In this post, we explore some broad guidelines for selecting machine learning models
The overall steps for Machine Learning/Deep Learning are:
- Collect data
- Check for anomalies, missing data and clean the data
- Perform statistical analysis and initial visualization
- Build models
- Check the accuracy
- Present the results
Machine learning tasks can be classified into
- Supervised learning
- Unsupervised learning
- Semi-supervised learning
- Reinforcement learning
PS – in this document – we do not focus on the last two
Below are some approaches on choosing a model for Machine Learning/Deep Learning
OVERALL APPROACHES
- Dealing with unbalanced data: Use resampling strategies
- Create new features : Principal component analysis (PCA) to reduce dimensionality, Autoencoders to create a latent space and possibly Clustering to create new features
- To prevent overfitting, outliers and noise in linear regression – use regularization techniques like lasso and ridge.
- Overcoming the Black-box AI problem – consider strategies for building interpretable models
- Algorithms not sensitive to outliers : Some discussion on choice of Random Forest to overcome outliers
MACHINE LEARNING MODELS
- First approach to predicting continuous values: Linear Regression is generally a good first approach for predicting continuous values (ex: prices)
- Binary classification: Logistic regression is a good starting point for Binary classification. Support Vector Machines SVM is also a good choice of two class classification
- Multi-class classification: Random forest is a choice for multi-class classification. See SVM vs Random Forest usage
- Is there a simplest or easiest model category to start off with? Decision trees are often seen as simple to understand and use. Decision trees are implemented through models such as Random forest or Gradient boosting.
- Which models are used in Kaggle? For supervised learning: Random forest and XGboost See note on Gradient boosted trees
DEEP LEARNING MODELS
- Complex features which cannot be easily specified but you have large number of labelled examples: Multi-layer perceptrons
- Vision based Machine Learning: Image classification, Object Detection, Image segmentation – Convolutional Neural Networks
- Sequence modelling tasks: RNNs (typically LSTM) for sequence modelling tasks ex text classification or language translation
Comments welcome
Image source: BMJ – what makes machine learning in healthcare so powerful
