Phases of a Machine Learning
A machine learning workflow's specific activities vary based on the learning context (supervised, unsupervised, or reinforcement), a generalized process can be established.
This workflow is generally divided into four primary phases:
Data Preparation
Learning and Modeling
Performance Evaluation
Performance Improvement
The process begins with the data.
For supervised learning, this involves a labeled training dataset and unlabeled test data.
Conversely, unsupervised learning relies on unlabeled data, as its primary goal is to discover inherent patterns within the dataset.
1. Data Preparation (Exploration and Pre-processing)
The success of a machine learning model is highly dependent on the quality of the data. This initial phase focuses on exploring and preparing the data to ensure its quality and suitability for modeling.
Typical preparation activities include:
Data Type Identification: Understand the types of data (e.g., categorical, numerical, text) in the dataset.
Data Exploration: Analyze the data to understand its statistical properties, distributions, and quality.
Relationship Exploration: Investigate relationships between data elements, such as inter-feature correlations.
Issue Identification: Detect potential problems such as missing values, outliers, or inconsistencies.
Data Remediation: Perform corrective actions, such as imputing missing values or removing outliers.
Data Pre-processing: Apply necessary transformations, such as normalization, scaling, or one-hot encoding.
Data exploration helps uncover problems like missing values or outliers (data elements with surprisingly different values from others).
2. Learning and Modeling
Once the data is prepared, the learning tasks commence. This phase involves selecting, training, and applying models.
Typical activities in this phase include:
Data Splitting: Partitioning the prepared data into training and testing (or holdout) sets. (Applicable to supervised learning).
Model Selection: Evaluating and choosing appropriate learning algorithms for the task.
Training and Application:
For supervised learning, the selected model is trained using the labeled training data. The process of fitting a model to a dataset is known as model training.
For unsupervised learning, the chosen model is applied directly to the input data to identify patterns.
Prediction/Application: The trained model is applied to new, unknown data (e.g., the test set) to generate predictions or classifications.
3. Performance Evaluation
After the model is applied, its effectiveness and accuracy must be rigorously assessed. This step determines whether the model's outputs are valid and reliable.
The primary activities involve:
Metric-Based Assessment: Evaluating the model's performance using appropriate metrics.
For classification: This may include using a confusion matrix, accuracy, precision, recall, F1-score, or visualizing trade-offs with ROC curves.
For regression: Common metrics include Mean Absolute Error (MAE) or Root Mean Square Error (RMSE).
Effectiveness Analysis: Determining if the model training (supervised) or grouping (unsupervised) was successful and meets the project's objectives.
4. Performance Improvement
Based on the results of the evaluation, the model's performance may require enhancement. This is an iterative process of refinement.
Common improvement strategies include:
Hyperparameter Tuning: Adjusting the model's internal settings (hyperparameters) to optimize performance.
Feature Engineering: Modifying, selecting, or creating new features from the existing data to improve the model's predictive power.
Model Re-selection: If performance is unsatisfactory, it may be necessary to revisit Phase 2 and select a different model or algorithm.