(Data & ML) 2022 — 2023 · Academic project
Six prediction models, tested side by side, to estimate how long a taxi ride takes.
A research project on New York taxi data: cleaning it, finding patterns in pickup times, and testing six prediction models side by side to see which one guesses best.
(01) In numbers
Each scored on the same features and the same split
The baseline every result has to beat to mean anything
Components in place of heavily correlated raw columns
(02) Under the hood
01 · RECORDS
Pickup and dropoff points, timestamps, passenger counts. Duration is the thing that has to be predicted from them.
02 · CLEAN
Trips that take a day, trips that take no time at all, coordinates that are not in New York. Nothing useful is learned from any of them.
03 · FEATURES
Distance, bearing, hour of day, day of week — the things a duration actually depends on, made explicit rather than left implicit in a timestamp.
04 · PCA
The engineered columns overlap heavily. PCA turns them into components that do not, and hands the models a smaller space to work in.
05 · MODELS
Six of them, on the same features and the same split, so that the comparison is about the model rather than about the setup.
06 · BASELINE
Predicting the mean, every single time. A regressor that cannot beat it has not learned anything worth having.
07 · SCORE
Each regressor read against the baseline instead of against an absolute number — the only reading that says whether the features carry signal.
(03) The story
01 — The Problem
Trip duration can be predicted, but an error figure on its own is not a result. With nothing to compare against, a model that has learned the average and a model that has learned the city look much the same on paper.
02 — What I Built
The records are cleaned, turned into features that describe a journey rather than a row, compressed with PCA, and fed to six regressors — all of them measured against a null model that predicts the mean.
03 — What Changed
Every model is reported relative to the baseline, so the number that matters is how much of the duration the features actually explain — not how small an error can be made to look.
(04) Architecture
01
The raw New York taxi trip dataset, as published.
02
Impossible durations and out-of-range coordinates removed first.
03
Distance, bearing, hour and day derived from the raw columns.
04
A correlated feature space reduced to components.
05
Six regressors, fitted on the same split.
06
Every score read against the null model.
(05) Decisions
The null model is written before the regressors, not after them. It sets the bar the rest of the work has to clear, and it stops a respectable-looking error from being mistaken for a result.
Features derived from impossible rows are impossible features. Cleaning comes first so that everything downstream is describing a journey that could have happened.
Six models, the same features, the same train and test split. Any difference in score is then a difference between the models and nothing else.
(06) Built with
A study, not a service — but the same instinct: measure against something.
For the technically curiousSix regressors benchmarked against a null model on PCA features. A prediction is only interesting once you know what doing nothing looks like. This study cleans New York taxi trip records into features that describe a journey, compresses them with PCA, and runs six regressors against a null model so that every score has something to be measured against.
PythonpandasNumPyscikit-learnPCARegressionFeature EngineeringBenchmarking
