Every term StudyDSA teaches, starting with machine learning words like overfitting, bias, and variance. Each one links to the lesson section that explains it.
the lowest error rate any classifier can reach on a problem, because some examples genuinely look like the other class
k-Nearest Neighborshow far off the predictions still are after averaging over every batch: an offset built into the model's shape
k-Nearest Neighborspredicting a category, like spam or not spam, cat or dog
k-Nearest Neighborsthe number of axes an example lives on, one per feature
k-Nearest Neighborsone thing you have data on: one email, one house, one motor
k-Nearest Neighborsthe numbers used to describe an example, like an email's link count and its exclamation mark count
k-Nearest Neighborsthe answer attached to an example, like spam or not spam, or a house's sale price
k-Nearest Neighborsa method that does all of its computing at prediction time instead of learning something up front
k-Nearest Neighborsrescaling a feature so its smallest value lands at 0 and its largest lands at 1
k-Nearest Neighborsfitting the training examples so tightly that the model also fits their mistakes and random quirks, which don't repeat on new data
The classic sign is a training error near 0 while the test error climbs. The model stopped learning the pattern and started memorizing the exact pile it was handed, quirks included. Both curves below run through the same points, but only one of them will hold up on tomorrow's data:
predicting a number, like a price or a lifetime in hours
k-Nearest Neighborsrescaling a feature so its average lands at 0 and the values are measured in standard deviations away from it
k-Nearest Neighborslearning from examples that come with their answers, supplied ahead of time by a person
k-Nearest Neighborsthe error rate measured on examples the model has never seen before
k-Nearest Neighborslabeled examples kept away from the model, so its mistakes on them predict its mistakes on new data
k-Nearest Neighborsthe pile of examples that already carry labels, the only thing the algorithm gets to learn from
k-Nearest Neighborsthe error rate measured on the examples the model has already seen
k-Nearest Neighborsa model too blunt to follow the real pattern, so it scores badly on the training examples and on new ones too
k-Nearest Neighborshow much a model's prediction for one point changes from one training batch to the next
k-Nearest Neighbors