Machine Learning

• Machine Learning
  • Introduction
  • Definition Machine Learning (ML)
  • Types of Machine Learning
    • 1. Supervised ML
    • 2. Unsupervised ML
    • 3. Reinforcement ML
  • The Machine Learning Pipeline
  • Getting Started
  • Data preparation and preprocessing
    • Data Collection
    • Data Preparation
    • Data Cleaning
    • Data Preprocessing
    • Data Visualization
  • Model Training
    • Algorithms
    • Formatting data
    • Testing and validation techniques
    • Splitting data
    • Cross-validation
    • Model training concepts
  • Model Evaluation
    • Bias Variance Tradeoff
    • Model Metrics
    • Classification Metrics
    • Regression Metrics
    • Validation Curve
    • Learning Curve
    • Model Debugging
  • Feature Engineering
    • Feature Extraction
    • Feature Selection
    • Feature Creation and transformation
    • Bagging/Boosting
  • Model Training/Tuning
    • Training Data Tuning
    • Regularization
    • Hyperparameter tuning
    • Grid Search
    • Randomized Search
    • Bayesian Search
  • Choosing the ML model
    • Supervised Machine Learning Algorithms
    • Neural Networks
      • Perceptron
      • Neural network architecture
      • Convolutional Neural Networks
      • Recurrent neural networks
    • K-nearest neighbor
    • Linear Model
  • ML Data Readiness
  • Productizing a ML Model
  • Aspects to consider
  • Types of Production environments
    • Batch predictions
    • Online Predictions
    • Online training
  • Business metrics vs Model Metrics
  • Storage
  • Model and Pipeline Persistence
  • Model Deployment
  • Monitoring
  • Maintenance
  • Common Mistakes

Introduction

Definition Machine Learning (ML)

• The study of computer programs (algorithms) that can learn by example.
• ML algorithms can generalize from existing examples of a task.

Types of Machine Learning

1. Supervised ML

• Learn to predict target values from labelled data.

1. Classification: Target values with discrete classes
   1. Binary Classification: To identify targets with two classes
      • eg: Fraud detection or Spam email identification
   2. Multi-class identification: To identify targets with more than two classes.
      • eg: Differentiate between fruit types
2. Regression: Target values are continuous values.
   • eg: Predict house price

2. Unsupervised ML

• Find structure in unlabeled data

1. Clustering: Find groups of similar instances in the data. eg: Finding clusters of similar users.

3. Reinforcement ML

• Learning through Trial and Error

The Machine Learning Pipeline

1. Business Problem:
   • Identify the problem that could benefit from ML
2. Business Formulation: Preparing the problem
   • Identify ML model type (Classification, Regression)
   • Frame the simplest solution without losing important information
   • Choosing the data:
     • How much is it?
     • Where is it?
     • Do I have access to it?
   • Get a domain expert:
     • An expert for the business case.
     • Can identify the important features
     • Decide whether the data is representative for the real world
   • Evaluate the data quality
   • Identifying the features and the label
   • Does the problem needs a lot of labeled data?
   • Identify the metrics:
     • Model performance metric
     • Used during model training and evaluation
     • Business goal metric
     • Used after model deployment
     • Measures how well the model is performing
3. Data preparation and preprocessing
   • Data collection and integration
   • Determine where data comes from
   • Data Preprocessing and visualization
   • Design data for the model
4. Model Training and Tuning
5. Model Evaluation:
   • Testing the model on new data and assess the results
6. Optimization
   • Data Augmentation:
     • Modify the data
   • Feature Engineering:
     • Create new features
7. Model Deployment

Getting Started

Data preparation and preprocessing

Data Collection

• It can be structured or unstructured data.
• It can be collected from many sources like: server, database, disk, clickstream, multimedia, IoT and sensors, Social media.
• When in the cloud, Data Lake is used. It can store and serve both structured and unstructured data.

Data Preparation

• Reformat the collected data from (CSV, JSON, Pickle, ..etc) into a tabular format.
• Make sure that the data is imported properly by importing the first few rows
• Understand the data dimensions, column names
• Checking for missing data, duplicates, wrong data types

Data Cleaning

• Missing Data:
  • Sources:
  • Undefined values, collection errors, left joins, etc..
  • Issues:
  • Many learning algorithms can't handle missing values.
  • Makes it hard to interpret a target relationship
  • Identifying the cause can determine whether to delete or impute them.
  • What were the mechanisms that caused the missing values?
  • Is it random and which kind of values are missing?
  • Are there rows or columns missing that not aware of?
  • Dropping:
  • Risk of dropping rows:
    • Not enough training samples (overfitting)
    • May bias sample
  • Risk of dropping columns:
    • May lose information in features (underfitting)
  • Imputation:
  • Unit Non-Response:
    • Refers to entire rows of missing data
    • Imputation Methods include: Weight-Class Adjustments
  • Item Non-Response
    • Specific cells of a column are missing
    • Types:
    • MCAR: stands for Missing Completely at Random.
      • This happens when missing values are missing independently from all the features as well as the target.
    • MAR: stands for Missing at Random.
      • This occurs when the missing value is dependant on a variable, but independent from itself.
    • MNAR: stands for Missing Not at Random.
      • This is the case where the missingness of a value is dependent on the value itself.
    • Methods:
    • Weight-Class Adjustments
    • Deductive Imputation
    • Mean/Median/Mode Imputation
    • Hot-Deck Imputation
    • Model-Based Imputation (Regression, Bayesian, etc)
    • Proper Multiple Stochastic Regression
    • Pattern Submodel Approach
  • Python libraries:
    • SciKit learn:
    • sklearn.preprocessing.SimpleImputer
    • Advanced methods for Imputation:
    • MICE (Multiple Imputation by chained Equations). sklearn.impute.MICEImpute (v0.20)
    • Python (not sklearn) fancyimpute package (KNN impute, SoftImpute, MICE, etc..)
  • References:
    • A Comprehensive Guide To Data Imputation
    • Matt Brems Repo
• Inconsistency
• Column values with different units
• Wrong or unrelated column values
• Outliers:
  • It can:
  • Add richness to the data
  • Make accurate predictions more difficult
  • Indicate that the data point belongs to another column
  • Types:
  • Artificial: when the outlier doesn't belong to the real world
    • eg: Age is 150
    • It needs to be deleted
  • Natural: when the outlier can be actually genuine
    • eg: Salary of CEO vs other employees
    • Transform the outlier
    • Ex: Use the natural log of each value in the column to reduce the extreme variation between the values.
    • Impute a new value for the outlier
    • Ex: Use the mean of that column

Data Preprocessing

• Descriptive Statistics:
  • Categorical vs Numerical stats
  • Understanding Mean and Median

• Encoding Labels for categorical features:

  • It is converting categorical variable into a numerical variable.
  • Ordinal:
  • SciKit Learn library, LabelEncoder, converts categorical variable to numerical variable that starts with 0 and increase with 1. But if this applied to non-nominal categorical type, may lead to wrong computations or wrong usages. So, it is recommended to be used with Ordinal type that has relationships with each other.

      from sklearn.preprocessing import LabelEncoder
  
      loan_enc = LabelEncoder()
      y = group_enc.fit_transform(df['loan_approved'])
  

  • Nominal:
  • While library OneHotEncoder is more likely to be used for Nominal variables.

    from sklearn.preprocessing import OneHotEncoder
  
    df = pd.DataFrame({"Fruits":["Apple","Banana","Banana","Mango","Banana"]})
    num_type = group_enc.fit_transform(df['Fruits'])
    type_enc = OneHotEncoder()
    type_enc.fit(num_type.reshape(-1,1))
  

  • Pandas library has get_dummies() function to do the same as OneHotEncoder

    import pandas as pd
  
    df = pd.DataFrame({"Fruits":["Apple","Banana","Banana","Mango","Banana"]})
  
    pd.get_dummies(df)
  

  • Encoding with many classes:
  • Define a hierarchy structure.
  • Try to group the levels by similarity to reduce the overall number of groups.

Data Visualization

• Categorical data visualization:
  • Bar Charts
• Numerical data:
  • Histograms
  • How many peaks
  • Is there any skewness
  • Is the data normally distributed
  • Density plot
  • Identify Skewness
  • Detect outliers
  • Box plot
  • Detect outliers
  • Visualize mean, std, IQR
• Multivariate stats
  • Benefits:
  • Identify correlation between features.
  • High correlation between features can sometimes lead to poor model performance
  • Visualization techniques:
  • Scatterplot
  • Scatterplot with labels identification
  • Scatterplot matrix
  • Correlation matrix with Heatmap

Model Training

• The goal of training is to create an accurate model that answers the business question accurately as often as you need it or more.

Algorithms

1. Supervised Learning
   1. Classification
      1. Binary
         1. Linear learner
         2. XGBoost
      2. Multi-class
         1. XGBoost
         2. KNN
   2. Regression
      1. XGBoost
      2. KNN
      3. Linear learner
      4. Factorization machines
   3. Recommendation
      1. Factorization machines
2. Unsupervised Learning
   1. Clustering
      1. K-means
      2. LDA
   2. Topic modeling
      1. LDA
   3. Embeddings
      1. Object2Vec
   4. Anomaly detection
      1. Random cut forest
      2. IP insights
   5. Dimensionality reduction
      1. PCA

Formatting data

• Common types for algorithms:
  • CSV: Comma separated values
  • Label on the left
  • rec: Record-IO protobuf

Testing and validation techniques

Splitting data

• We split the data to avoid overfitting and get generalized performance. The data is split into three sets:
  • Training set:
  • It is used in the model training phase to see patterns.
  • It ranges around 80% of the full data
  • Validation/Evaluation set:
  • It is used also in the training phase but used to give an estimate of model performance and/or compare performance across different models.
  • It ranges around 10% of the full data.
  • Testing set:
  • It is used to evaluate the predictive quality of the model.
  • It ranges around 10% of the full data.
• Python SciKit learn sklearn.model_selection.train_test_split can be used for splitting

Cross-validation

• It compares the performance of multiple models
• It basically gives more stable and reliable estimates of how the classifiers likely to perform on average by running multiple different training test splits and then averaging the results.
• It gives information about how sensitive the model is, to the nature of the specific training set.
• It does take more time and computation to do cross-validation.
• The records might need shuffling to avoid possible bias in the ordering by class label. For example, the first 20% of the data has the same label.

1. K-Fold Cross Validation:
   • Used when there is a small dataset
   • Most commonly used with K set to 5 or 10.
   • As K increases, TIME and Variance in test data increases and BIAS decreases
   • Eg: to do five-fold cross-validation, the original dataset is partitioned into five parts of equal or close to equal sizes. Each of these parts is called a "fold". Then a series of five models is trained and validated using a different fold. In such that Model one, is trained using folds 2 through 5 as the training set and evaluated using fold 1 as the validation set. Model 2, is trained using Folds 1, 3, 4, and 5 as the training set, and evaluated using Fold 2 as the validation set, and so on. When this process is done, we have five accuracy values, one per fold.
2. Iterated K-Fold validation with shuffling
   • It is the same as K-Fold validation but with extra K iteration. The difference between each iteration is that the data set is shuffled differently in each iteration.
   • Eg: In K = 3, the first iteration is occurred same as 3-Fold validation which the average of the 3 models is taken, then two more iteration of this 3-Fold validation but with shuffling the data before splitting them. Then the average of the 3 iterations is taken.
3. Leave-one-out cross-validation:
   • It is used for very small datasets.
   • Test set is one data point.
4. Stratified K-Fold cross-validation:
   • Distributes label class across training and testing datasets.
   • For imbalanced data.

• SciKit Learn library sklearn.model_selection.cross_val_score can be used for cross validation:
  • In Classification problem, SciKit Learn do "Stratified K-fold Cross-validation". The Stratified Cross-validation means that when splitting the data, the proportions of classes in each fold are made as close as possible to the actual proportions of the classes in the overall data set.
  • In regression, SciKit Learn uses regular k-fold cross-validation.

Model training concepts

• How training the model works:
  • The model is given a specific set of features.
  • The model predicts the classes for these features based upon the weights that was given to the features.
  • Then the predictions are compared with the actual labels to compute the loss.
  • Based on the loss computed, the model parameters (features weights) are updated to minimize the loss.
• What is loss function
  • Loss function is the measure of the error in the model predictions, given a certain weights.
• Types of loss functions:
  • RMSE (Root mean square error):
  • Describes the sample standard deviation of the differences between predicted and observed values.
  • $$\sqrt{\frac{\sum_{i=1}^n(Y_{\text{target},i}-Y_{\text{pred},i})^2}{n}}$$
  • Logs Likelihood loss:
  • Considers the logarithm of probabilities of each class.
  • In Binary classification: $-(y\log{p} + (1-y) \log{(1-p)})$
• Optimization:
  • It is used to find the Minima.
  • Minima is the minimum point in the plot between loss and parameters or the point with the least amount of error.
  • It is often that there might more than one local Minima, in which the model might get stuck in local Minima instead of the Global Minima.
  • Ways to find the Global Minima:
  • Comparing all possible values which is inefficient way.
  • Gradient Descent
    • It is the searching for the Minima by taking a step into the direction where the point on the graph with negative gradient. Until it finds an positive gradient, then it reverses the direction to another point where has negative gradient. This process occurs until the Minima is found
    • Learning Rate:
    • It is how big is the step to be taken.
    • If it too big, the local Minima will be hard to be found.
    • If it too small, it will too much time to be found.
    • Drawbacks:
    • Updates the parameters only after a pass through all the data (one epoch)
    • Can't be used when data is too large to fit entirely in memory.
    • Can get stuck at local Minima or fail to reach Global Minima.
  • Stochastic Gradient Descent:
    • It is the same as gradient descent except that the weights is updated at every data point.
    • It is very fast to converge.
    • The drawback is that it is very noisy, in such that the steps might be in several directions.
  • Mini-Batch Gradient Descent:
    • It uses mini batch of records, and then the parameters is updated.
    • It is slower than SGD but faster than Gradient Descent.
    • It doesn't consume much memory as SGD
  • Gradient Descent Variations:
  • To find the Minima, an equation is calculated which is the derivative of the plot when it is equals to zero. Which is the point where the slope is neither increase nor decreases.

Model Evaluation

Bias Variance Tradeoff

• Machine learning models depend on input data, output data, and understanding the relationship between the two. bias and variance affect the relationship between input and output data.

• Bias:

  • It is the gap between predicted value and actual value.
  • It is an error from flawed assumptions in the algorithm.
  • High bias can cause an algorithm to miss important relationships between features and target outputs resulting in underfitting.
  • Bias = $E[\hat{f}(x)] - f(x)$, Where $f(x)$ is the true model, $\hat{f}(x)$ is the estimated model.
  • Solution:
  • Increase the number of features
  • Decrease degree of regularization

• Variance:

  • How dispersed the predicted values are
  • It is an error from sensitivity to small variations in the training data. High variance can cause an algorithm to model random noise in the training set, resulting in overfitting.
  • Variance = $E[(\hat{f}(x)-E[\hat{f}(x)])^2]$, Where $f(x)$ is the true model, $\hat{f}(x)$ is the estimated model.
  • Solution:
  • Increase training data
  • Reduce model complexity
    • Decrease the number of features
    • Increase the degree of regularization

• Total Error $(x) = \text{Bias}^2 + \text{Variance} + \text{Irreducible Error}$

Model Metrics

Classification Metrics

• Confusion Matrix:
  • True Positive (TP): When model predicts Positive outcome as Positive
  • True Negative (TN): When model predicts Negative Outcome as Negative
  • False Positive (FP): When model predicts Positive outcome as Negative
  • False Negative (FN): When model predicts Negative outcome as Positive

To determine how well is the Logistic Model, the are some some metrics:

• Accuracy:
  • Accuracy (also called Score) is the proportion of correctly labeled rows divided by the total number of rows in data set. There are some cases when Accuracy won't work well, when there are large class imbalances in the dataset.

$$\text{Accuracy} = \frac{TP+TN}{TP+TN+FP+FN}$$

• Precision:
  • Out of all the items labeled as positive, how many truly belong to the positive class.

$$\text{Precision} = \frac{TP}{TP+FP}$$

• Recall:
  • Out of all the items that are truly positive, how many were correctly classified as positive. Or simply, how many positive items were 'recalled' from the dataset.

$$\text{Recall} = \frac{TP}{TP+FN}$$

• F1 score:
  • It helps express precision and recall with a single value.

$$\text{F1 Score} = \frac{2 . \text{Precision} . \text{Recall}}{\text{Precision} + \text{Recall}}$$

• AUC - ROC:
  • AUC: Area-under-curve (degree or measure of separability).
  • ROC: Receiver-operator characteristic curve (probability curve).
  • AUC - ROC curve:
  • A performance measurement for a classification problem at various threshold settings
  • The optimum model has 1 AUC value

Regression Metrics

• MSE (Mean Squared Error)
  • Average squared error over entire dataset
  • Mean squared error(MSE) = $\frac{1}{N}\sum_{i=1}^N(\widehat{y} - y_i)^2$
  • Very commonly used
  • SciKit-learn: sklearn.metrics.mean_squared_error
• $R^2$ error
  • $R^2 = 1 - \frac{\text{Sum of Squared Error (SSE)}}{\text{Var}(y)}$ which is between 0 and 1
  • Interpretation: Fraction of variance accounted for by the model
  • Basically, standardized version of MSE
  • Good $R^2$ are determined by actual problem
  • $R^2$ always increase when more variables are added to the model.
  • Highest $R^2$ may not be the best model.
• Adjusted $R^2$
  • Adjusted $R^2= 1-(1-R^2)\frac{\text{no. of data pts.} -1}{\text{no. of data pts. - no. of variables}-1}$
  • Takes into account of the effect of adding more variables such that it only increases when the added variables have significant effect in prediction.
  • It is a better metric for multiple variates regression.
  • SciKit-learn: sklearn.metrics.r2_score
• Confidence Interval
  • An average computed on a sample is merely an estimate of the true population mean.
  • Confidence interval: Quantifies margin-of-error between sample metric and true metric due to sampling randomness
  • Informal interpretation: with x% confidence, true metric lies within the interval.
  • Precisely: If the true distribution is as stated, then with x% probability the observed value is in the interval.
  • Z-score: Quantifies how much the value is above or below the mean in terms of its standard deviation

Validation Curve

• A validation curve is typically drawn between some parameter of the model and the model’s score. Two curves are present in a validation curve – one for the training set score and one for the cross-validation score.
• To evaluate the effect that an important parameter of a model has on the cross-validation scores.
• Like cross-value score, validation curve will do threefold cross-validation by default but it can be adjusted with the CV parameter as well.
• Unlike cross_val_score, you can also specify a classifier, parameter name, and set of parameter values, you want to sweep across.
• So you first pass in the estimator object, or that is the classifier or regression object to use, followed by the data set samples X and target values Y, the name of the parameter to sweep, and the array of parameter values that that parameter should take on in doing the sweep.

• It will return two two-dimensional arrays corresponding to evaluation on the training set and the test set. Each array has one row per parameter value in the sweep, and the number of columns is the number of cross-validation folds that are used.
  

  from sklearn.svm import SVC
  from sklearn.model_selection import validation_curve
  
  param_range = np.logspace(-3, 3, 4)
  train_scores, test_scores = validation_curve(SVC(), X, y,
                                  param_name='gamma',
                                  param_range=param_range, cv=3)
  

Learning Curve

• It used to detect if the model is underfitting or overfitting, and impact of training data size the error.
• It plots the training dataset and validation dataset error or accuracy against training set size.
• scikit-learn: sklearn.learning_curve.learning_curve
  • Uses stratified k-fold cross-validation by default if output is binary or multiclass (preserves percentage of samples in each class)
  • Note: sklearn.model_selection.learning_curve in v0.18

Model Debugging

• Filter on failed predictions and manually look for patterns.
  • Data problems (eg, many variants for same word)
  • Labeling errors (eg, data mislabelled)
  • Under/over-represented subclasses (eg, too many examples of one type)
  • Discriminating information is not captured in features (eg, customer location)

• This helps pivot on target, key attributes, and failure type, and build histograms of error counts.
  

  pred = clf.predict(train[col])
  error_df = test[pred != test['target']]
  

Feature Engineering

• It is the science/art of extracting more information from the existing data in order to improve the model's predictive power.
• There are two ways:
  • Reduce the dataset dimensionality using Feature Extraction and Feature Selection.
  • Increase the dataset dimensionality using Feature Creation and Transformation.
• This leads to Curse of Dimensionality where the model performance reaches its optimal state at certain number of features, however it drastically decreases by increasing the number of features any further.

Feature Extraction

• It is a technique used to reduce the dimensionality of the dataset.
  • a.k.a data compression
• It is the extraction of new features from the existing features in the dataset.
• It is considered old school and deep learning techniques are more efficient now.
• Motivation:
  • Improves computational efficiency
  • Reduces curse of dimensionality
• It is used in:
  • Images: extracting a certain components of the image to identify the image.
  • NLP: Popular words excluding articles and prepositions.
  • Structured data: Principle component analysis (PCA) or t-distributed stochastic neighbor embedding (T-SNE)

• Techniques

  • Principle component analysis (PCA)
  • It is unsupervised linear approach to feature extraction
  • Finds pattern based on correlations between features
  • Constructs principal components: orthogonal axes in directions of maximum variance.

  • scikit-learn: sklearn.decomposition.PCA
    

    pca = PCA(n_components=2)
    X_train_pca = pca.fit_transform(X_train_std)
    lr = LogisticRegression()
    lr.fit(X_train_pca)
    

  • Linear discriminant analysis (LDA)

  • A supervised linear approach to feature extraction

  • Transforms to subspace that maximizes class separability

  • Assumes data is normally distributed

  • Used for dimensionality reduction of features

  • Can reduce to at most #classes-1 components

  • scikit-learn: sklearn.discriminant_analysis.LinearDiscriminantAnalysis

  • Kernel versions of these for fundamentally non-linear data

Feature Selection

• It is a technique used to reduce the dimensionality of the dataset.
• It is by leaving only one of the highly correlated features.
• Filtering is a technique in Feature selection, which is eliminating the irrelevant data.

Feature Creation and transformation

• It is a technique used to increase the dimensionality of the dataset.
• It is the creation of new features using already existing data.
  • For example:
  • Creating a feature which is the multiplication of two other features.
  • Or creating new features by splitting an already existing feature.
• Techniques for Numerical Data:
  • Logarithmic transformation:
  • It is about changing the shape of the distribution plot.
  • It can be applied to right skewed plots.
  • It can not be applied to 0 or -ve values.
  • Square:
  • It can not be applied to negative values
  • It has moderate impact on distribution
  • Ex: Area of apartment
  • Cube:
  • It can be applied to negative numbers
  • It has high impact on distribution
  • Volume of rainfall in a year.
  • Binning:
  • It is the converting the values of a feature into features of ranges of the numbers.:
    • Ex: Converting Age column into columns of ranges. (0 to 7 years), (8 - 16 years) ....
  • Scaling:
  • It is the normalization of the feature values:
  • Techniques:
    • Mean/Variance standardization:
    • Centering the values around mean $\mu_j = 0$ with standard deviation $\sigma_j = 1$ for each column.
    • This can be achieved by removing the mean from the variable and divide it with the standard variance.
    • $$x_{i,j}^* = \frac{x_{i,j} - \mu_j}{\sigma_j}$$
    • Advantages:
      • Many algorithms behave better with smaller values
      • Keeps outlier information, but reduces impact.
    • SciKit-learn: sklearn.preprocessing.StandardScalar
    • MinMax Scaling:
    • It is the transformation of all the features, so they're all on the same scale between zero and one.
    • $$x_{i,j}^* = \frac{x_{i} - \min x_j}{\max x_j - \min x_j}$$
    • Advantages:
      • Robust to small standard deviations
    • SciKit-learn: sklearn.preprocessing.MinMaxScaler
    • MaxAbs scaling:
    • Divides each element by the maximum absolute value in the feature. $$x_{i,j}^* = \frac{x_{i,j}}{\max (|x_j|)}$$
    • Advantages:
      • It doesn't destroy sparsity, because there observations are not centered around any measurement
    • SciKit-learn: sklearn.preprocessing.MaxAbsScaler
    • Robust scaling
    • It is applied to particular features. $$x_{i}^* = \frac{x_{i} - Q_{50}(x)}{Q_{75}(x) - Q_{25}(x)}$$
    • Advantages:
      • Minimizes the impact of large marginal outliers.
      • After transformation, it will be robust outliers
    • SciKit-learn: sklearn.preprocessing.RobustScaler
    • Normalizer:
    • It is applied to rows.
    • Scaled values are scaled with standard deviation $\sigma_j = 1$ $$x_{i,j}^* = \frac{x_{i,j}}{\sigma_j}$$
    • SciKit-learn: sklearn.preprocessing.Normalizer
    • Rescales $x_j$ to unit norm based on:
      • L1 norm
      • L2 norm
      • Max norm
    • It is widely used in text analysis.
  • Critical aspects to Feature Engineering:
    • The scalar is fit to the training data only, then transform both train and validation data.
    • Apply the same scalar object to both training and testing data.
    • Training the scalar object on the training data and not on the test data. If it trained on the test data, it will cause a phenomena called Data Leakage, where the training phase has information that is leaked from the test set.
    • The scaling should applied to the real world numbers not only to that available in the dataset. For example: If the dataset has Age column that ranges from 20 to 50, this doesn't neglect the fact that Ages in real world can range from 0 to 80 or 100. So, this should be taken in consideration while scaling in order to generalize the model to the real world.
    • Scaling is applied differently to each column
  • Polynomial Features
    • Generate new features consisting of all polynomial combinations of the original two features $𝑥_0,𝑥_1$.
    • The degree of the polynomial specifies how many variables participate at a time in each new feature (above example: degree 2).
    • This is still a weighted linear combination of features, so it's still a linear model, and can use same least-squares estimation method for $w$ and $b$.
    • Adding these extra polynomial features allows us a much richer set of complex functions that we can use to fit to the data.
    • This intuitively as allowing polynomials to be fit to the training data instead of simply a straight line, but using the same least-squares criterion that minimizes mean squared error.
    • We want to transform the data this way to capture interactions between the original features by adding them as features to the linear model.
    • Polynomial feature expansion with high as this can lead to complex models that over-fit.
    • Polynomial feature expansion is often combined with a regularized learning method like ridge regression.
• Techniques for categorical data:
  • Ordinal categories:
  • Convert binary classifications to 0 and 1
  • Mapping multi categorical features to numerics with the assistance of the domain expert. For example: mapping Small, Medium, Large to 5, 10, and 20.
  • Nominal categories:
  • One Hot Encoding:
    • It is creating a binary column for each of the classes in the feature.
    • Pandas: pandas.get_dummies()
  • Grouping:
    • Create a binary column for a group of features together.
• Other techniques:
  • Radial Basis Function
  • Transform: $f(x) = f(||x - c||)$
  • Widely used in Support Vector Machine as a kernel and in Radial Basis Networks (RBNNs)
  • Gaussian RBF is the most common RBF used.
  • Text-based Features
  • Bag-of-words model
    • Represent document as vector of numbers, one for each word (tokenize, count, and normalize)
    • Note:
    • Sparse matrix implementation is typically used, ignores relative position of words.
    • Can be extended to bag of n-grams of words or of characters
  • Count Vectorizer
    • Per-word value is count (also called term frequency)
    • Note: Includes lowercasing and tokenization on white space and punctuation
    • scikit-learn: sklearn.feature_extraction.text.CountVectorizer
  • TfidfVectorizer
    • Term-Frequency times Inverse Document-Frequency
    • Per-word value is downweighted for terms common across documents (eg. "the")
    • scikit-learn: sklearn.feature_extraction.text.TfidfVectorizer
  • Hashing Vectorizer
    • Stateless mapper from text to term index
    • scikit-learn: sklearn.feature_extraction.text.HashingVectorizer

Bagging/Boosting

Feature extraction and selection are relatively manual processes. Bagging and boosting are automated or semi-automated approaches to determining which features to include.

• Bagging (Bootstrapping Aggregation)

  • Generate a group of weak learners that when combined together generate higher accuracy
  • Reduces variance
  • Keeps bias the same
  • sklearn:
  • sklearn.ensemble.BaggingClassifier
  • sklearn.ensemble.BaggingRegressor

• Boosting

  • Assign strengths to each weak learner
  • Iteratively train learners using misclassified examples by previous weak learners.
  • It is used for models that have a high bias and accepts weights on individual samples
  • sklearn:
  • sklearn.ensemble.AdaBoostClassifier
  • sklearn.ensemble.AdaBoostRegressor
  • sklearn.ensemble.GradientBoostingClassifier
  • XGBoost library

Model Training/Tuning

Training Data Tuning

• If training set too small, then Sample and Label more data if possible
• If training set biased against or missing some important scenarios, then Sample and Label more data for those scenarios if possible.
• If it is not easy to sample or label more, then consider creating synthetic data (duplication or techniques like SMOTE)
• IMPORTANT: Training data doesn't need to be exactly representative, but yor test set does.

Regularization

• Overfitting often caused by overly-complex models capturing idiosyncrasies in training set.
• Regularization is a technique used to reduce the errors by fitting the function appropriately on the given training set and avoid overfitting.
• Adding penalty score for complexity to cost function.
• $\text{cost}_{reg} = \text{cost} + \frac{\alpha}{2}\text{penalty}$
• Idea: Large weights corresponds to higher complexity.
• Two standard types:
  • L1 regularization, Lasso
  • L2 regularization, Ridge

Hyperparameter tuning

It is an Estimator parameter that is NOT fitted in the data

• Hyperparameter types:
  • Model Hyperparameter:
  • It helps to define the model itself.
  • Ex: Filter size, pooling, stride, padding
  • Optimizer Hyperparameter:
  • It is how the model learns patterns on data
  • Ex: Gradient Descent, Stochastic Gradient Descent
  • Data Hyperparameter:
  • It defines attributes for the data itself
  • Useful for small/homogenous datasets
• Hyperparameters must be optimized separately
• Methods for tuning hyperparameters:
  • Manually:
  • Manually select Hyperparameters based on one's intuition and experience.
  • Often too shallow and inefficient of an approach
  • Grid Search
  • Random Search
  • Bayesian Search
• Hyperparameter tuning doesn't always improve the model.
• Best practices:
  • Don't adjust every hyperparameter
  • Limit range of values to what's most effective.
  • Run one training job at a time rather in parallel.

Grid Search

• It finds the optimum combination of hyperparameters by exhaustive search over specified parameter values.
• Compute intensive
• scikit-learn: sklearn.grid_search.GridSearchCV
  • GridSearchCV(estimator, param_grid, scoring=None):
  • estimator is the ML model types. ex: tree for decision tree
  • scoring is your choice of model performance metric
  • param_grid is the hyperparameters values
    • ex: param_grid ={ max_depth: [5, 10, 50, 100, 250], min_samples_leaf: [15, 20, 25, 30, 35]}
  • GridSearchCV can be used as an estimator, with a fit and predict methods, it is also performing 5-fold Cross Validation by default for each combination of hyperparameters.
  • For the above example, 25 combinations of hyperparameters and with 5-fold CV for each combination, that would train 125 models. That will take time to complete.
  • Once all the combinations are evaluated, the model with the set of parameters which give the top metric is considered to be the best.
  • GridSearchCV returns the best combination of the hyperparameters, the best estimator equipped with these best hyperparameters, and will also report the performance metric of this best estimator.

Randomized Search

• Trained and scored on random combinations of hyperparameters
• Each setting is sampled from a distribution over possible parameter values.
• A more efficient implementation of hyperparameter tuning.
• RandomizedSearchCV(estimator, param_distributions, n_iter=10, scoring=None)
  • estimator is the ML model types. ex: tree for decision tree
  • scoring is your choice of model performance metric
  • param_grid is the hyperparameters values
  • ex: param_grid ={max_depth: [5, 10, 50, 100, 250], min_samples_leaf: uniform(15,35,5)}
  • n_tier (default 10) is the number of random parameter settings that are sampled. It trades off runtime vs quality of the solution.
• In contrast to GridSearchCV, not all parameter values are tried out, but rather a fixed number of parameter settings is sampled from the specified distributions. The number of parameter settings that are tried is given by n_iter.
• RandomizedSearchCV also performs 5-fold CV by default for each combination of hyperparameters.
• Can sample from distributions (sampling with replacement is used), if at least one parameter is given as a distribution.
• If all parameters are presented as a list, sampling without replacement is performed.
• If at least one parameter is given as a distribution, sampling with replacement is used.
• It is highly recommended to use continuous distributions for continuous parameters.

Bayesian Search

• Make guesses about hyperparameter combinations, then uses regressions to refine the combinations.
• It keeps track of previous hyperparameter evaluations and builds a probabilistic model.
• It tries to balance exploration (uncertain hyperparameter set) and exploitation (hyperparameters with a good chance of being optimum)
• It prefers points near the ones that worked well.
• AWS SageMaker uses Bayesian Search for hyperparameter optimization.

Choosing the ML model

Supervised Machine Learning Algorithms

The supervised aspect refers to the need for each training example to have a label in order for the algorithm to learn how to make accurate predictions on future examples.
This is in contrast to unsupervised machine learning where we don't have labels for the training data examples, and we'll cover unsupervised learning in a later part of this course.

Neural Networks

Perceptron

• It is the simplest neural network.
• It is a single layer neural network that uses one layer of a list of input features and one output.
• One of the features is a bias, same as intercept in linear regression, that gets combined a long with the other features.
• After having this linear combination, an activation function is applied. This function is usually non-linear and depends on the problem being tried to solve.

Neural network architecture

• Generally hard to interpret.
• Expensive to train, fast to predict
• Scikit-learn: sklearn.neural_network.MLPClassifier.
• Deep Learning Frameworks:
  • MXNet
  • TensorFlow
  • Caffe
  • PyTorch

Convolutional Neural Networks

• It is very useful in image analysis
• The input is an image or a sequence of images.
• Kernel is used as filter to extract local features.

Recurrent neural networks

• It consists of multiple inputs layer, multiple hidden layers and multiple output layer.
• Each node outputs to the next input node.

K-nearest neighbor

• It is a type of machine learning algorithm.
• It can be used for classification and regression.
• It is an example of what's called instance based or memory based supervised learning.
• It Non-parametric, instance based, lazy:
  • Non-parametric: Model is not defined by fixed set of parameters.
  • Instance-based or lazy learning: Model is the result of effectively memorizing training data.
• Requires keeping the original dataset.
• It returns a classifier that classifies the input with respect to the nearest n_neighbors neighbors that is the most predominant. K-nearest neighbors doesn't make a lot of assumptions about the structure of the data and gives potentially accurate, but sometimes unstable predictions
• Space complexity and prediction-time complexity grow with size of training data.
• Suffers from curse of dimensionality: points became increasingly isolated with more dimensions, for a fixed-size training dataset.
• It can be sensitive to small changes in the training data.

• It can be used in python as below:

  1. Initiate a variable
     

     from sklearn.neighbors import KNeighborsClassifier
     knn = KNeighborsClassifier(n_neighbors)
     

2. Train the model to memorize all its features and labels

    ```python
    knn.fit()
    ```

3. To predict a label use the below function with 1 parameter that has the same number of feature as the trained data

    ```python
    knn.predict(param)
    ```

4. The accuracy can be tested by passing testing data and testing labels

    ```python
    knn.score(X_test, y_test)
    ```

Linear Model

• Linear models make strong assumptions about the structure of the data.
• The target value can be predicted just using a weighted sum of the input variables, a linear function.
• It can get stable, but potentially inaccurate predictions.

1. Linear Regression
   
   The hat(^) is an indication that the parameter is estimated during training process.
   
   y: the predicted output.
   
   w_i: the model coefficients or feature weights.
   
   b: the biased term or intercept of the model.

   w, b parameters are estimated by:

- They are estimated from training the data.
- There are different methods correspond to different 'fit' criteria and goals and ways to control complexity.
- `Squared loss function` returns the squared difference between the target value and the  actual value as the penalty.
- The learning algorithm then computes or searches for the set of w, b parameters that optimize an objective function, typically to minimize the total of this loss function over all training points.

1. Least Squares:  
    - The most popular way to estimate w and b parameters is using what's called least-squares linear regression or ordinary least-squares. Least-squares finds the values of w and b that minimize the total sum of squared differences between the predicted y value and the actual y value in the training set. Or equivalently it minimizes the mean squared error of the model.
    - This technique is designed to find the slope, the w value, and the b value of the y intercept, that minimize this squared error, this mean squared error.
    - The mean squared error is the square difference between predicted and actual values, and then all these are added up, and then divided by the number of training points, take the average, that will be the mean squared error of the model.
    - One thing to note about this linear regression model is that there are no parameters to control the model complexity. No matter what the value of w and b, the result is always going to be a straight line. This is both a strength and a weakness of the model.

        ```python
        from sklearn.linear_model import LinearRegression

        X_train, X_test, y_train, y_test=train_test_split(X_R1, y_R1, random_state=0)

        linreg = LinearRegression().fit(X_train,y_train)

        # w_i: coefficients
        linreg.coef_
        # b: the intercept term
        linreg.intercept_

        #! In Scikit-Learn object attribute endswith an underscore, this means that thisattribute is derived from the trainingdata, not quantities that set by the user.
        ```

2. Ridge Regression:
    - Ridge regression uses the same least-squares criterion, but with one difference. During the training phase, it adds a penalty for large feature weights in w parameters.
    - Once the parameters are learned, its prediction formula is the same as ordinary least-squares.
    - The addition of a parameter penalty is called regularization. Regularization prevents over fitting by restricting the model, typically to reduce its complexity.
    - It uses L2 regularization: minimize sum of squares of w entries.
    - If ridge regression finds two possible linear models that predict the training data values equally well, it will prefer the linear model that has a smaller overall sum of squared feature weights.
    - The amount of regularization to apply is controlled by the alpha parameter. Larger alpha means more regularization and simpler linear models with weights closer to zero.(default 1.0)

        ```python
        from sklearn.preprocessing import MinMaxScaler
        scaler = MinMaxScaler()

        from sklearn.linear_model import Ridge
        X_train, X_test, y_train, y_test =train_test_split(X_crime, y_crime,random_state=0)

        X_train_scaled = scaler.fit_transform(X_train)
        X_test_scaled = scaler.transform(X_test)

        linridge = Ridge(alpha = 20.0).fit(X_train_scaled, y_train)
        ```

3. Lasso Regression
    - Like ridge regression, lasso regression adds a regularization penalty term to the ordinary least-squares objective, that causes the model W-coefficients to shrink towards zero.
    - Lasso regression is another form of regularized linear regression that uses an L1 regularization penalty for training (instead of ridge's L2 penalty).
    - L1 Penalty: minimizes the sum of the absolute values of the coefficients.
    - This has the effect of setting parameter weights in w to zero for the least influential variables. This called a sparse solution: a kind of feature selection.
    - The parameter alpha controls the amount of L1 regularization (default = 1.0).
    - The prediction formula is the same as ordinary least-squares.

        ```python
        from sklearn.preprocessing import MinMaxScaler
        scaler = MinMaxScaler()

        from sklearn.linear_model import Ridge
        X_train, X_test, y_train, y_test = train_test_split(X_crime, y_crime,random_state=0)

        X_train_scaled = scaler.fit_transfor(X_train)
        X_test_scaled = scaler.transform(X_test)

        linlasso = Lasso(alpha = 2.0, max_iter =10000) .fit(X-train_scaled, y_train)
        ```

        - When to use ridge vs lasso:
          - Many small/medium sized effects: use ridge.
          - Only a few variables with medium/large effect: use lasso.

1. Linear Classification
   1. Support Vector Machines (SVC):
      • Linear models are also used for classification, starting with binary classification.
      • This approach uses the same linear functional form as for regression. But instead of predicting a continuous target value, we take the output of the linear function and apply the sine function to produce a binary output with two possible values, corresponding to the two possible class labels.
      • One way to define a good classifier is to reward classifiers for the amount of separation that can provide between the two classes (classifier margin). The margin is the width that the decision boundary can be increased perpendicularly before hitting a data point. The classifier that has the maximum margin is called the Linear Support Vector Machine, also known as an LSVM or a support vector machine with linear kernel.
      • How tolerant the support vector machine is of misclassifying training points, as compared to its objective of minimizing the margin between classes is controlled by a regularization parameter called C which by default is set to 1.0. Larger values of C represent less regularization and will cause the model to fit the training set with these few errors as possible, even if it means using a small immersion decision boundary. Very small values of C on the other hand use more regularization that encourages the classifier to find a large marge on decision boundary, even if that decision boundary leads to more points being misclassified.

- **Linear Model Pros**:
  - Simple and easy to train.
  - Fast prediction.
  - Scales well to very large dataset.
  - Works well with sparse data.
  - Reasons for prediction are relatively easy to interpret.
- **Linear Model Cons**:
  - For lower-dimensional data, other models may have superior generalization performance.
  - For classification, data may not be linearly separable.

1. Logistic Regression

   • It is a kind of generalized linear model.
   • In spite of being called a regression measure, it is actually used for classification
   • like ordinary least squares and other regression methods, logistic regression takes a set input variables, the features, and estimates a target value.
   • Unlike ordinary linear regression, in it's most basic form logistic repressions target value is a binary variable instead of a continuous value.
   • There are flavors of logistic regression that can also be used in cases where the target value to be predicted is a multi class categorical variable, not just binary.
   • Logistic regression is similar to linear regression, but with one critical addition. The logistic regression model still computes a weighted sum of the input features xi and the intercept term b (like in linear regression), but it runs this result through a special non-linear function f, the logistic function represented by this new box in the middle of the diagram to produce the output y. The effect of applying the logistic function is to compress the output of the linear function so that it's limited to a range between 0 and 1. Below the diagram, you can see the formula for the predicted output y hat which first computes the same linear combination of the inputs xi, model coefficient weights wi hat and intercept b hat, but runs it through the additional step of applying the logistic function to produce y hat.
   • If we pick different values for b hat and the w hat coefficients, we'll get different variants of this s shaped logistic function, which again is always between 0 and 1.

   • To perform logistic, regression in Scikit-Learn, you import the logistic regression class from the sklearn.linear model module, then create the object and call the fit method using the training data just as you did for other class files like k nearest neighbors.
     

     from sklearn.linear_model import LogisticRegression
     
     X_train, X_test, y_train, y_test = train_test_split(X_C2,   y_C2,random_state = 0)
     clf = LogisticRegression(C=1).fit(X_train, y_train)
     

 - L2 regularization is 'on' by default (like ridge regression)
 - Parameter C controls amount of regularization (default 1.0)
 - As with regularized linear regression, it can be important to normalize all features so that they are on the same scale.

1. Kernelized Support Vector Machines (SVMs)

   • It is a very powerful extension of linear support vector machines, it can provide more complex models that can go beyond linear decision boundaries.
   • SVMs can be used for both classification and regression.
   • one way to think about what kernelized SVMs do, is they take the original input data space and transform it to a new higher dimensional feature space, where it becomes much easier to classify the transform to data using a linear classifier. (eg. instead of y(x) it became y(x,x^2) like polynomial feature). The above figure shows at the right that the points can be separated by a straight line after converting it to a two dimensional space, while on the left is the original one dimensional points in which the straight line is converted to a parabola.

   • An example of how it can be done using scikit-learn in Python.
     

     from sklearn.svm import SVC
     from adspy_shared_utilities import              plot_class_regions_for_classifier
     
     X_train, X_test, y_train, y_test = train_test_split(X_D2, y_D2, random_state = 0)
     
     # The default SVC kernel is radial basis function (RBF)
     #! SVC() = SVC(kernel = 'rbf', gamma=1, C=1)
     
     plot_class_regions_for_classifier(SVC().fit(X_train, y_train), X_train, y_train, None, None, 'Support Vector Classifier: RBF kernel')
     
     # Compare decision boundaries with polynomial kernel, degree = 3
     plot_class_regions_for_classifier(SVC(kernel = 'poly', degree = 3).fit(X_train, y_train), X_train, y_train, None, None, 'Support Vector Classifier: Polynomial kernel, degree = 3')
     

     • Calling the fit method with the training data to train the model.
     • There is an SVC parameter called kernel, that allows us to set the kernel function used by the SVM. The polynomial kernel takes additional parameter degree that controls the model complexity and the computational cost of this transformation.
     • Small gamma means a larger similarity radius (give broader, smoother decision regions). So that points farther apart are considered similar . Which results in more points being group together. Small values of gamma. While larger values of gamma give smaller, more complex decision regions, tightly constrained decision boundaries.
     • SVMs also have a regularization parameter, C, that controls the tradeoff between satisfying the maximum margin criterion to find the simple decision boundary, and avoiding misclassification errors on the training set. The C parameter is also an important one for kernelized SVMs, and it interacts with the gamma parameter.

• Pros:
  • Can perform well on a range of datasets.
  • Versatile: different kernel functions can be specified, or custom kernels can be defined for specific data types.
  • Works well for both low-and high-dimensional data.
• Cons:
  • Efficiency (runtime speed and memory usage) decreases as training set size increases (e.g. over 50000 samples).
  • Needs careful normalization of input data and parameter tuning.
  • Does not provide direct probability estimates (but can be estimated using e.g. Platt scaling).
  • Difficult to interpret why a prediction was made.

1. Decision tree

- It can be used for both regression and classification.
- It learns a series of explicit `if then` rules on feature values that result in a decision that predicts the target value. In this way any given object can be categorized as either matching the target object the first person is thinking of or not, according to its features as determined by asking the series of yes or no questions. We can form these questions into a tree with a node representing one question and the yes or no possible answers as the left and right branches from that node that connect the node to the next level of the tree. One question being answered at each level. At the bottom of the tree are nodes called leaf nodes that represent actual objects as the possible answers. For any object there's a path from the root of the tree to that object that is determined by the answers to the specific yes or no questions at each level.

    ```python
    from sklearn.datasets import load_iris
    from sklearn.tree import DecisionTreeClassifier
    from sklearn.model_selection import train_test_split

    iris = load_iris()

    X_train, X_test, y_train, y_test = train_test_split(iris.data, iris.target, random_state = 3)
    clf = DecisionTreeClassifier(max_depth = 3, min_samples_leaf = 8).fit(X_train, y_train)
    ```

    - max_depth: controls maximum depth (number of split points). Most common way to reduce tree complexity and overfitting.

    - min_samples_leaf: threshold for the minimum # of data instances a leaf can have to avoid further splitting.

    - max_leaf_nodes: limits total number of leaves in the tree.

    - In practice, adjusting only one of these (e.g. max_depth) is enough to reduce overfitting.



Over Fitting


There is a problem with decision tree which is overfitting, due to its complexity and essentially memorizing the training data.
One strategy to prevent overfitting is to prevent the tree from becoming really detailed and complex by stopping its growth early. This is called pre-pruning.
Another strategy is to build a complete tree with pure leaves but then to prune back the tree into a simpler form. This is called post-pruning or sometimes just pruning.



Feature Importance


Another way of analyzing the tree instead of looking at the whole tree at once is to do what's called a feature important calculation.
one of the most useful and widely used types of summary analysis you can perform on a supervised learning model.
typically a number between 0 and 1 that's assigned to an individual feature.
It indicates how important that feature is to the overall prediction accuracy.
A feature importance of zero means that the feature is not used at all in the prediction. A feature importance of one, means the feature perfectly predicts the target.
Typically, feature importance numbers are always positive and they're normalized so they sum to one.
In scikit-learn, feature importance values are stored as a list in an estimated property called feature_importances_.



Pros:


Easily visualized and interpreted.
No feature normalization or scaling typically needed.
Work well with datasets using a mixture of feature types (continuous, categorical, binary)



Cons:


Even after tuning, decision trees can often still overfit.
Usually need an ensemble of trees for better generalization performance.

ML Data Readiness

• ML data readiness is the capability to evaluate readiness, or worthiness, of datasets for use in an ML based predictive solution.

• ML data readiness evaluation is typically done prior to embarking on an ML project. It can help:

  • Evaluate the predictive potential of the dataset
  • Identify predictive outcomes that can be supported
  • Build initial ML models and understand relative performance

Productizing a ML Model

Aspects to consider

• Model Hosting
• Model deployment
• Pipeline to provide features vectors
• Code to provide low-latency and/or high-volume predictions
• Model and data updating and versioning
• Quality monitoring and alarming
• Data and model security and encryption
• Customer privacy, fairness, and trust
• Data provider contractual constraints (eg., attribution, cross-fertilization)

1. Overfitting

   Informally, overfitting typically occurs when we try to fit a complex model with an inadequate amount of training data. And overfitting model uses its ability to capture complex patterns by being great at predicting lots and lots of specific data samples or areas of local variation in the training set. But it often misses seeing global patterns in the training set that would help it generalize well on the unseen test set.

2. Underfitting

   The model is too simple for the actual trends that are present in the data. It doesn't even do well on the training data and thus, is not at all likely to generalize well to test data.

• To avoid these, the below points would help:
  1. First, try to draw the data with respect to the labels and try to figure out the relationship between, whether its linear, quadratic, polynomial and so on.
  2. Reduce the number of features.
     1. Manually select which features to keep.
     2. Use a model selection algorithm.
  3. Regularization
     1. Keep all the features, but reduce the magnitude of parameters.
     2. It works well when there are a lot of slightly useful features.

Types of Production environments

Batch predictions

• Useful if all possible inputs known a priori (eg., all product categories for which demand is to be forecast, all keywords to bid)
• Predictions can still be served real-time, simply read from pre-computed values

Online Predictions

• Useful if input space is large (e.g., customer's utterances or photos, detail pages to be translated)
• Low latency requirement (e., at most 100ms)

Online training

• Sometimes training data patterns change often, so need to train online (eh., fraud detection)

Business metrics vs Model Metrics

• Business metrics may not be the same as the performance metrics that are optimized during training. Why?
  • click-through rate
• Ideally, performance metrics are highly correlated with business metrics.

Storage

• Row-oriented formats:
  • comma/tab-separated values (CSV/TSV)
  • Read-only DB (RODB): Internal read-only file-based store with fast key-based access
  • Avro: allows schema evolution for Hadoop
• Column-oriented formats:
  • Parquet: Type-aware and indexed for Hadoop
  • Optimized row columnar (ORC): Type-aware, indexed, and with statistics for Hadoop
• User-defined formats:
  • JavaScript object notation (JSON): For key-value objects
  • Hierarchical data format 5 (HDF5): Flexible data model with chunks
• Compression can be applied to all formats
• Usual trad-offs: Read/write speeds, size, platform-dependency, ability for schema to evolve, schema/data separability, type richness

Model and Pipeline Persistence

• Predictive Model Markup Language (PMML):

  • Vendor-independent XML-based language for storing ML models
  • Support varies in different libraries:
  • KNIME (analytics/ML library): Full Support
  • Scikit-learn: Extensive support
  • Spark MLlib: Limited Support

• Custom methods:

  • Scikit-learn: Uses the Python pickle method to serialize/deserialize Python objects
  • Spark MLlib: Transformers and Estimators implement ML Writable
  • TensorFlow (deep learning library): Allows saving of MetaGraph
  • MxNet (deep learning library): Saves into JSON

Model Deployment

• It is the integration of the model and its resources into a production environment so that it can be used to create predictions.
• Technology transfer: Experimental framework may not suffice for production
  • A/B testing or shadow testing: Helps catch production issues early

Monitoring

• It is important to monitor quality metrics and business impacts with dashboards, alarms, user feedback, etc.:
  • The real-world domain may change over time.
  • The software environment may change.
  • High profile special cases may fail.
  • There may be a change in business goals.

Maintenance

• Performance deterioration may require acquire new tuning:
  • Changing goals may require new metrics.
  • A changing domain may require changes to validation set.
  • Your validation set may be replaced over time to avoid overfitting.

Common Mistakes

• You solved the wrong problem
• The data was flawed
• The solution didn't scale
• Final result doesn't match with the prototype's results.
• It takes too long to fail
• The solution was too complicated
• There weren't enough allocated engineering resources to try out long-term science ideas.
• There was a lack of a true collaboration.