JDAI - Machine Learning

1. Introduction

1.1 Machine Learning overview

1.2 Algorithms and hyperpartameters

Algorithms

  • Stochastic Gradient Descent(SGD) 随机梯度下降
  • Binary classification: Logistic loss function + SGD
  • Multiclass classifiction: Multinomial logistic loss function + SGD
  • Regression: Linear regession(squared loss fucntion + SGD)

    Hyperparameters

  • Learning rate
  • Model size
  • Number of passes
  • Data shuffling
  • Regularization

    1.3 Steps in Machine Learning

    1. Prepare data
    2. Create a trainging datasource
    3. Create an ML model
    4. Review the ML model's predictive performance
    5. Use the ML model to generate predictions
    6. Clean up

2. Binary Model

2.1 Exploring dataset

2.2 Predict Response to a Marketing offer

  • Data set
  • Our objective

在实际的工作前，我们用一组过去的银行贷款数据为例，通过模型决定出谁还款是没有问题的（你的目标客户）

2.3 Prepare data

step

2.4 Create a datasource

2.5 Confirm schema

2.6 Create a binary classification Model

2.7 Understanding binary model's perdictive performance

performance badorgood

2.8 Setting binary model's prodictive performance

2.9 Use the ML Model to Gernaerate Predictions

  • Realtime: for a single observation
  • Batch: for a group observation step

    2.10 Create batch predictions

    Accuracy 结果会保存在S3，有一个result的目录生成

3. Multiclass Model

3.1 Exploring Multiclass model data set

这个例子里，我们用森林覆盖率作为例子，做一个训练 CoverType：森林覆盖的情况

3.2 Multiclass model data preparation

  • binary
  • categorical
  • numeric
  • text

    3.3 Multiclass Machine Learning model

    3.4 Predictions and evaluations of multiclass learning model

    evaluations 同时，我们也使用F-Score来判断模型的Auccracy evaluations

    3.5 Generate predictions for mutilclass

    我们的目录结构是这样的 evaluations

    3.6 create mulitclass batch predictions

    3.7 Interpreting batch predictions

    结果是这样的，我们从中选出最尤的几个解 evaluations

4. Regression Model

这个例子中，我们使用了house pricing的data，来预测房价

4.1 Exploring regression model data set

so we're given almost 80 different features including things like the number of

  • kitchens,
  • the bathrooms,
  • the square footage, all of the things that could determine the price of a house and we need to try and predict the house price based on that. data

    4.2 Regression data preparation

    4.3 create ML model

    4.4 Predictions and evaluations of ML model

    4.5 Regression batch predictions

    RegressionAccuracy 中文叫残差，如果回归模型正确的话， 我们可以将残差看作误差的观测值。 它应符合模型的假设条件，且具有误差的一些性质。利用残差所提供的信息，来考察模型假设的合理性及数据的可靠性称为残差分析。正数，负数都是不理想的结果。nagative rasidual indicates overestimating the target.

If we didn't have a zero-centered bell-shaped curve, then there's some definite error with the model's prediction and we might then try and add some more features to our model, so in the example of house prices, we might try and include new features such as the say the area of the garden or does the house have under floor heating or the energy rating of the house and we'll use these to try and determine what is not being captured by our model. Residuals

4.6 Regression batch perdictions

  • ML model for batch prediction
  • data for prediction
  • Batch prediction results
  • review

Other methods

  • Sagemaker
  • AWS Deep Learing AMI(CPU, and GPU) » for deep learning
  • Apache MXNet
    • Python
    • R