Skip to content

Commit

Permalink
LSH Ensemble Optimal Partitioning (#79)
Browse files Browse the repository at this point in the history
Use optimal partitioning instead of equi-depth so the index can be adaptive to any set size distribution.
  • Loading branch information
ekzhu committed Dec 27, 2018
1 parent 82d9639 commit 54d053c
Show file tree
Hide file tree
Showing 47 changed files with 975 additions and 577 deletions.
230 changes: 160 additions & 70 deletions benchmark/lshensemble_benchmark.py
Original file line number Diff line number Diff line change
@@ -1,102 +1,192 @@
"""
Benchmark dataset from:
https://github.com/ekzhu/set-similarity-search-benchmark.
Use "Canada US and UK Open Data":
Indexed sets: canada_us_uk_opendata.inp.gz
Query sets (10 stratified samples from 10 percentile intervals):
Size from 10 - 1k: canada_us_uk_opendata_queries_1k.inp.gz
Size from 10 - 10k: canada_us_uk_opendata_queries_10k.inp.gz
Size from 10 - 100k: canada_us_uk_opendata_queries_100k.inp.gz
"""
import time, argparse, sys, json
from hashlib import sha1
import numpy as np
import nltk
import scipy.stats
import random
import collections
import gzip
import random
import os
import pickle
import pandas as pd
from SetSimilaritySearch import SearchIndex

from datasketch import MinHashLSHEnsemble, MinHash
from lshforest_benchmark import bootstrap_data

def benchmark_lshensemble(threshold, num_perm, num_part, l, index_data, query_data):

def bootstrap_sets(sets_file, sample_ratio, num_perms, skip=1):
print("Creating sets...")
sets = collections.deque([])
random.seed(41)
with gzip.open(sets_file, "rt") as f:
for i, line in enumerate(f):
if i < skip:
# Skip lines
continue
if random.random() > sample_ratio:
continue
s = np.array([int(d) for d in \
line.strip().split("\t")[1].split(",")])
sets.append(s)
sys.stdout.write("\rRead {} sets".format(len(sets)))
sys.stdout.write("\n")
sets = list(sets)
keys = list(range(len(sets)))
print("Creating MinHash...")
minhashes = dict()
for num_perm in num_perms:
print("Using num_parm = {}".format(num_perm))
ms = []
for s in sets:
m = MinHash(num_perm)
for word in s:
m.update(str(word).encode("utf8"))
ms.append(m)
sys.stdout.write("\rMinhashed {} sets".format(len(ms)))
sys.stdout.write("\n")
minhashes[num_perm] = ms
return (minhashes, sets, keys)


def benchmark_lshensemble(threshold, num_perm, num_part, m, index_data,
query_data):
print("Building LSH Ensemble index")
lsh = MinHashLSHEnsemble(threshold=threshold, num_perm=num_perm, num_part=num_part, l=l)
lsh.index((key, minhash, len(set))
for key, minhash, set in \
zip(index_data.keys, index_data.minhashes[num_perm], index_data.sets))
(minhashes, indexed_sets, keys) = index_data
lsh = MinHashLSHEnsemble(threshold=threshold, num_perm=num_perm,
num_part=num_part, m=m)
lsh.index((key, minhash, len(set))
for key, minhash, set in \
zip(keys, minhashes[num_perm], indexed_sets))
print("Querying")
(minhashes, sets, keys) = query_data
times = []
results = []
for qs, minhash in zip(query_data.sets, query_data.minhashes[num_perm]):
start = time.clock()
for qs, minhash in zip(sets, minhashes[num_perm]):
start = time.perf_counter()
result = list(lsh.query(minhash, len(qs)))
duration = time.clock() - start
duration = time.perf_counter() - start
times.append(duration)
results.append(sorted([[key, _compute_containment(qs, index_data.sets[key])]
for key in result],
key=lambda x : x[1], reverse=True))
results.append(result)
# results.append(sorted([[key, _compute_containment(qs, indexed_sets[key])]
# for key in result],
# key=lambda x : x[1], reverse=True))
sys.stdout.write("\rQueried {} sets".format(len(results)))
sys.stdout.write("\n")
return times, results


def benchmark_ground_truth(threshold, index_data, query_data):
def benchmark_ground_truth(threshold, index, query_data):
(_, query_sets, _) = query_data
times = []
results = []
for q in query_data.sets:
start = time.clock()
result = [key for key, a in zip(index_data.keys, index_data.sets)
if _compute_containment(q, a) >= threshold]
duration = time.clock() - start
for q in query_sets:
start = time.perf_counter()
result = [key for key, _ in index.query(q)]
duration = time.perf_counter() - start
times.append(duration)
results.append(sorted([[key, _compute_containment(q, index_data.sets[key])]
for key in result],
key=lambda x : x[1], reverse=True))
results.append(result)
sys.stdout.write("\rQueried {} sets".format(len(results)))
sys.stdout.write("\n")
return times, results


def _compute_containment(x, y):
if len(x) == 0 or len(y) == 0:
return 0.0
intersection = 0
for w in x:
if w in y:
intersection += 1
intersection = len(np.intersect1d(x, y, assume_unique=True))
return float(intersection) / float(len(x))


if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("--output", type=str, default="lshensemble_benchmark.json")
parser = argparse.ArgumentParser(
description="Run LSH Ensemble benchmark using data sets obtained "
"from https://github.com/ekzhu/set-similarity-search-benchmark.")
parser.add_argument("--indexed-sets", type=str, required=True,
help="Input indexed set file (gzipped), each line is a set: "
"<set_size> <1>,<2>,<3>..., where each <?> is an element.")
parser.add_argument("--query-sets", type=str, required=True,
help="Input query set file (gzipped), each line is a set: "
"<set_size> <1>,<2>,<3>..., where each <?> is an element.")
parser.add_argument("--query-results", type=str,
default="lshensemble_benchmark_query_results.csv")
parser.add_argument("--ground-truth-results", type=str,
default="lshensemble_benchmark_ground_truth_results.csv")
parser.add_argument("--skip-ground-truth", action="store_true")
args = parser.parse_args(sys.argv[1:])

threshold = 0.5
num_perms = [32, 64, 96, 128, 160, 192, 224, 256]
num_part = 16
l = 8
output = {"threshold" : threshold,
"num_perms" : num_perms,
"num_part" : 16,
"l" : l,
"lsh_times" : [], "lsh_results" : [],
"ground_truth_times" : None, "ground_truth_results" : None}

population_size = 500

class zipfian:
def __init__(self):
self.rv = scipy.stats.zipf(1.25)
def rvs(self):
x = int(self.rv.rvs())
if x > population_size:
return population_size
return x

index_data, query_data = bootstrap_data(num_perms, 100, population_size, zipfian())
thresholds = [0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0]
num_parts = [8, 16, 32]
#num_perms = [32, 64, 96, 128, 160, 192, 224, 256]
num_perms = [256,]
m = 8

index_data, query_data = None, None
index_data_cache = "{}.pickle".format(args.indexed_sets)
query_data_cache = "{}.pickle".format(args.query_sets)
if os.path.exists(index_data_cache):
print("Using cached indexed sets {}".format(index_data_cache))
with open(index_data_cache, "rb") as d:
index_data = pickle.load(d)
else:
print("Using indexed sets {}".format(args.indexed_sets))
index_data = bootstrap_sets(args.indexed_sets, 0.1, num_perms)
with open(index_data_cache, "wb") as d:
pickle.dump(index_data, d)
if os.path.exists(query_data_cache):
print("Using cached query sets {}".format(query_data_cache))
with open(query_data_cache, "rb") as d:
query_data = pickle.load(d)
else:
print("Using query sets {}".format(args.query_sets))
query_data = bootstrap_sets(args.query_sets, 1.0, num_perms, skip=0)
with open(query_data_cache, "wb") as d:
pickle.dump(query_data, d)

if not args.skip_ground_truth:
rows = []
# Build search index separately, only works for containment.
print("Building search index...")
index = SearchIndex(index_data[1], similarity_func_name="containment",
similarity_threshold=0.1)
for threshold in thresholds:
index.similarity_threshold = threshold
print("Running ground truth benchmark threshold = {}".format(threshold))
ground_truth_times, ground_truth_results = \
benchmark_ground_truth(threshold, index, query_data)
for t, r, query_set, query_key in zip(ground_truth_times,
ground_truth_results, query_data[1], query_data[2]):
rows.append((query_key, len(query_set), threshold, t,
",".join(str(k) for k in r)))
df_groundtruth = pd.DataFrame.from_records(rows,
columns=["query_key", "query_size", "threshold",
"query_time", "results"])
df_groundtruth.to_csv(args.ground_truth_results)

rows = []
for threshold in thresholds:
for num_part in num_parts:
for num_perm in num_perms:
print("Running LSH Ensemble benchmark "
"threshold = {}, num_part = {}, num_perm = {}".format(
threshold, num_part, num_perm))
lsh_times, lsh_results = benchmark_lshensemble(
threshold, num_perm, num_part, m, index_data, query_data)
for t, r, query_set, query_key in zip(lsh_times, lsh_results,
query_data[1], query_data[2]):
rows.append((query_key, len(query_set), threshold,
num_part, num_perm, t, ",".join(str(k) for k in r)))
df = pd.DataFrame.from_records(rows,
columns=["query_key", "query_size", "threshold", "num_part",
"num_perm", "query_time", "results"])
df.to_csv(args.query_results)

for num_perm in num_perms:
print("Use num_perm = %d" % num_perm)
result = {}
print("Running LSH Ensemble benchmark")
lsh_times, lsh_results = benchmark_lshensemble(threshold, num_perm, num_part, l, index_data, query_data)
output["lsh_times"].append(lsh_times)
output["lsh_results"].append(lsh_results)

print("Running ground truth benchmark")
output["ground_truth_times"], output["ground_truth_results"] =\
benchmark_ground_truth(threshold, index_data, query_data)

average_cardinality = np.mean([len(s) for s in
index_data.sets + query_data.sets])
print("Average cardinality is", average_cardinality)

with open(args.output, 'w') as f:
json.dump(output, f)
100 changes: 72 additions & 28 deletions benchmark/lshensemble_benchmark_plot.py
Original file line number Diff line number Diff line change
@@ -1,40 +1,84 @@
import json, sys, argparse
import numpy as np
import pandas as pd
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from lsh_benchmark_plot import get_precision_recall, fscore, average_fscore

if __name__ == "__main__":
parser = argparse.ArgumentParser()
parser.add_argument("benchmark_output")
parser.add_argument("query_results")
parser.add_argument("ground_truth_results")
args = parser.parse_args(sys.argv[1:])
df = pd.read_csv(args.query_results,
converters={"results": lambda x: x.split(",")})
df_groundtruth = pd.read_csv(args.ground_truth_results,
converters={"results": lambda x: x.split(",")})
df_groundtruth["has_result"] = [len(r) > 0
for r in df_groundtruth["results"]]
df_groundtruth = df_groundtruth[df_groundtruth["has_result"]]
df = pd.merge(df, df_groundtruth, on=["query_key", "threshold"],
suffixes=("", "_ground_truth"))
prs = [get_precision_recall(result, ground_truth)
for result, ground_truth in \
zip(df["results"], df["results_ground_truth"])]
df["precision"] = [p for p, _ in prs]
df["recall"] = [r for _, r in prs]
df["fscore"] = [fscore(*pr) for pr in prs]

with open(args.benchmark_output) as f:
benchmark = json.load(f)
thresholds = sorted(list(set(df["threshold"])))
num_perms = sorted(list(set(df["num_perm"])))
num_parts = sorted(list(set(df["num_part"])))

num_perms = benchmark["num_perms"]
lsh_times = benchmark["lsh_times"]
ground_truth_results = [[x[0] for x in r] for r in benchmark["ground_truth_results"]]
lsh_fscores = []
for results in benchmark["lsh_results"]:
query_results = [[x[0] for x in r] for r in results]
lsh_fscores.append(average_fscore(query_results, ground_truth_results))

lsh_times = np.array([np.percentile(ts, 90)
for ts in lsh_times])*1000

fig, axes = plt.subplots(1, 2, figsize=(5*2, 4.5), sharex=True)
# Plot query fscore vs. num perm
axes[0].plot(num_perms, lsh_fscores, marker="+", label="LSH Ensemble")
axes[0].set_ylabel("Average F-Score")
axes[0].set_xlabel("# of Permmutation Functions")
axes[0].grid()
# Plot query time vs. num perm
axes[1].plot(num_perms, lsh_times, marker="+", label="LSH Ensemble")
axes[1].set_xlabel("# of Permutation Functions")
axes[1].set_ylabel("90 Percentile Query Time (ms)")
axes[1].grid()
axes[1].legend(loc="lower right")
plt.tight_layout()
fig.savefig("lshensemble_benchmark.png", pad_inches=0.05, bbox_inches="tight")
for i, num_perm in enumerate(num_perms):
# Plot precisions
for j, num_part in enumerate(num_parts):
sub = df[(df["num_part"] == num_part) & (df["num_perm"] == num_perm)].\
groupby("threshold")
precisions = sub["precision"].mean()
plt.plot(thresholds, precisions, label="num_part = {}".format(num_part))
plt.ylim(0.0, 1.0)
plt.xlabel("Thresholds")
plt.ylabel("Average Precisions")
plt.grid()
plt.legend()
plt.savefig("lshensemble_num_perm_{}_precision.png".format(num_perm))
plt.close()
# Plot recalls
for j, num_part in enumerate(num_parts):
sub = df[(df["num_part"] == num_part) & (df["num_perm"] == num_perm)].\
groupby("threshold")
recalls = sub["recall"].mean()
plt.plot(thresholds, recalls, label="num_part = {}".format(num_part))
plt.ylim(0.0, 1.0)
plt.xlabel("Thresholds")
plt.ylabel("Average Recalls")
plt.grid()
plt.legend()
plt.savefig("lshensemble_num_perm_{}_recall.png".format(num_perm))
plt.close()
# Plot fscores.
for j, num_part in enumerate(num_parts):
sub = df[(df["num_part"] == num_part) & (df["num_perm"] == num_perm)].\
groupby("threshold")
fscores = sub["fscore"].mean()
plt.plot(thresholds, fscores, label="num_part = {}".format(num_part))
plt.ylim(0.0, 1.0)
plt.xlabel("Thresholds")
plt.ylabel("Average F-Scores")
plt.grid()
plt.legend()
plt.savefig("lshensemble_num_perm_{}_fscore.png".format(num_perm))
plt.close()
# Plot query time.
times = []
for num_part in num_parts:
t = np.percentile(df[df["num_part"] == num_part]["query_time"], 90)
times.append(t * 1000.0)
plt.bar(num_parts, times)
plt.xlabel("Number of Partitions")
plt.ylabel("90 Percentile Query Time (ms)")
plt.grid()
plt.savefig("lshensemble_num_perm_{}_query_time.png".format(num_perm))
plt.close()
Loading

0 comments on commit 54d053c

Please sign in to comment.