Loading subject…
Activation function
The function a node applies to its weighted sum to produce its output. It introduces non-linearity, which is what allows a network of layers to model more than a single straight boundary.
Apriori algorithm
A method for association rule learning that first finds every itemset meeting a minimum support, then builds rules from those itemsets and keeps the ones meeting a minimum confidence. It prunes the search using the property that any set containing an infrequent itemset must itself be infrequent.
Artificial Neural Network (ANN)
A computational model built from interconnected nodes, or neurons, loosely inspired by the brain.
Association rule
An if-then statement reporting that two sets of items tend to appear in the same transaction, written $\{\text{antecedent}\} \Rightarrow \{\text{consequent}\}$. It records co-occurrence in the data and makes no claim about cause.
Backpropagation
The training step that works backwards from the loss at the output layer, calculating how much each weight and bias contributed to the error so gradient descent can adjust them.
Bias (neural network)
A learned constant added to a node's weighted sum, which shifts the threshold at which the node fires so its decision boundary need not pass through the origin.
Centroid
The mean position of all the data points assigned to a cluster. It is a calculated position rather than an observation, so usually no data point sits exactly on it.
Clustering
Clustering is a technique used to group a set of objects so that objects in the same group (or cluster) are more similar to each other than to those in other groups.
Coefficient of determination
A measure of model fit, written $r^2$, giving the share of the variation in the dependent variable that the model accounts for. It runs from 0, no better than always predicting the mean, to 1, a line passing through every observed point.
Collaborative filtering
A recommendation method that treats each user as a point defined by their ratings, finds the users nearest to them, and suggests items those neighbours rated highly.
Example
A streaming service suggests a film you have not watched because the three users whose ratings most resemble yours all rated it highly.
Confidence
For an association rule, the share of the transactions containing the antecedent that also contain the consequent. It equals the support of the whole itemset divided by the support of the antecedent alone.
Confusion matrix
A table of the four outcome counts of a two-class prediction: true positives, false positives, false negatives and true negatives.
Example
On 1000 admissions a sepsis model records 80 true positives, 10 false positives, 20 false negatives and 890 true negatives.
Convolutional Neural Network (CNN)
A neural network designed for data with spatial structure, above all images.
Cross-validation
A way of estimating performance that splits the data several times, uses a different slice as the validation set each round, and averages the results.
Example
Five-fold cross-validation trains five times on 800 records and validates on the remaining 200, then averages the five scores.
Crossover
Crossover, also called recombination, is the operator that combines parts of two parent solutions to produce offspring, so features that already score well can be mixed into one candidate.
Cumulative reward
The sum of every reward a reinforcement learning agent collects across a complete episode, which is the quantity the agent tries to maximise.
DBSCAN
A density-based clustering algorithm that groups points packed closely together, using a neighbourhood radius epsilon and a minimum neighbour count minPts, and labels points in sparse regions as outliers. It finds clusters of any shape and does not need the number of clusters in advance.
Decision tree
A supervised learning model that classifies an example by testing one feature at a time and following the branch that matches, until it reaches a leaf carrying a label.
Example
A triage tree that tests whether temperature exceeds 38 degrees, then whether a cough is present, and ends at a leaf labelled urgent or routine.
Discount factor
A value between 0 and 1, written $\gamma$, that sets how much a later reward is worth to a reinforcement learning agent now. A reward $k$ steps ahead counts as $\gamma^k$ times its face value.
Epsilon-greedy
A strategy that takes a random action with probability $\epsilon$ and the highest-valued known action the rest of the time, giving the agent a set rate of exploration.
Exploration and exploitation trade-off
The choice a reinforcement learning agent faces at every step between taking the action it currently believes is best and trying an action whose value it is still uncertain about.
F1 score
The harmonic mean of precision and recall, used as one number when a false positive and a false negative both carry real cost.
Example
A precision of 0.89 and a recall of 0.80 give an F1 score of 0.84.
Feature map
The grid of values produced when one filter is applied at every position of a layer's input. Each value records how strongly that filter's pattern matched at that position.
Fitness function
A fitness function is the rule that scores how good a candidate solution is, and it is the only part of a genetic algorithm that carries what the problem is trying to achieve.