Skip to content

Prediction and evaluation

PredictionPipeline ties the mapping, the features and a scikit-learn estimator together.

Fitting

from sklearn.ensemble import GradientBoostingRegressor
from predspot import PredictionPipeline, PandasFeatureUnion
from predspot.crime_mapping import KDE, create_gridpoints
from predspot.feature_engineering import AR, Seasonality, Trend

pipeline = PredictionPipeline(
    mapping=KDE(tfreq="M", grid=create_gridpoints(study_area, resolution=1)),
    fextraction=PandasFeatureUnion([
        ("ar", AR(lags=3)),
        ("seasonal", Seasonality(lags=12)),
        ("trend", Trend(lags=12)),
    ]),
    estimator=GradientBoostingRegressor(random_state=0),
    random_state=0,
)
pipeline.fit(dataset)

fit computes the spatio-temporal series (pipeline.stseries), the feature matrix (pipeline.features) and trains the estimator on every (t, places) row whose target is known. Rows are shuffled (random_state) since the estimator sees them as independent samples.

Evaluating

pipeline.evaluate("r2", cv=5)            # or "mse", or ["r2", "mse"] for a DataFrame

Periods are ordered and split with scikit-learn's TimeSeriesSplit, so each fold trains on earlier periods and tests on later ones — no leakage from the future. One score per fold is returned; the estimator is refitted on all the data afterwards.

Note

Scores measure how well the density surface of the next period is predicted at every place. For patrol planning you may care more about the ranking of places (e.g. hit rate of the top-k cells); compute it from pipeline.predict() against the series of the following period.

Forecasting

forecast = pipeline.predict()        # next period
forecast.head()

                   crime_density
t          places
2021-01-31 0               27.30
           1               46.43

Each predict() call appends its forecast to pipeline.stseries, recomputes the features and advances pipeline.next_time, so repeated calls produce a multi-step, recursive forecast. Join the result with pipeline.grid to map it:

hot = pipeline.grid.join(forecast.droplevel("t"))
top = hot.nlargest(20, "crime_density")     # top-20 places for patrol planning
hot.plot(column="crime_density", cmap="magma", markersize=12, legend=True)
Forecast

Feature importances

When the estimator (or the last step of a Pipeline) exposes feature_importances_, pipeline.feature_importances returns them by feature name, accounting for a FeatureSelection step if present:

pipeline.feature_importances.head()

Modelling one crime type at a time

Different crime types have different dynamics; the usual approach is one pipeline per tag:

pipelines = {}
for tag, events in crimes.groupby("tag"):
    pipelines[tag] = build_default_pipeline(study_area).fit(Dataset(events, study_area))

One-call helpers

predspot.pipeline offers build_default_pipeline (KDE on points, seasonal/trend/diff features, quantile scaling, RFE selection and a random forest) and run_prediction_pipeline, which also filters by tag and time of day and fits in one go.