test_algorithms.py

"""
Module for testing prediction algorithms.
"""


import os

import pytest

from surprise import (
    accuracy,
    BaselineOnly,
    CoClustering,
    Dataset,
    KNNBaseline,
    KNNBasic,
    KNNWithMeans,
    KNNWithZScore,
    NMF,
    NormalPredictor,
    Reader,
    SlopeOne,
    SVD,
    SVDpp,
)
from surprise.model_selection import train_test_split


def test_unknown_user_or_item(toy_data):
    """Ensure that all algorithms act gracefully when asked to predict a rating
    of an unknown user, an unknown item, and when both are unknown.
    """

    trainset = toy_data.build_full_trainset()

    klasses = (
        NormalPredictor,
        BaselineOnly,
        KNNBasic,
        KNNWithMeans,
        KNNBaseline,
        SVD,
        SVDpp,
        NMF,
        SlopeOne,
        CoClustering,
        KNNWithZScore,
    )
    for klass in klasses:
        algo = klass()
        algo.fit(trainset)
        algo.predict("user0", "unknown_item", None)
        algo.predict("unknown_user", "item0", None)
        algo.predict("unknown_user", "unknown_item", None)

    # unrelated, but test the fit().test() one-liner:
    trainset, testset = train_test_split(toy_data, test_size=2)
    for klass in klasses:
        algo = klass()
        algo.fit(trainset).test(testset)


def test_knns(u1_ml100k, pkf):
    """Ensure the k and min_k parameters are effective for knn algorithms."""

    # Actually, as KNNWithMeans and KNNBaseline have back up solutions for when
    # there are not enough neighbors, we can't really test them...
    klasses = (KNNBasic,)  # KNNWithMeans, KNNBaseline)

    k, min_k = 20, 5
    for klass in klasses:
        algo = klass(k=k, min_k=min_k)
        for trainset, testset in pkf.split(u1_ml100k):
            algo.fit(trainset)
            predictions = algo.test(testset)
            for pred in predictions:
                if not pred.details["was_impossible"]:
                    assert min_k <= pred.details["actual_k"] <= k


def test_nearest_neighbors():
    """Ensure the nearest neighbors are different when using user-user
    similarity vs item-item."""

    reader = Reader(
        line_format="user item rating", sep=" ", skip_lines=3, rating_scale=(1, 5)
    )

    data_file = os.path.dirname(os.path.realpath(__file__)) + "/custom_train"
    data = Dataset.load_from_file(data_file, reader)
    trainset = data.build_full_trainset()

    algo_ub = KNNBasic(sim_options={"user_based": True})
    algo_ub.fit(trainset)
    algo_ib = KNNBasic(sim_options={"user_based": False})
    algo_ib.fit(trainset)
    assert algo_ub.get_neighbors(0, k=10) != algo_ib.get_neighbors(0, k=10)


@pytest.mark.parametrize(
    "algo, expected_rmse",
    [
        (BaselineOnly(), 1.0268524031297395),
        (KNNBasic(sim_options={"name": "cosine"}), 1.1495072745332722),
        (KNNBasic(sim_options={"name": "msd"}), 1.1337265249554591),
        (KNNBasic(sim_options={"name": "pearson"}), 1.1218719178944856),
        (KNNBasic(sim_options={"name": "pearson_baseline"}), 1.1242139598121241),
        (KNNWithMeans(), 1.1043129441881696),
        (KNNBaseline(), 1.0700718041752253),
        (KNNWithZScore(), 1.11179436167853),
        (SVD(random_state=0), 1.0077323320656948),
        (SVDpp(cache_ratings=False, random_state=0), 1.00284553561452),
        (SVDpp(cache_ratings=True, random_state=0), 1.00284553561452),
        (NMF(random_state=0), 1.0865370266372372),
        (SlopeOne(), 1.1559939123891685),
        (CoClustering(random_state=0), 1.0841941385276614),
    ],
)
def test_sanity_checks(u1_ml100k, pkf, algo, expected_rmse):
    """Basic sanity checks for all algorithms: check that RMSE stays the same."""

    trainset, testset = next(pkf.split(u1_ml100k))
    algo.fit(trainset)
    predictions = algo.test(testset)
    assert not all(pred.details["was_impossible"] for pred in predictions)
    assert accuracy.rmse(predictions, verbose=False) == expected_rmse