def test_python_ndcg_at_k(rating_true, rating_pred, rating_nohit):
assert (ndcg_at_k(
rating_true=rating_true,
rating_pred=rating_true,
col_prediction=DEFAULT_RATING_COL,
k=10,
) == pytest.approx(1.0, TOL))
assert ndcg_at_k(rating_true, rating_nohit, k=10) == 0.0
assert ndcg_at_k(rating_true, rating_pred,
k=10) == pytest.approx(0.38172, TOL)
def on_epoch_end(self, batch, logs={}):
"""At the end of each epoch calculate NDCG@k of the validation set.
If the model performance is improved, the model weights are saved.
Update the list of validation NDCG@k by adding obtained value
"""
# recommend top k items based on training part of validation set
top_k = self.recommend_k_items(x=self.val_tr, k=self.k, remove_seen=True)
# convert recommendations from sparse matrix to dataframe
top_k_df = self.mapper.map_back_sparse(top_k, kind="prediction")
test_df = self.mapper.map_back_sparse(self.val_te, kind="ratings")
# calculate NDCG@k
NDCG = ndcg_at_k(test_df, top_k_df, col_prediction="prediction", k=self.k)
# check if there is an improvement in NDCG, if so, update the weights of the saved model
if NDCG > self.best_ndcg:
self.best_ndcg = NDCG
# save the weights of the optimal model
if self.save_path is not None:
self.model.save(self.save_path)
self._data.append(NDCG)
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_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 test_recommend_k_items(rating_true):
train_set = cornac.data.Dataset.from_uir(
rating_true.itertuples(index=False), seed=42)
bpr = cornac.models.BPR(k=100, max_iter=10000, seed=42).fit(train_set)
preds = predict_ranking(bpr, rating_true, remove_seen=False)
n_users = len(rating_true["userID"].unique())
n_items = len(rating_true["itemID"].unique())
assert preds.shape[0] == n_users * n_items
assert set(preds.columns) == {"userID", "itemID", "prediction"}
assert preds["userID"].dtypes == rating_true["userID"].dtypes
assert preds["itemID"].dtypes == rating_true["itemID"].dtypes
# perfect ranking achieved
assert 1e-10 > 1 - ndcg_at_k(rating_true, preds)
assert 1e-10 > 1 - recall_at_k(rating_true, preds)
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_ndcg = ndcg_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_ndcg}")
# Log to AzureML dashboard
run = Run.get_context()
run.parent.log("nDCG at {}".format(k), eval_ndcg)
score_result = pd.DataFrame({"ndcg_at_k": [eval_ndcg]})
save_data_frame_to_directory(
args.score_result,
