End to end Machine Learning Project

Machine Learning Practice Course

Outline

               1. Steps in ML projects

               2. Illustration through practical set up

ML Project

  •  Excellent wine company wants to develop ML model for predicting wine quality on certain physiochemical characteristics in order to replace expensive quality sensor.

  •  Let's understand steps involved in addressing this problem. 

Steps in ML projects

  1.  Look at the big picture.
  2.  Get the data.
  3.  Discover and visualize the data to gain insights.
  4.  Prepare the data for Machine Learning algorithms.
  5.  Select a model and train it.
  6.  Fine-tune your model.
  7.  Present your solution.
  8.  Launch, monitor and maintain your system.

A few words of wisdom

  •  ML is usually a small piece in a big project. e.g. wine quality prediction is a small piece in setting up the manufacturing process.
  •  Typically 10-15% of time is spent on ML.
  •  A lot more time is spent on capturing and processing data needed for ML and taking decisions based on output of ML module. 
  • Needs strong collaboration with domain experts, product managers and eng-teams for successful execution. 

Step 1: Look at the big picture

  1. Frame the problem
  2. Select a performance measure
  3. List and check the assumptions

 1.1 Frame the problem

  •  What is input and output?
  •  What is the business objective? How does company expects to use and benefit from the model?  
    •  Useful in problem framing 
    •  Algorithm and performance measure selection
    •  Overall effort estimation
  •  What is the current solution (if any)?
    •  Provides a useful baseline

Design consideration in problem framing

  • Is this a supervised, unsupervised or a RL problem?
  • Is this a classification, regression or some other task?
  • What is the nature of the output: single or multiple outputs?
  • Does system need continuous learning or periodic updates?
  • What would be the learning style: batch or online?

1.2 Selection of performance measure

  • Regression
    • Mean Squared Error (MSE) or
    • Mean Absolute Error (MAE)
  • Classification
    • Precision
    • Recall
    • F1-score
    • Accuracy

1.3 Check the assumptions

  • List down various assumptions about the task.
  • Review with domain experts and other teams that plan to consume ML output.
  • Make sure all assumptions are reviewed and approved before coding!

Step 2: Get the data

  • Data spread across multiple tables, files or documents with access control.
  • Obtain appropriate access controls and authorizations.
  • Get familiarized with data by looking at schema and a few rows. (Familiarity with SQL would be useful here.)
import pandas as pd
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np

Load basic libraries

  • Let's first access our data - in this case, we need to download it from the web.
  • It's a good practice to create a function for downloading and extracting the data.
data_url = 'https://archive.ics.uci.edu/ml/machine-learning-databases/wine-quality/winequality-red.csv'
data = pd.read_csv(data_url, sep=";")

Now that the data is loaded, let's examine it.

2.1 Check data samples

Let's look at a few data samples with head() method.

data.head()

2.2 Features

It's a good idea to understand significance of each feature by consulting the experts.

Feature Significance
Fixed acidity Most acids involved with wine or fixed or nonvolatile (do not evaporate readily)
Volitile acidity The amount of acetic acid in wine, which at too high of levels can lead to an unpleasant, vinegar taste
Citric acid Found in small quantities, citric acid can add 'freshness' and flavor to wines
Residual sugar it's rare to find wines with less than 1 gram/liter and wines with greater than 45 grams/liter are considered sweet.
Chlorides The amount of salt in the wine.


 
Alcohol The percentage of alcohol contents in the wine.

 (Credits:https://www.kaggle.com/uciml/red-wine-quality-cortez-et-al-2009)

\vdots
\vdots
feature_list = data.columns[:-1].values
label = [data.columns[-1]]

print ("Feature list:", feature_list)
print ("Label:", label)
Feature list: ['fixed acidity' 'volatile acidity' 'citric acid' 'residual sugar' 'chlorides' 'free sulfur dioxide' 'total sulfur dioxide' 'density' 'pH' 'sulphates' 'alcohol'] Label: ['quality']

2.3 Data statistics

Let's use info() method to get quick description of data.

data.info()
  • Total entries: 1599 (Tiny dataset by ML standard)
  • There are total 12 columns: 11 features + 1 label
    • Label column: quality
    • Features: [fixed acidity, volitile acidity, citric acid, residual sugar, cholrides, free sulphur dioxide, total sulphur dioxide, density, pH, sulphates, alcohol]
  • All columns are numeric (float64) and label is an integer.

2.3 Data statistics

In order to understand nature of numeric attribites, we use describe() method.

data.describe()

This one prints count and statistical properties - mean, standard deviations and quartiles.

  • The wine quality can be between 0 and 10, but in this dataset, the quality values are between 3 and 8. Let's look at the distribution of examples by the wine quality.
data['quality'].value_counts()
  • High quality value \( \rightarrow \) better quality of wine
  • You can see that there are lots of samples of average wines than good or the poor quality ones.
    • Many examples with quality = 5 or 6

The information can be viewed through histogram plot.

  • A Histogram gives the count of how many samples occurs within a specific range (bins).
  • The x-axis denotes the range of values in a feature and
  • The y-axis denotes the frequency of samples with those specific values.
sns.set()
data.quality.hist()
plt.xlabel('Wine Quality')
plt.ylabel('Count')

Note taller bars for quality 5 and 6 compared to the other qualities.

In a similar manner, we can plot all numerical attributes with histogram plot for quick examination.

A few observations based on these plots:

  1. Features are at different scales.
  2. Features have different distributions -
    • A few are tail heavy. e.g. residual sugar, free so2
    • A few have  multiple modes. e.g. volitile acidity, citric acid

Before any further exploration, it's a good idea to separate test set and do not look at it in order to have a clean evaluation set.

2.4 Create test set

  • When we look at the test set, we are likely to notice patterns in that and based on that we may select certain models.
  • This leads to biased estimation on test set, which may not generalize well in practice. This is called data snooping bias.

Let's write a function to split the data into training and test. Make sure to set the seed so that we get the same test set in the next run.

def split_train_test(data, test_ratio):
  # set the random seed.
  np.random.seed(42)

  # shuffle the dataset.
  shuffled_indices = np.random.permutation(len(data))

  # calculate the size of the test set.
  test_set_size = int(len(data) * test_ratio)

  # split dataset to get training and test sets.
  test_indices = shuffled_indices[:test_set_size]
  train_indices = shuffled_indices[test_set_size:]
  return data.iloc[train_indices], data.iloc[test_indices]
train_set, test_set = split_train_test(data, 0.2)

Scikit-Learn provides a few functions for creating test sets based on               

  1. Random sampling, which randomly selects \(k\)% points in the test set.         
  2. Stratified sampling, which samples test examples such that they are representative of overall distribution.

Random sampling

  • train_test_split() function performs random sampling with

    • random_state parameter to set the random seed, which ensures that the same examples are selected for test sets across runs.
    • test_size parameter for specifying size of the test set.
    • shuffle flag to specify if the data needs to be shuffled before splitting.

  • Provision for processing multiple datasets with an identical number of rows and selecting the same indices from these datasets.

    • Useful when labels are in different dataframe.
from sklearn.model_selection import train_test_split

We can read the documentation for this function by using the following line of code:

?train_test_split
from sklearn.model_selection import train_test_split

Let's perform random sampling on our dataset:

train_set, test_set = train_test_split(data, test_size=0.2, random_state=42)

Stratified sampling

  • Data distribution may not be uniform in real world data.
  • Random sampling - by its nature - introduces biases in such data sets.

Recall the label distribution in our dataset: It's not uniform!

sns.set()
data.quality.hist()
plt.xlabel('Wine Quality')
plt.ylabel('Count')
  • Many examples of class 5 and 6 compared to the other classes.
  • This causes a problem while random sampling. The test distribution may not match with the overall distribution.

How do we sample in such cases?

  • We divide the population into homogenous groups called strata.
  • Data is sampled from each stratum so as to match it with the overall data distribution.
  • Scikit-Learn provides a class StratifiedShuffleSplit that helps us in stratified sampling.
from sklearn.model_selection import StratifiedShuffleSplit
split = StratifiedShuffleSplit(n_splits=1, test_size=0.2, random_state=42)
for train_index, test_index in split.split(data, data["quality"]):
  strat_train_set = data.loc[train_index]
  strat_test_set = data.loc[test_index]

Let's examine the test set distribution by the wine quality that was used for stratified sampling.

strat_dist = strat_test_set["quality"].value_counts() / len(strat_test_set)

Now compare this with the overall distribution:

overall_dist = data["quality"].value_counts() / len(data)

Let's look at them side-by-side:

dist_comparison = pd.DataFrame({'overall': overall_dist, 'stratified': strat_dist})
dist_comparison['diff(s-o)'] = dist_comparison['stratified'] - dist_comparison['overall']
dist_comparison['diff(s-o)_pct'] = 100*(dist_comparison['diff(s-o)']/dist_comparison['overall'])

You can notice that there is a small difference in most strata.

dist_comparison
random_dist = test_set["quality"].value_counts() / len(test_set) 
random_dist

Let's contrast this with random sampling:

Sampling bias comparison

    Compare the difference in distribution of stratified and uniform sampling:

  • Stratified sampling gives us test distribution closer to the overall distribution than the random sampling.
dist_comparison.loc[:, ['diff(s-o)_pct', 'diff(r-o)_pct']]

Step 3: Data visualization

  • Performed on training set.

  • In case of large training set -

    • Sample examples to form exploration set.

  • Enables to understand features and their relationship among themselves and with output label.

In our case, we have a small training data and we use it all for data exploration. There is no need to create a separate exploration set.

It's a good idea to create a copy of the training set so that we can freely manipulate it without worrying about any manipulation in the original set.

exploration_set = strat_train_set.copy()

Scatter Visualization

With seaborn library:

sns.scatterplot(x='fixed acidity', y='density', hue='quality',
                data=exploration_set)

With matplotlib:

exploration_set.plot(kind='scatter', x='fixed acidity', y='density', alpha=0.5,
                     c="quality", cmap=plt.get_cmap("jet"))

Relationship between features

  • Standard correlation coefficient between features.
    • Ranges between -1 to +1
      • Correlation = +1: Strong positive correlation between features
      • Correlation = -1: Strong negative correlation between features
      • Correlation = 0: No linear correlation between features
    • Visualization with heat map
  • Only captures linear relationship between features.
    • For non-linear relationship, use rank correlation

Let's calculate correlations between our features.

corr_matrix = exploration_set.corr()

Let's check features that are correlated with the label, which is quality in our case.

corr_matrix['quality']

Notice that quality has strong positive correlation with alcohol content [0.48] and strong negative correlation with volitile acidity [-0.38].

Let's visualize correlation matrix with heatmap:

plt.figure(figsize=(14,7))
sns.heatmap(corr_matrix, annot=True)

You can notice:

  • The correlation coefficient on diagonal is +1.
  • Darker colors represent negative correlations, while fainter colors denote positive correlations. For example
    • citric acid and fixed acidity have strong positive correlation.
    • pH and fixed acidity have strong negative correlation.

Another option to visualize the relationship between the feature is with scatter matrix.

 

from pandas.plotting import scatter_matrix
attribute_list = ['citric acid', 'pH', 'alcohol', 'sulphates', 'quality']
scatter_matrix(exploration_set[attribute_list])

For convenience of visualization, we show it for a small number of attributes.

  • Similar analysis can be carried out with combined features - features that are derived from the original features.

Note of wisdom

  1. Visualization and data exploration do not have to be absolutely thorough.
  2. Objective is to get quick insight into features and its relationship with other features and labels.
  3. Exploration is an iterative process: Once we build model and obtain more insights, we can come back to this step.

Step 4: Prepare data for ML algorithm

We often need to preprocess the data before using it for model building due to variety of reasons:

Typical steps in data preprocessing are as follows:

It's a good practice to make a copy of the data and apply preprocessing on that copy. This ensures that in case something goes wrong, we will at least have original copy of the data intact.

  • Due to errors in data capture, data may contain outliers or missing values.
  • Different features may be at different scales.
  • The current data distribution is not exactly amenable to learning.
  1. Separate features and labels.
  2. Handling missing values and outliers.
  3. Feature scaling to bring all features on the same scale.
  4. Applying certain transformations like log, square root on the features.

4.1 Separate features and labels from the training set.

# Copy all features leaving aside the label.
wine_features = strat_train_set.drop("quality", axis=1)

# Copy the label list
wine_labels = strat_train_set['quality'].copy()

4.2 Data cleaning

Let's first check if there are any missing values in feature set: One way to find that out is column-wise.

wine_features.isna().sum() # counts the number of NaN in each column of wine_features

In this dataset, we do not have any missing values.

In case, we have non-zero numbers in any columns, we have a problem of missing values.

  • These values are missing due to errors in recording or they do not exist.
  • If they are not recorded:
    • Use imputation technique to fill up the missing values.
    • Drop the rows containing missing values.
  • If they do not exists, it is better to keep it as NaN.

Sklearn provides the following methods to drop rows containing missing values:

  • dropna()
  • drop()

It provides SimpleImputer class for filling up missing values with. say, median value.

from sklearn.impute import SimpleImputer
imputer = SimpleImputer(strategy="median")

The strategy contains instructions as how to replace the missing values. In this case, we specify that the missing value should be replaced by the median value.

imputer.fit(wine_features)

In case, the features contains non-numeric attributes, they need to be dropped before calling the fit method on imputer object.

SimpleImputer(add_indicator=False, copy=True, fill_value=None, missing_values=nan, strategy='median', verbose=0)

Let's check the statistics learnt by the imputer on the training set:

imputer.statistics_
array([ 7.9 , 0.52 , 0.26 , 2.2 , 0.08 , 14. , 39. , 0.99675, 3.31 , 0.62 , 10.2 ])

Note that these are median values for each feature. We can cross-check it by calculating median on the feature set:

wine_features.median()

Finally we use the trained imputer to transform the training set such that the missing values are replaced by the medians:

tr_features = imputer.transform(wine_features)

This returns a Numpy array and we can convert it to the dataframe if needed:

tr_features.shape
(1279, 11)
wine_features_tr = pd.DataFrame(tr_features, columns=wine_features.columns)

4.3 Handling text and categorical attributes

4.3.1 Converting categories to numbers:

from sklearn.preprocessing import OrdinalEncoder
ordinal_encoder = OrdinalEncoder()
  • Call fit_transform() method on ordinal_encoder object to convert text to numbers.
  • The list of categories can be obtained via categories_ instance variable.

One issue with this representation is that the ML algorithm would assume that the two nearby values are closer than the distinct ones.

4.3.2 Using one hot encoding

  • Here we create one binary feature per category - the feature value is 1 when the category is present else it is 0.
  • Only one feature is 1 (hot) and the rest are 0 (cold).
  • The new features are referred to as dummy features.
  • Scikit-Learn provides a OneHotEncoder class to convert categorical values into one-hot vectors.
from sklearn.preprocessing import OneHotEncoder
cat_encoder = OneHotEncoder()
  • We need to call fit_transform() method on OneHotEncoder object.
  • The output is a SciPy sparse matrix rather than NumPy array. This enables us to save space when we have a huge number of categories.
  • In case we want to convert it to dense representation, we can do so with toarray() method.
  • The list of categories can be obtained via categories_ instance variable.
  • As we observed that when the number of categories are very large, the one-hot encoding would result in a very large number of features.
  • This can be addressed with one of the following approaches:
    • Replace with categorical numerical features
    • Convert into low-dimensional learnable vectors called embeddings

4.4 Feature Scaling

  • Most ML algorithms do not perform well when input features are on very different scales.
  • Scaling of target label is generally not required.

4.5.1 Min-max scaling or Normalization

  • We subtract minimum value of a feature from the current value and divide it by the difference between the minimum and the maximum value of that feature.
  • Values are shifted and scaled so that they range between 0 and 1.
  • Scikit-Learn provides MinMaxScalar transformer for this.
  • One can specify hyperparameter feature_range to specify the range of the feature.

4.5.2 Standardization

  • We subtract mean value of each feature from the current value and divide it by the standard deviation so that the resulting feature has a unit variance.
  • While normalization bounds values between 0 and 1, standardization does not bound values to a specific range.
  • Standardization is less affected by the outliers compared to the normalization.
  • Scikit-Learn provides StandardScalar transformation for feature standardization.
  • Note that all these transformers are learnt on the training data and then applied on the training and test data to tranform them.
  • Never learn these transformers on the full dataset.

Transformation Pipeline

  • Scikit-Learn provides a Pipeline class to line up transformations in an intended order.
  • Here is an example pipeline:
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
transform_pipeline = Pipeline([
                               ('imputer', SimpleImputer(strategy="median")),
                               ('std_scaler', StandardScaler()),])
wine_features_tr = transform_pipeline.fit_transform(wine_features)

Let's understand what is happening here:

  • Pipeline has a sequence of transformations - missing value imputation followed by standardization.
  • Each step in the sequence is defined by name, estimator pair.
  • Each name should be unique and should not contain __ (double underscore).
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import StandardScaler
transform_pipeline = Pipeline([
                               ('imputer', SimpleImputer(strategy="median")),
                               ('std_scaler', StandardScaler()),])
wine_features_tr = transform_pipeline.fit_transform(wine_features)
  • The output of one step is passed on the next one in sequence until it reaches the last step.
    • Here the pipeline first performs imputation of missing values and its result is passed for standardization.
  • The pipeline exposes the same method as the final estimator.
    • Here StandardScalar is the last estimator and since it is a transformer, we call fit_transform() method on the Pipeline object.

How to transform mixed features?

  • The real world data has both categorical as well as numerical features and we need to apply different transformations to them.
  • Scikit-Learn introduced ColumnTransformer for this purpose.
from sklearn.compose import ColumnTransformer

In our dataset, we do not have features of mixed types. All our features are numeric.

 

num_attribs = list(wine_features)
cat_attribs = ["place_of_manufacturing"]
full_pipeline = ColumnTransformer([
 ("num", num_pipeline, num_attribs),
 ("cat", OneHotEncoder(), cat_attribs),
 ])
wine_features_tr = full_pipeline.fit_transform(wine_features)

For the illustration purpose, here is an example code snippet:

  • Here we apply num_pipeline on numerical features and OneHotEncoder transformation on the categorical features.
  • The ColumnTransformer applies each transformation to the appropriate columns and then concatenates the outputs along the columns.
  • Note that all transformers must return the same number of rows.
  • The numeric transformers return dense matrix while the categorical ones return sparse matrix. The ColumnTransformer automatically determines the type of the output based on the density of the resulting matrix.

Step 5: Select and train ML model

  • It's a good practice to build a quick baseline model on the preprocessed data and get an idea about model performance.
from sklearn.linear_model import LinearRegression

lin_reg = LinearRegression()
lin_reg.fit(wine_features_tr, wine_labels)
LinearRegression(copy_X=True, fit_intercept=True, n_jobs=None, normalize=False)

Now that we have a working model of a regression, let's evaluate performance of the model on training as well as test sets.

  • For regression models, we use mean squared error as an evaluation measure.
from sklearn.metrics import mean_squared_error

quality_predictions =  lin_reg.predict(wine_features_tr)
mean_squared_error(wine_labels, quality_predictions)
0.4206571060060278

Let's evaluate performance on the test set.

  • We need to first apply transformation on the test set and then apply the model prediction function.
# copy all features leaving aside the label.
wine_features_test = strat_test_set.drop("quality", axis=1)

# copy the label list
wine_labels_test = strat_test_set['quality'].copy()

# apply transformations
wine_features_test_tr = transform_pipeline.fit_transform(wine_features_test)

# call predict function and calculate MSE.
quality_test_predictions = lin_reg.predict(wine_features_test_tr)
mean_squared_error(wine_labels_test, quality_test_predictions)
0.39759130875015164

Let's visualize the error between the actual and predicted values.

plt.scatter(wine_labels_test, quality_test_predictions)
plt.plot(wine_labels_test, wine_labels_test, 'r-')
plt.xlabel('Actual quality')
plt.ylabel('Predicted quality')

The model seem to be making errors on the best and poor quality wines.

Let's try another model: DecisionTreeRegressor.

from sklearn.tree import DecisionTreeRegressor
tree_reg = DecisionTreeRegressor()
tree_reg.fit(wine_features_tr, wine_labels)
DecisionTreeRegressor(ccp_alpha=0.0, criterion='mse', max_depth=None, max_features=None, max_leaf_nodes=None, min_impurity_decrease=0.0, min_impurity_split=None, min_samples_leaf=1, min_samples_split=2, min_weight_fraction_leaf=0.0, presort='deprecated', random_state=None, splitter='best')

Notice similarity between two code snippets.

Linear regression Decision Trees
lin_reg.fit(wine_features_tr, wine_labels) tree_reg.fit(wine_features_tr, wine_labels) 
quality_predictions =  tree_reg.predict(wine_features_tr)
mean_squared_error(wine_labels, quality_predictions)
0.0
quality_test_predictions = tree_reg.predict(wine_features_test_tr)
mean_squared_error(wine_labels_test, quality_test_predictions)
0.58125

Note that the training error is 0, while the test error is 0.58. This is an example of an overfitted model.

plt.scatter(wine_labels_test, quality_test_predictions)
plt.plot(wine_labels_test, wine_labels_test, 'r-')
plt.xlabel('Actual quality')
plt.ylabel('Predicted quality')

We can use cross-validation (CV) for robust evaluation of model performance.

from sklearn.model_selection import cross_val_score
  • Cross validation provides a separate MSE for each validation set, which we can use to get a mean estimation of MSE as well as the standard deviation, which helps us to determine how precise is the estimate.
  • The additional cost we pay in cross validation is additional training runs, which may be too expensive in certain cases.
def display_scores(scores):
  print("Scores:", scores)
  print("Mean:", scores.mean())
  print("Standard deviation:", scores.std())

Linear Regression CV

scores = cross_val_score(lin_reg, wine_features_tr, wine_labels,
                         scoring="neg_mean_squared_error", cv=10)
lin_reg_mse_scores = -scores
display_scores(lin_reg_mse_scores)
Scores: [0.56364537 0.4429824 0.38302744 0.40166681 0.29687635 0.37322622 0.33184855 0.50182048 0.51661311 0.50468542] 
Mean: 0.431639217212196 
Standard deviation: 0.08356359730413976

Decision tree CV

scores = cross_val_score(tree_reg, wine_features_tr, wine_labels,
                         scoring="neg_mean_squared_error", cv=10)
tree_mse_scores = -scores
display_scores(tree_mse_scores)
Scores: [0.6171875 0.6875 0.6328125 0.5078125 0.4609375 0.640625 0.65625 0.7109375 0.859375 1.07874016] 
Mean: 0.6852177657480315 
Standard deviation: 0.16668343331737054

Let's compare scores of Linear regression (LinReg) and decision tree (DT)regressions:

  • LinReg has better MSE and more precise estimation compared to DT.

Random forest CV

  • Random forest model builds multiple decision trees on randomly selected features and then average their predictions.
  • Building a model on top of other model is called ensemble learning, which is often used to improve performance of ML models.
from sklearn.ensemble import RandomForestRegressor

forest_reg = RandomForestRegressor()
forest_reg.fit(wine_features_tr, wine_labels)

scores = cross_val_score(forest_reg, wine_features_tr, wine_labels,
                         scoring="neg_mean_squared_error", cv=10)
forest_mse_scores = -scores
display_scores(forest_mse_scores)
Scores: [0.36989922 0.41363672 0.29063438 0.31722344 0.21798125 0.30233828 0.27124922 0.38747344 0.42379219 0.46229449] 
Mean: 0.34565226131889765 
Standard deviation: 0.0736322184302973
quality_test_predictions = forest_reg.predict(wine_features_test_tr)
mean_squared_error(wine_labels_test, quality_test_predictions)
0.34449875
plt.scatter(wine_labels_test, quality_test_predictions)
plt.plot(wine_labels_test, wine_labels_test, 'r-')
plt.xlabel('Actual quality')
plt.ylabel('Predicted quality')

Random forest looks more promising than the other two models.

  • It's a good practice to build a few such models quickly without tuning their hyperparameters and shortlist a few promising models among them.
  • Also save the models to the disk in Python pickle format.

What to do next?

Model diagnosis
 		 Remedy
Underfitting  Models with more capacity
Less constraints/regularization
Overfitting More data
Simpler model
More constraints/regularization

Step 6: Finetune your model

  • Usually there are a number of hyperparameters in the model, which are set manually.
  • Tuning these hyperparameters lead to better accuracy of ML models.
  • Finding the best combination of hyperparameters is a search problem in the space of hyperparameters, which is huge.

Grid search

  • Scikit-Learn provives a class GridSearchCV that helps us in this pursuit.
from sklearn.model_selection import GridSearchCV
  • We need to specify a list of hyperparameters along with the range of values to try.
  • It automatically evaluates all possible combinations of hyperparameter values using cross-validation.

For example, there are number of hyperparameters in RandomForest regression such as:

  • Number of estimators
  • Maximum number of features
param_grid = [
 {'n_estimators': [3, 10, 30], 'max_features': [2, 4, 6, 8]},
 {'bootstrap': [False], 'n_estimators': [3, 10], 'max_features': [2, 3, 4]},
]

Here the parameter grid contains two combinations:

  1. The first combination contains n_estimators with 3 values and max_features with 4 values.
  2. The second combination has an additional bootstrap parameter, which is set to False. Note that it was set to its default value, which is True, in the first grid.

Let's compute the total combinations evaluated here:

Let's create an object of GridSearchCV:

grid_search = GridSearchCV(forest_reg, param_grid, cv=5, 
                           scoring='neg_mean_squared_error',
                           return_train_score=True)

  1. The first one results in \( 3 \times 4 =12 \) combinations.

  2. The second one has 2 values of n_estimators and 3 values of max_features, thus resulting \( 2 \times 3 =6 \) in total of values.

The total number of combinations evaluated by the parameter grid \( 12 + 6 =18 \)

Let's create an object of GridSearchCV:

grid_search = GridSearchCV(forest_reg, param_grid, cv=5, 
                           scoring='neg_mean_squared_error',
                           return_train_score=True)
  • In this case, we set cv=5 i.e. using 5 fold cross validation for training the model.
  • We need to train the model for 18 parameter combinations and each combination would be trained 5 times as we are using cross-validation here.
  • The total model training runs = \( 18 \times 5 = 90 \)

Let's launch the hyperparameter search:

grid_search.fit(wine_features_tr, wine_labels)

The best parameter combination can be obtained as follows:

 grid_search.best_params_
{'max_features': 6, 'n_estimators': 30}

Let's find out the error at different parameter settings:

cvres = grid_search.cv_results_
for mean_score, params in zip(cvres["mean_test_score"], cvres["params"]):
  print(-mean_score, params)
cvres = grid_search.cv_results_
for mean_score, params in zip(cvres["mean_test_score"], cvres["params"]):
  print(-mean_score, params)

As you can notice the lowest MSE is obtained for the best parameter combination.

 

Let's obtain the best estimator as follows:

grid_search.best_estimator_

Note: GridSearchCV is initialized with refit=True option, which retrains the best estimator on the full training set. This is likely to lead us to a better model as it is trained on a larger dataset.

Randomized Search

  • When we have a large hyperparameter space, it is desirable to try RandomizedSearchCV.
  • It selects a random value for each hyperparameter at the start of each iteration and repeats the process for the given number of random combinations.
  • It enables us to search hyperparameter space with appropriate budget control.
from sklearn.model_selection import RandomizedSearchCV

Analysis of best model and its errors

Analysis of the model provides useful insights about features. let's obtain the feature importance as learnt by the model:

 feature_importances = grid_search.best_estimator_.feature_importances_
sorted(zip(feature_importances, feature_list), reverse=True)
  • Based on this information, we may drop features that are not so important.
  • It is also useful to analyze the errors in prediction and understand its causes and fix them.

Evaluation on test set

Now that we have a reasonable model, we evaluate its performance on the test set. The following steps are involved in the process:

# copy all features leaving aside the label.
wine_features_test = strat_test_set.drop("quality", axis=1)

# copy the label list
wine_labels_test = strat_test_set['quality'].copy()

# apply transformations
wine_features_test_tr = transform_pipeline.fit_transform(wine_features_test)
  1. Transform the test features.

2. Use the predict method with the trained model and the test set.

quality_test_predictions = grid_search.best_estimator_.predict(
    wine_features_test_tr)

3.Compare the predicted labels with the actual ones and report the evaluation metrics.

mean_squared_error(wine_labels_test, quality_test_predictions)
0.35345138888888883
from scipy import stats
confidence = 0.95
squared_errors = (quality_test_predictions - wine_labels_test) ** 2
stats.t.interval(confidence, len(squared_errors) - 1,
                 loc=squared_errors.mean(),
                 scale=stats.sem(squared_errors))
(0.29159276569581916, 0.4153100120819586)

4.It's a good idea to get 95% confidence interval of the evaluation metric. It can be obtained by the following code:

Step 7: Present your solution

Once we have satisfactory model based on its performance on the test set, we reach the prelaunch phase.

Before launch,

  1. We need to present our solution that highlights learnings, assumptions and systems limitation.
  2. Document everything, create clear visualizations and present the model.
  3. In case, the model does not work better than the experts, it may still be a good idea to launch it and free up bandwidths of human experts.

Step 8: Launch, monitor and maintain your system

Launch

  • Plug in input sources and
  • Write test cases

Monitoring

  • System outages
  • Degradation of model performance
  • Sampling predictions for human evaluation
  • Regular assessment of data quality, which is critical for model performance

Maintenance

  • Train model regularly every fixed interval with fresh data.
  • Production roll out of the model.

Summary

In this module, we studied steps involved in end to end machine learning project with an example of prediction of wine quality.

x_1
x_2
x_3
x_{100}
\vdots
\vdots
\vdots
w_1
w_{100}
w_2
\hat{y}
a
b
c
d

MLP_Week1

By Arun Prakash

MLP_Week1

  • 429

More from Arun Prakash