About AutoML
The following sections focus on the low-level details on the AutoML provider. For general information about choosing AutoML as the provider for IntergratedML, see AutoML.
AutoML is an automated machine learning system developed by InterSystems, housed within InterSystems IRIS® data platform. It is designed to quickly build accurate predictive models using your data, automating several key components of the machine learning process:
After you train your model with AutoML, you can easily deploy your model by using the SQL syntax provided by IntegratedML.
Key Features
Natural Language Processing
AutoML leverages natural language processing (NLP) to turn your text features into numeric features for your predictive models. AutoML uses Term frequency-inverse document frequency (TFIDF) to evaluate key words in text and list columns.
Multi-Hot Encoding
While most of our data is sparse, machine learning algorithms can only understand dense data. In most data modeling workflows, data scientists are burdened with performing difficult, and cumbersome, manual transformations to convert their sparse data into dense data.
Unlike many workflows that require this manual step, AutoML performs this conversion seamlessly. Lists and one-to-many relationships are smartly “multi-hot encoded” to account for columns that are representing more than a single value.
For instance, assume a table that contains a list of medical conditions for each person:
| Person | Conditions |
|---|---|
| Person A | [‘diabetes’, ‘osteoporosis’, ‘asthma’] |
| Person B | [‘osteoporosis’, ‘hypertension’] |
| Person C | [‘asthma’, ‘hypertension’] |
| Person D | [‘hypertension’, ‘asthma’] |
Many machine learning functions treat these lists as separate entities, with one-hot encoding resulting in the following conversion:
| Person | [‘diabetes’, ‘osteoporosis’, ‘asthma’] | [‘osteoporosis’, ‘hypertension’] | [‘asthma’, ‘hypertension’] | [‘hypertension’, ‘asthma’] |
|---|---|---|---|---|
| Person A | 1 | 0 | 0 | 0 |
| Person B | 0 | 1 | 0 | 0 |
| Person C | 0 | 0 | 1 | 0 |
| Person D | 0 | 0 | 0 | 1 |
Instead, AutoML uses bag-of-words to create a separate column for each value in each list:
| Person | ‘diabetes’ | ‘osteoporosis’ | ‘asthma’ | ‘hypertension’ |
|---|---|---|---|---|
| Person A | 1 | 1 | 1 | 0 |
| Person B | 0 | 1 | 0 | 1 |
| Person C | 0 | 0 | 1 | 1 |
| Person D | 0 | 0 | 1 | 1 |
While other functions would have treated each person as having a separate list of medical conditions, AutoML’s method allows a model to properly find patterns between each of these persons’ set of medical conditions.
AutoML assumes that order does not matter. Person C and Person D share the same set of medical conditions, but just ordered differently. While other functions treat those two lists differently, AutoML identifies that they are the same.
Feature Engineering
AutoML performs two key steps of feature engineering:
These steps help make the data compatible with the utilized machine learning models, and can greatly improve performance.
Column Type Classification
AutoML first examines the columns in the dataset and classifies them as a particular Python data type. For information about the conversion from DDL to Python data types, see DDL to Python Type Conversion.
The column types, along with how their classifications are made, are listed below:
Numeric columns are those that have the numeric pandas datatype, including int8, int64, float32, etc. All columns meeting this condition are included, except:
Columns of the timedelta datatype.
Columns with only one unique value.
Some columns with seemingly numeric data may be inappropriately classified as numeric columns. For example, an ID number of 1000 is not “half of” an ID number of 2000. You can properly treat these columns as category columns by recasting the numeric data with VARCHAR values.
Category columns are those that contain categorical values, meaning there are a relatively small, fixed number of values that appear. They satisfy the following criteria:
Must be of category or object pandas datatype.
Must not include Ignored Columns.
Must not include List Columns.
The number of unique values is less than 10% the total number of values.
Text columns are columns where the values look like sentences. AutoML looks for values that contain 4 or more words. They satisfy the following criteria:
Must be of category or object pandas datatype.
Must not include Ignored Columns.
Must not include Category Columns.
Must not include List Columns.
The number of unique values is less than 10% the total number of values.
List columns are those that contain list values. They satisfy the following criteria:
Must be of category or object pandas datatype.
Must not include Ignored Columns.
Must be, or contain, one of the following types:
InterSystems IRIS data type %Library.String:list
InterSystems IRIS data type %Library.String:array
Python list. This is determined by checking the first 10 non-empty values of the column to see if the type of each value is a Python list.
String array. This is determined by checking the first 10 non-empty values of the column to see if the type of each value is a string, with starting character [, ending character ], and of length at least 2.
Boolean columns are those that have the bool pandas datatype. They additionally satisfy the condition that they do not include Ignored Columns.
Ignored columns are those that are to be disregarded and removed before training. These include:
The ID column.
The label column.
Columns with only one unique value (except for columns of datetime pandas datatype).
Date/Time columns are those that have the datetime pandas datatype. They additionally satisfy the condition that they do not include Ignored Columns.
See below for discussion of additional date/time columns created.
DDL to Python Type Conversion
The following table maps DDL data types to the Python data types that AutoML uses to classify data columns.
| DDL Data Type | Python Data Type |
|---|---|
| BIGINT | integer |
| BINARY | bytes |
| BIT | Boolean |
| DATE | datetime64 (numpy) |
| DECIMAL | decimal |
| DOUBLE | float |
| INTEGER | integer |
| NUMERIC | float |
| REAL | float |
| SMALLINT | integer |
| TIME | datetime64 (numpy) |
| TIMESTAMP | datetime64 (numpy) |
| TINYINT | integer |
| VARBINARY | bytes |
| VARCHAR | string |
For information about DDL data types, and their associated InterSystems IRIS® data types, see “Data Types” in the InterSystems SQL Reference.
Data Transformation
The Transform Function transforms the entire dataset into the form to be used by the machine learning models. It is applied on the training set before training, and on any future datasets before predictions are made.
Additional Date/Time columns are created. For every datetime column, the following separate columns are added whenever applicable:
Hour of day.
Day of week.
Month of year.
AutoML also creates duration columns. Each column added represents one of the original date/time columns, and each value in this column is the duration between the dates of that particular date/time column and all other date/time columns. For example, consider patient data that has three date/time columns:
Date of birth.
Time of admission.
Time of exit.
AutoML creates two useful duration columns from these columns: age (duration between date of birth and time of admission) and length of stay (duration between time of admission and exit).
Finally, for each list column present, another column is added simply with the size of the lists. That is, each value in the new column is the length of the corresponding list in the old column.
Datasets can often be incomplete, with missing values in some of their columns. To help compensate for this and improve performance, AutoML fills in missing/NULL values:
For categorical and date columns, AutoML replaces missing values with the mode (most popular value) of the column.
For numeric and duration columns, AutoML replaces missing values with the mean (average) of the column.
For list and text columns, AutoML replaces missing values with an empty string.
For each numeric column, a standard scalar is fit. These include the original numeric columns, along with the duration and list size columns as well.
Numerical column values are also binned and then used as categorical values. These new categorical bin columns are added on separately in addition to the already present numerical columns. Each numerical column is separated into four bins, each representing a quartile of the values in that column. The new binned columns are treated as categorical columns.
For each text and list column, a vectorizer is fit to transform the data to the appropriate form needed for training. This is done with SciKit Learn’s TFIDF Vectorizer. Please see their documentation
Opens in a new window.
The following parameters are used:
| Parameter | Value |
|---|---|
| Convert to lowercase | True |
| Stop Words | None |
| N-Gram Range | (1,1) |
| Max Features | 10000 |
| Norm | L2 |
Binary columns are simply transformed to be composed of 1’s and 0’s, with true values mapping to 1’s.
Categorical columns are one-hot encodedOpens in a new window before being used for training.
As the last step before training, feature elimination is performed to remove redundancy, improve training speed, and improve the accuracy of models. This is done using Scikit Learn’s SelectFPROpens in a new window function.
The following parameters are used:
| Parameter | Value |
|---|---|
| Scoring function | f_classif |
| alpha | 0.2 |
Algorithms Used
AutoML uses four different models to make predictions.
For regression models:
For classification models:
XGBRegressor
For regression models, AutoML uses XGBoost’s XGBRegressorOpens in a new window class.
The model hyperparameters are detailed below:
| Hyperparameter | Value |
|---|---|
| Max Depth | 3 |
| Learning Rate | 0.1 |
| Number of Estimators | 100 |
| Objective | Squared Error |
| Booster | Gbtree |
| Tree Method | Auto |
| Number of Jobs | 1 |
| Gamma (min loss reduction for partition on leaf) | 0 |
| Min Child Weight | 1 |
| Max Delta Step | 0 |
| L2 Regularization Lambda | 1 |
| Scale Positive Weight | 1 |
| Base/Initial Score | 0.5 |
Neural Network
For the Neural Network model, AutoML uses TensorFlow with KerasOpens in a new window as a wrapper.
The input layer has its size based on the number of features. This layer is then densely connected to a single hidden layer composed of 100 neurons, which implement the ReLU Activation Function. This hidden layer is densely connected to the final output layer, which implements the Softmax Activation Function. The number of neurons in the output layer is equivalent to the number of classes present for classification.
The model hyperparameters are detailed below:
| Hyperparameter | Value |
|---|---|
| Optimizer (name) | Adam |
| Beta_1 | 0.9 |
| Beta_2 | 0.999 |
| Epsilon | 1e-07 |
| Amsgrad | False |
| Loss | Sparse Categorical Crossentropy |
Logistic Regression
For the Logistic Regression Model, AutoML uses SciKit Learn’s Logistic RegressionOpens in a new window class.
The model hyperparameters are detailed below:
| Hyperparameter | Value |
|---|---|
| Penalty | L2 |
| Dual Formulation | False |
| Tolerance | 1e-4 |
| C (Inverse Regularization Parameter) | 1 |
| Fit Intercept | True |
| Intercept Scaling | 1 |
| Class Weight | Balanced |
| Solver | liblinear |
| Max Iterations | 100 |
| Multiclass | One-vs-Rest |
| Warm Start | False |
| Number of Jobs | 1 |
Random Forest Classifier
For the Random Forest Classifier model, AutoML uses SciKit Learn’s Random Forest ClassifierOpens in a new window class.
The model hyperparameters are detailed below:
| Hyperparameter | Value |
|---|---|
| Number of Estimators | 100 |
| Criterion | Gini Impurity |
| Max Depth | None |
| Min Samples to Split | 2 |
| Min Samples to be Leaf Node | 1 |
| Min Fraction of Total Sum of Weights to be Leaf | 0 |
| Max Features | Square root of number of features |
| Max Leaf Nodes | None |
| Min Impurity Decrease for Split | 0 |
| Bootstrap | True |
| Number of Jobs | 1 |
| Warn Start | False |
| Class Weight | Balanced |
Model Selection Process
If the label column is of type float or complex, AutoML trains a regression model using XGBRegressor.
For classification models, AutoML uses the following selection process to determine the most accurate model:
If the dataset is too large, AutoML samples down the data to speed up the model selection process. The full dataset is still used for training after model selection.
The size of the dataset is calculated by multiplying the number of columns by the number of rows. If this calculated size is larger than the target size, sampling is needed. The number of rows that can be utilized is calculated by dividing the target size by the number of columns. This number of rows is randomly selected from the entire dataset to be used only for the purposes of model selection.
AutoML determines if the dataset presents a binary classification problem, or if multiple classes are present.
If it is a binary classification problem, the ROC AUC scoring metric is used.
Otherwise, the F1 scoring metric is used.
These scoring metrics are then computed for each model using Monte Carlo cross validation, with three training/testing splits of 70%/30%, to determine the best model.
