Glossary
Plain explanations of the words that come back in many notes. Each note also explains its own terms the first time it uses them.
Places and data
Stay cluster (stay). A group of GPS points recorded while a person stayed in one place, for example sitting in a cafe for 20 minutes.
Entity. One place on the map: a building, a street, or an open space (a park, a plaza).
Entity type. Building, street or open space. The model predicts the type of place, not the exact building.
Candidate. An entity close enough to a stay to be a possible answer. Each stay has a list of candidates.
Ground truth (GT). The correct answers. In this study, participants wrote down where they really were, so each stay has a known correct type.
MILES. A large GPS dataset from a smartphone app. The model is meant to be used on MILES later.
Model terms
Model. A formula that turns the scores of each candidate into a probability for building, street and open space.
MNL / logit / conditional logit. The type of model used here. It gives each candidate a value, and turns those values into probabilities that add up to 1.
Weight (coefficient, w). How much one score counts in the model. The model learns the weights from the ground truth.
Parameter. A number the model learns. "12 parameters" means the model learns 12 weights.
Generic vs alternative-specific. A generic weight is the same for all three types. An alternative-specific weight is different for building, street and open space.
ASC (alternative-specific constant). A fixed extra value for each type, used to see if one type is preferred in general.
Calibration. Finding the best weights and settings by fitting the model to the ground truth.
α (alpha). A setting that controls how fast the distance score drops as a candidate gets farther away.
σ (sigma). How spread out the GPS points of a stay are. A large σ means the GPS position was uncertain.
R_buffer. The search distance around each GPS point. Entities inside this distance become candidates. The study uses 10 m.
How results are measured
Accuracy. The share of stays where the model's top answer is the correct type. 0.716 means 71.6% correct.
Recall (per type). Of all stays that really were, for example, street, the share the model called street.
In-sample. Accuracy measured on the same data the model learned from. It is usually too optimistic.
Cross-validation (CV). The data is split into 5 parts (folds). The model learns from 4 parts and is tested on the 5th, five times. This shows how well the model works on data it has not seen.
Grouped CV. CV where all recordings of the same stop stay in the same part. This study needs it because three phones recorded each stop at the same time.
± value. How much the result changes from fold to fold (one standard deviation). In 0.716 ± 0.076, the 0.076 shows that some folds are clearly easier than others.
Log-likelihood (LL). A number that shows how well the model's probabilities fit the correct answers. Closer to 0 is better.
ρ² (rho-squared). How much better the model is than a simple baseline. Negative means worse than the baseline.
t-value. How clearly a weight differs from zero. Above about 2 (or below -2) usually counts as significant.
p-value. The chance of seeing a result this strong if there were no real effect. Below 0.05 usually counts as significant.
LR test (likelihood-ratio test). A test that checks whether extra parameters improve the fit enough to be worth keeping.
Overfitting. When a model learns the training data too closely, including noise, and then does worse on new data.
Warning signs in a model
Correlation (r). How closely two numbers move together, from -1 to 1. Near 1 means they almost always rise and fall together.
Multicollinearity. When several scores in a model are strongly correlated, so the model cannot tell their effects apart well.
VIF (variance inflation factor). A number for each score that shows multicollinearity. Above 10 is the usual warning line.
Condition number. One number for the whole model that shows how stable the fit is. Large values (thousands) mean the individual weights are unreliable.
Random Forest, XGBoost. Two machine-learning methods built from many decision trees. They are flexible but hard to interpret.
Part of Thesis summary.
Linked from MNL model development · Model evaluation metrics · Thesis summary