Files
semantica/benchmarks/quality_assurance/test_deduplication.py
T

338 lines
9.6 KiB
Python

import random
import string
import time
from typing import Any, Dict, List
from unittest.mock import patch
import numpy as np
import pytest
from semantica.deduplication.cluster_builder import ClusterBuilder
from semantica.deduplication.duplicate_detector import DuplicateDetector
from semantica.deduplication.entity_merger import EntityMerger
from semantica.deduplication.similarity_calculator import SimilarityCalculator
# Infra
class NullTracker:
"""
Discards all data to prevent memory leaks
"""
def start_tracking(self, *args, **kwargs):
return "dummy_id"
def update_tracking(self, *args, **kwargs):
pass
def stop_tracking(self, *args, **kwargs):
pass
def register_pipeline_modules(self, *args, **kwargs):
pass
def clear_pipeline_context(self, *args, **kwargs):
pass
def update_progress(self, *args, **kwargs):
pass
@property
def enabled(self):
return False
@enabled.setter
def enabled(self, value):
pass
@pytest.fixture(autouse=True)
def kill_io_overhead():
"""
Replaces ProgressTracker with NullTracker globally.
"""
with patch("semantica.utils.logging.get_logger"), patch(
"semantica.utils.progress_tracker.get_progress_tracker"
) as mock_getter:
mock_getter.return_value = NullTracker()
with patch(
"semantica.deduplication.similarity_calculator.get_progress_tracker",
return_value=NullTracker(),
), patch(
"semantica.deduplication.duplicate_detector.get_progress_tracker",
return_value=NullTracker(),
), patch(
"semantica.deduplication.cluster_builder.get_progress_tracker",
return_value=NullTracker(),
):
yield
# Sim data
def generate_entity_cluster(base_name: str, size: int) -> List[Dict[str, Any]]:
"""
Generates a cluster of similar entities based on a seed name.
Example: "Apple" -> ["Apple Inc", "Apple Corp", etc.]
"""
entities = []
suffixes = ["Inc", "Corp", "Ltd", "Gmbh", "LLC", "Group", "Systems"]
for i in range(size):
if random.random() < 0.8:
name = f"{base_name} {random.choice(suffixes)}"
else:
# Generating a typo for our calc to work on
chars = list(base_name)
if len(chars) > 2:
idx = random.randint(0, len(chars) - 2)
chars[idx], chars[idx + 1] = chars[idx + 1], chars[idx]
name = "".join(chars)
entities.append(
{
"id": f"{base_name.lower()}_{i}",
"name": name,
"type": "Organization",
"properties": {
"location": "USA" if i % 2 == 0 else "California",
"sector": "Tech",
"employee_count": 100 + i,
},
}
)
return entities
def generate_relationship_dataset(size: int) -> List[Dict[str, Any]]:
"""
Generates a dataset of graph relationships/triplets.
Includes exact matches, synonym predicates, and dirty literal strings.
"""
relationships = []
predicates = ["works_for", "employed_by", "is_employee_of", "has_employer"]
for i in range(size):
# Base relationship
rel = {
"subject": f"Person_{i % 50}",
"predicate": random.choice(predicates),
"object": f"Company_{i % 10}"
}
relationships.append(rel)
# Inject semantic duplicates (dirty literals / synonym predicates)
if random.random() < 0.4:
dirty_rel = {
"subject": f"Person_{i % 50}",
"predicate": random.choice(predicates),
"object": f" Company_{i % 10} Inc. "
}
relationships.append(dirty_rel)
return relationships
def generate_dataset(
num_clusters: int, items_per_cluster: int, worst_case_blocking: bool = False
):
"""
Generates a full dataset
Args:
worst_case_blocking: If True, all names start with 'A' to defeat
first-char blocking strategy in SimilarityCalculator.
"""
dataset = []
for i in range(num_clusters):
if worst_case_blocking:
# All starts with 'A'
base_name = f"A_Company_{i}"
else:
start_char = random.choice(string.ascii_uppercase)
base_name = f"{start_char}_company_{i}"
cluster = generate_entity_cluster(base_name, items_per_cluster)
dataset.extend(cluster)
return dataset
# ~~ Benchmarks ~~
@pytest.mark.parametrize("method", ["levenshtein", "jaro_winkler"])
def test_string_metric_speed(benchmark, method):
"""
Measures the speed of string comparison algos.
"""
calc = SimilarityCalculator()
s1 = "International Business Machines Corporation"
s2 = "International Business Machine Corp."
benchmark.pedantic(
lambda: calc.calculate_string_similarity(s1, s2, method=method),
iterations=1000,
rounds=100,
)
def test_full_similarity_calculation(benchmark):
"""
Measures weighted multi-factor calculation overhead.
(String + Property + Relationship + Weights).
"""
calc = SimilarityCalculator(
string_weight=0.5, property_weight=0.3, relationship_weight=0.2
)
e1 = {
"name": "Acme Corp",
"properties": {"loc": "NY", "id": "123"},
"relationships": [{"target": "t1"}, {"target": "t2"}],
}
e2 = {
"name": "Acme Inc",
"properties": {"loc": "NY", "id": "123"},
"relationships": [{"target": "t1"}, {"target": "t2"}],
}
benchmark.pedantic(
lambda: calc.calculate_similarity(e1, e2), iterations=1000, rounds=50
)
@pytest.mark.parametrize("dataset_size", [100, 500])
def test_duplicate_detection_scaling_opt(benchmark, dataset_size):
"""
Tests duplication on a 'Distributed' dataset (Best Case)
Now utilizing V2 Candidate Generation to ensure no regressions.
"""
data = generate_dataset(
num_clusters=dataset_size // 10, items_per_cluster=10, worst_case_blocking=False
)
detector = DuplicateDetector(
similarity_threshold=0.8,
similarity={
"candidate_strategy": "blocking_v2",
"max_candidates_per_entity": 50,
"prefilter_enabled": True,
"score_breakdown_enabled": True,
"prefilter_thresholds": {
"min_length_ratio": 0.4,
"require_shared_token": True
}
}
)
benchmark.pedantic(lambda: detector.detect_duplicates(data), iterations=1, rounds=5)
@pytest.mark.parametrize("dataset_size", [100, 500])
def test_duplicate_detection_worst_Case(benchmark, dataset_size):
"""
Tests detection on a 'Clustered' dataset (Worst Case).
Now utilizing V2 Candidate Generation to cut the pair explosion.
"""
data = generate_dataset(
num_clusters=dataset_size // 10, items_per_cluster=10, worst_case_blocking=True
)
detector = DuplicateDetector(
similarity_threshold=0.8,
similarity={
"candidate_strategy": "blocking_v2",
"max_candidates_per_entity": 50,
"prefilter_enabled": True,
"score_breakdown_enabled": True,
"prefilter_thresholds": {
"min_length_ratio": 0.4,
"require_shared_token": True
}
}
)
benchmark.pedantic(lambda: detector.detect_duplicates(data), iterations=1, rounds=5)
def test_incremental_detection_speed(benchmark):
"""
Measures performance of adding new data to existing index.
"""
existing = generate_dataset(num_clusters=50, items_per_cluster=5)
new_data = generate_dataset(num_clusters=5, items_per_cluster=2)
detector = DuplicateDetector()
benchmark.pedantic(
lambda: detector.incremental_detect(new_data, existing), iterations=5, rounds=10
)
@pytest.mark.parametrize("algo", ["graph", "hierarchical"])
def test_clustering_strategy_performance(benchmark, algo):
"""
Comapres Union-Fund (Graph) vs Hierarchical Clustering.
"""
data = generate_dataset(num_clusters=20, items_per_cluster=10)
use_hierarchical = algo == "hierarchical"
builder = ClusterBuilder(use_hierarchical=use_hierarchical)
benchmark.pedantic(lambda: builder.build_clusters(data), iterations=1, rounds=5)
def test_merge_entity_benchmark(benchmark):
"""
Measures the cost of fusing entities / res conflicts.
"""
group = generate_entity_cluster("MegaCorp", 50)
merger = EntityMerger()
benchmark.pedantic(
lambda: merger.merge_entity_group(group, strategy="keep_most_complete"),
iterations=10,
rounds=10,
)
@pytest.mark.parametrize("mode", ["legacy", "semantic_v2"])
def test_relationship_dedup_speed(benchmark, mode):
"""
Measures the speed of relationship/triplet deduplication.
Compares the O(N^2) legacy fallback vs the fast canonical hash path.
"""
# Yields ~280 relationships (approx 39,000 comparisons in O(N^2))
relationships = generate_relationship_dataset(200)
detector = DuplicateDetector()
options = {
"threshold": 0.85,
"relationship_dedup_mode": mode,
"predicate_synonym_map": {
"works_for": "employed_by",
"is_employee_of": "employed_by",
"has_employer": "employed_by"
},
"literal_normalization_enabled": True
}
benchmark.pedantic(
lambda: detector.detect_relationship_duplicates(relationships, **options),
iterations=5,
rounds=10,
)