def ranking_metrics_python(test, predictions, k=DEFAULT_K):
return {
"MAP": map_at_k(test, predictions, k=k, **COL_DICT),
"nDCG@k": ndcg_at_k(test, predictions, k=k, **COL_DICT),
"Precision@k": precision_at_k(test, predictions, k=k, **COL_DICT),
"Recall@k": recall_at_k(test, predictions, k=k, **COL_DICT),
}
def test_python_precision(rating_true, rating_pred, rating_nohit):
assert (
precision_at_k(
rating_true=rating_true,
rating_pred=rating_true,
col_prediction=DEFAULT_RATING_COL,
k=10,
)
== 0.6
)
assert precision_at_k(rating_true, rating_nohit, k=10) == 0.0
assert precision_at_k(rating_true, rating_pred, k=10) == pytest.approx(0.26666, TOL)
# Check normalization
single_user = pd.DataFrame(
{
DEFAULT_USER_COL: [1, 1, 1],
DEFAULT_ITEM_COL: [1, 2, 3],
DEFAULT_RATING_COL: [5, 4, 3],
}
)
assert (
precision_at_k(
rating_true=single_user,
rating_pred=single_user,
col_rating=DEFAULT_RATING_COL,
col_prediction=DEFAULT_RATING_COL,
k=3,
)
== 1
)
same_items = pd.DataFrame(
{
DEFAULT_USER_COL: [1, 1, 1, 2, 2, 2],
DEFAULT_ITEM_COL: [1, 2, 3, 1, 2, 3],
DEFAULT_RATING_COL: [5, 4, 3, 5, 5, 3],
}
)
assert (
precision_at_k(
rating_true=same_items,
rating_pred=same_items,
col_prediction=DEFAULT_RATING_COL,
k=3,
)
== 1
)
# Check that if the sample size is smaller than k, the maximum precision can not be 1
# if we do precision@5 when there is only 3 items, we can get a maximum of 3/5.
assert (
precision_at_k(
rating_true=same_items,
rating_pred=same_items,
col_prediction=DEFAULT_RATING_COL,
k=5,
)
== 0.6
)
def test_python_errors(rating_true, rating_pred):
with pytest.raises(ValueError):
rmse(rating_true, rating_true, col_user="******")
with pytest.raises(ValueError):
mae(
rating_pred,
rating_pred,
col_rating=DEFAULT_PREDICTION_COL,
col_user="******",
)
with pytest.raises(ValueError):
rsquared(rating_true, rating_pred, col_item="not_item")
with pytest.raises(ValueError):
exp_var(
rating_pred,
rating_pred,
col_rating=DEFAULT_PREDICTION_COL,
col_item="not_item",
)
with pytest.raises(ValueError):
precision_at_k(rating_true, rating_pred, col_rating="not_rating")
with pytest.raises(ValueError):
recall_at_k(rating_true, rating_pred, col_prediction="not_prediction")
with pytest.raises(ValueError):
ndcg_at_k(rating_true, rating_true, col_user="******")
with pytest.raises(ValueError):
map_at_k(
rating_pred,
rating_pred,
col_rating=DEFAULT_PREDICTION_COL,
col_user="******",
)
def run_eval(self):
"""Run evaluation on self.data.test.
Returns:
dict: Results of all metrics in `self.metrics`.
"""
topk_scores = self.recommend_k_items(self.data.test,
top_k=self.top_k,
use_id=True)
ret = []
for metric in self.metrics:
if metric == "map":
ret.append(
map_at_k(self.data.test,
topk_scores,
relevancy_method=None,
k=self.top_k))
elif metric == "ndcg":
ret.append(
ndcg_at_k(self.data.test,
topk_scores,
relevancy_method=None,
k=self.top_k))
elif metric == "precision":
ret.append(
precision_at_k(self.data.test,
topk_scores,
relevancy_method=None,
k=self.top_k))
elif metric == "recall":
ret.append(
recall_at_k(self.data.test,
topk_scores,
relevancy_method=None,
k=self.top_k))
return ret
def test_spark_python_match(python_data, spark):
# Test on the original data with k = 10.
df_true, df_pred = python_data
dfs_true = spark.createDataFrame(df_true)
dfs_pred = spark.createDataFrame(df_pred)
eval_spark1 = SparkRankingEvaluation(dfs_true, dfs_pred, k=10)
assert recall_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark1.recall_at_k(), TOL)
assert precision_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark1.precision_at_k(),
TOL)
assert ndcg_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark1.ndcg_at_k(), TOL)
assert map_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark1.map_at_k(), TOL)
# Test on the original data with k = 3.
dfs_true = spark.createDataFrame(df_true)
dfs_pred = spark.createDataFrame(df_pred)
eval_spark2 = SparkRankingEvaluation(dfs_true, dfs_pred, k=3)
assert recall_at_k(df_true, df_pred,
k=3) == pytest.approx(eval_spark2.recall_at_k(), TOL)
assert precision_at_k(df_true, df_pred,
k=3) == pytest.approx(eval_spark2.precision_at_k(),
TOL)
assert ndcg_at_k(df_true, df_pred,
k=3) == pytest.approx(eval_spark2.ndcg_at_k(), TOL)
assert map_at_k(df_true, df_pred,
k=3) == pytest.approx(eval_spark2.map_at_k(), TOL)
# Remove the first row from the original data.
df_pred = df_pred[1:-1]
dfs_true = spark.createDataFrame(df_true)
dfs_pred = spark.createDataFrame(df_pred)
eval_spark3 = SparkRankingEvaluation(dfs_true, dfs_pred, k=10)
assert recall_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark3.recall_at_k(), TOL)
assert precision_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark3.precision_at_k(),
TOL)
assert ndcg_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark3.ndcg_at_k(), TOL)
assert map_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark3.map_at_k(), TOL)
# Test with one user
df_pred = df_pred.loc[df_pred["userID"] == 3]
df_true = df_true.loc[df_true["userID"] == 3]
dfs_true = spark.createDataFrame(df_true)
dfs_pred = spark.createDataFrame(df_pred)
eval_spark4 = SparkRankingEvaluation(dfs_true, dfs_pred, k=10)
assert recall_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark4.recall_at_k(), TOL)
assert precision_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark4.precision_at_k(),
TOL)
assert ndcg_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark4.ndcg_at_k(), TOL)
assert map_at_k(df_true, df_pred,
k=10) == pytest.approx(eval_spark4.map_at_k(), TOL)
logger.debug(f"Prediction: {col_prediction}")
logger.debug(f"Relevancy: {relevancy_method}")
logger.debug(f"K: {k}")
logger.debug(f"Threshold: {threshold}")
logger.debug(f"Rating True path: {args.rating_true}")
logger.debug(f"Shape of loaded DataFrame: {rating_true.shape}")
logger.debug(f"Rating Pred path: {args.rating_pred}")
logger.debug(f"Shape of loaded DataFrame: {rating_pred.shape}")
eval_precision = precision_at_k(
rating_true,
rating_pred,
col_user=col_user,
col_item=col_item,
col_rating=col_rating,
col_prediction=col_prediction,
relevancy_method=relevancy_method,
k=k,
threshold=threshold,
)
logger.debug(f"Score: {eval_precision}")
# Log to AzureML dashboard
run = Run.get_context()
run.parent.log("Precision at {}".format(k), eval_precision)
score_result = pd.DataFrame({"precision_at_k": [eval_precision]})
save_data_frame_to_directory(
args.score_result,