
   ],
   "source": [
    "X.todense()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Podemos volcar este resultado en un `DataFrame`. Para obtener el encabezado de las columnas, vamos a utilizar el método `get_feature_names()`, propio del vectorizador."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "import pandas as pd\n",
    "pd.DataFrame(data = X.todense(), columns = vec.get_feature_names())\n",
    "\n",
    "Una vez que tenemos nuestros datos ordenados, podemos ajustar nuestro modelo de ML aproximadamente así:\n",
    "\n",
    "```python\n",
    "from sklearn.familia_de_modelos import Modelo\n",
    "mi_modelo = Modelo(hiperparámetros)\n",
    "mi_modelo.fit(X, y)\n",
    "\n",
    "predicción = mi_modelo.predict(nuevo_X)\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Para ajustar nuestros modelos de machine learning, vamos a tener nuestra matriz de features **X** y un vector **y** (en el caso de ML supervisado), donde **X** tiene la forma ``(n_observaciones, n_variables)`` e **y** ``(n_observaciones)``"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "```python\n",
    "from sklearn.linear_model import LinearRegression\n",
    "\n",
    "modelo = LinearRegression() # modelo = b_0 + b_1 * x\n",
    "modelo.fit(X, y) # se calculanlos b_0 y b_1 optimos, de mejor ajuste\n",
    "```"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Ejemplo base:\n",
    "\n",
    "## Regresión Lineal\n",
    "\n",
    "En estadística, la regresión lineal o es un modelo matemático usado para aproximar la relación de dependencia entre una variable dependiente $Y$, y  $x$. Este modelo puede ser expresado a grandes rasgos como:\n",
    "\n",
    "$$Y=\\beta_0+\\beta_1x_1$$\n",
    "\n",
    "donde:\n",
    "\n",
    "* <math>Y</math> es la variable dependiente o variable de respuesta.\n",
    "* $x$, es las variables explicativas, independientes o regresoras. Podemos tener cualquier número.\n",
    "* $\\beta_0,\\beta_1$ son los parámetros del modelo, miden la influencia que las variables explicativas tienen sobre el regrediendo.\n",
    "\n",
    "el término $\\beta_0$ es la intersección o término \"constante\", las $\\beta_i $ son los parámetros respectivos a cada variable independiente$. \n",
    "\n",
    "<img width='600px' src='https://dphi.tech/blog/wp-content/uploads/2020/04/example-regression-simulation.gif'></img>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Entrenamiento"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Ya tengo mis datos representados como vectores numéricos o “features” y etiquetas o “labels”. \n",
    "\n",
    "¿Cómo entreno un modelo que capture las relaciones entre los mismos?"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Generalización y sobreajuste"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Ya tengo un modelo que me permite predecir, clusterizar o reducir dimensiones.\n",
    "\n",
    "¿Cómo puedo saber cuán bien va a funcionar sobre datos nuevos?"
   ]
  },
  {
   "cell_type": "code",
   "execution_count": 3,
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "outputs": [
    {
     "data": {
      "image/png": "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",
      "text/plain": [
       "<Figure size 432x288 with 1 Axes>"
      ]
     },
     "metadata": {
      "needs_background": "light"
     },
     "output_type": "display_data"
    }
   ],
   "source": [
    "import numpy as np\n",
    "from matplotlib.cm import get_cmap\n",
    "import matplotlib.pyplot as plt\n",
    "from sklearn.linear_model import LinearRegression\n",
    "from sklearn.preprocessing import PolynomialFeatures\n",
    "\n",
    "def f(x):\n",
    "    \"\"\"Función a aproximar por interpolación polinomial\"\"\"\n",
    "    return x * np.sin(x)\n",
    "\n",
    "grados = [1, 3, 20]\n",
    "\n",
    "# Generamos algunos puntos de la función\n",
    "x = np.linspace(0, 10, 100)\n",
    "np.random.shuffle(x)\n",
    "x = np.sort(x[:35])\n",
    "y = f(x) + np.random.normal(size=x.shape) * 2 \n",
    "\n",
    "# Puntos para plotear\n",
    "x_plot = np.linspace(0, 10, 100)\n",
    "\n",
    "# Pasamos a formato matriz\n",
    "X = x[:, np.newaxis]\n",
    "X_plot = x_plot[:, np.newaxis]\n",
    "\n",
    "plt.scatter(x, y, color='navy', s=10, marker='o')\n",
    "\n",
    "for i, grado in enumerate(grados):\n",
    "    poly = PolynomialFeatures(grado)\n",
    "    X_poly = poly.fit_transform(X)\n",
    "    \n",
    "    model = LinearRegression()\n",
    "    model.fit(X_poly, y) \n",
    "    \n",
    "    y_plot = model.predict(poly.fit_transform(X_plot))\n",
    "    \n",
    "    plt.plot(x_plot, y_plot, color=get_cmap('Set1')(i), label=f\"grado {grado}\")\n",
    "\n",
    "plt.legend(bbox_to_anchor=(1.05, 1), loc='upper left')\n",
    "plt.ylim(-15,15)\n",
    "plt.show()"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Modelos no supervisados"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "# Clustering\n",
    "\n",
    "*Clustering* significa **agrupar** en inglés, o **encontrar grupos**.\n",
    "\n",
    "#### Modelo ejemplo: KMeans (clustering)\n",
    "\n",
    "* Asignar centroides al azar\n",
    "* Calcular qué punto pertenece a cada cluster\n",
    "* Promediar todos los puntos en cada dimensión para recalcular el centroide\n",
    "* Repetir hasta que los puntos dejen de cambiar de cluster.\n",
    "\n",
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image34.png width=500></center>\n",
    "\n",
    "Ver animación en: \n",
    "\n",
    "https://stanford.edu/class/engr108/visualizations/kmeans/kmeans.html"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "PCA (Reducción de dimensiones)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image35.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Evaluación"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "¡Cuidado! Es imposible usar el futuro para predecir el presente. \n",
    "\n",
    "Cuando los modelos incorporan información que naturalmente no debería estar disponible este problema se llama “information leak”.\n",
    "\n",
    "Aquí un caso famoso de information leak que tomó conocimiento público: https://liaa.dc.uba.ar/es/sobre-la-prediccion-automatica-de-embarazos-adolescentes/"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Cross Validation"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/cv.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Holdout sets para la evaluación (series de tiempo y datos de panel)"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image41.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Métricas de evaluación en problemas supervisados"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "Regresión"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {},
   "source": [
    "<img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image47.png width=400>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Clasificación\n",
    "\n",
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image42.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image43.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Matriz de confusión"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image44.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "Clasificación binaria: AUC (área debajo de la curva)\n",
    "\n",
    "Modelo perfecto:\n",
    "AUC = 1\n",
    "\n",
    "Ordenamiento de los casos al azar:\n",
    "AUC = 0.5\n",
    "\n",
    "Modelo que predice exactamente lo contrario a lo real:\n",
    "AUC = 0"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<center><img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image45.png width=400></center>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image46.png width=400>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "# En resumen:"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Para entrenar un modelo se necesita:\n",
    "\n",
    "- 1) Crear buenas características numéricas o “features” que representen las observaciones \n",
    "- 2) Conservar un conjunto de datos para evaluación\n",
    "- 3) Elegir un algoritmo que resuelva nuestro problema\n",
    "- 4) Evaluar los resultados\n",
    "- 5) Desplegar el modelo “en producción”"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Casos: problemas supervisados"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "**Predicción de propensión a conversión online**\n",
    "\n",
    "Objetivo: predecir qué tan probable es que un visitante a una app realice una conversión. Ésta puede ser una compra online, un clic, completar un formulario, etc.\n",
    "Esto se usa para decidir cuánto pagar en una subasta, asignar un precio o producto dinámicamente, etc."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "**Modelos de upselling/cross-selling**"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Objetivo: predecir la probabilidad de que un cliente compre un producto más caro (upselling) o complementario (cross-selling). \n",
    "Esto se puede integrar a un CRM o una herramienta de marketing para realizar campañas automáticamente."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "**Predicción de churn**"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Objetivo: predecir la probabilidad de que un cliente se dé de baja en determinado período.\n",
    "Sabiendo quiénes son los más propensos podemos generar un incentivo para impedirlo.\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "**Detección de fraude**"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "Objetivo: predecir la probabilidad de que una transacción sea fraudulenta\n"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "**Sistemas de recomendación**"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "\n",
    "Objetivo: optimizar qué productos ofrecer en una plataforma de ventas online"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<div class=\"row\">\n",
    "    <img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image53.png width=100>\n",
    "    <img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image54.png width=100>\n",
    "    <img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/ml-logo.png width=100>\n",
    "    <img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image56.png width=100>\n",
    "    <img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image57.jpg width=100>\n",
    "</div>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "## Casos: problemas no supervisados"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "**Segmentación automática**\n",
    "\n",
    "Generación de segmentos a partir de información demográfica y de comportamiento.\n",
    "Con esta técnica se pueden generar K segmentos automáticamente, de modo de accionar de manera distinta sobre cada uno de ellos."
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "subslide"
    }
   },
   "source": [
    "**Clustering geográfico**\n",
    "\n",
    "Encontrar patrones geográficos automáticamente.\n",
    "Por ejemplo: paradas de colectivos, transacciones de clientes,etc.\n",
    "\n",
    "<img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/geo.png width=400>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### El proceso real de los productos basados en ML"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image60.png width=800>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "### Equipo de data"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "fragment"
    }
   },
   "source": [
    "<img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image61.png width=800>"
   ]
  },
  {
   "cell_type": "markdown",
   "metadata": {
    "slideshow": {
     "slide_type": "slide"
    }
   },
   "source": [
    "<img src=https://github.com/institutohumai/cursos-python/raw/ml0/MachineLearning/media/image12.png width=800>"
   ]
  }
 ],
 "metadata": {
  "celltoolbar": "Slideshow",
  "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": "3.8.8"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 4
}