Description
Trains a layered network of weighted nodes on example records to predict a numeric or categorical outcome, tuning layers, cycles and learning rate against cross-validated accuracy.
A multilayer perceptron has an input layer with one node per normalised descriptor variable, one or more hidden layers, and an output layer with one node per response. Every connection carries a weight, initially random, and each node combines its weighted inputs through an activation function. During learning, training examples are presented one at a time and backpropagation adjusts the weights to reduce the error between predicted and actual outputs; one presentation is a cycle (epoch). The analyst sets the number of hidden layers and nodes, the number of cycles and the learning rate. Myatt treats model building as a designed experiment: try combinations of inputs, hidden layers and cycle counts, score each with cross-validated r squared or error, watch for the point where accuracy stops rising or falls (overtraining), and pick the simplest model that performs well. Networks capture non-linear relationships and tolerate noise, but they are a black box, so for UX work they suit prediction tasks where accuracy matters more than explanation, and should be paired with an explanation method or a simpler comparison model.
Attributes
- Discipline
- Computational & Machine Learning
- Loss type
- Siloed
- Stage
- Post-launch, Strategy
- Data
- Quantitative
- Effort
- Weeks
Steps
- Define the outcome to predict (for example 30-day retention or task time) and the accuracy the decision needs.
- Prepare and normalise descriptor variables; dummy-code categories.
- Hold out a test set or set up k-fold cross-validation.
- Design a tuning grid over input sets, number of hidden layers and nodes, number of cycles and learning rate.
- Train each configuration and record cross-validated accuracy, plotting accuracy against cycles to spot overtraining.
- Choose the simplest configuration with near-best accuracy and compare it with a transparent model such as a regression or decision tree.
- Inspect mispredicted cases and translate outputs back to original units before reporting.
Sources
- Myatt, G., Making Sense of Data: A Practical Guide to Exploratory Data Analysis and Data Mining (2007)
Also known as
- Backpropagation neural network
- Feed-forward neural network
- MLP regression or classification
- Artificial neural network model
Categories
- Analytics & Behaviour
- Computational & Machine Learning