profanity-detector

Profanity Detector

We use this example to test the efficiency of Bloom Filter on Word Dictionary. We used the words_alpha.txt from the repository dwyl/english-words that holds 370105 words.

We split the words into two segments in the ratio 9:1.

1. corpus - these would be correct english words
2. test - these would be words that are considered as profane (for benchmarking)

The total size of the dataset is 3864811 bytes ~ 3.8 MB. Corpus contains 333096 words while the test (profane list) contains 37009 words.

For set (map) and bloom filters of various length, we compute and analyze the false positivity rate.

How to run?

Download the words_alpha.txt file and store it at location examples/profanity-detector/words.txt.

$ go run examples/profanity-detector/main.go
$ pip install graph-cli
$ graph size-bench.csv --xcol 2 --ycol 4 -o size-fp.png
$ graph hashfn-bench.csv --xcol 3 --ycol 4 -o hash-fp.png
$ graph hashfn-bench.csv --xcol 3 --ycol 5 -o hash-time.png
$ graph dbf-region-bench.csv --xcol 6 --ycol 4 -o dbf-region.png

Results

The following result shows

1. storage - regular set or bloom filter
2. len - size occupied by the set or bloom filter
3. frac size - fraction of the size occupied as compared to the size of the data
4. false positivity rate on the test (profane) list.

We can see, how bloom filter's false positivity rate drops significantly with the increase in the length of the filter. At the bloom filter length of 512 KB which is mere 13.25% of the data size, we get a false positivity rate of 2.33%.

We also see how the false positivity rate drops by inncreasing the number of hash functions but after a certain limit it adversely affects the false positivity. Time taken to check the existence increases with increasing the number of hash functions.

Bloom Filter Size vs False Positive Rate

We can see how the size of bloom filter is related to the observed false positivity rate. Larger the size, smaller the false positivity rate.

2022/11/08 14:11:15 total length of data 3864812 bytes
2022/11/08 14:11:15 number of words in corpus 333097
2022/11/08 14:11:15 number of words in test 37009
2022/11/08 14:11:15 size=1024,size_frac=0.026495,num_hash=2,fp_rate=100.000000,time_taken_ms=98
2022/11/08 14:11:15 size=2048,size_frac=0.052991,num_hash=2,fp_rate=100.000000,time_taken_ms=89
2022/11/08 14:11:15 size=3072,size_frac=0.079486,num_hash=2,fp_rate=100.000000,time_taken_ms=79
2022/11/08 14:11:15 size=4096,size_frac=0.105982,num_hash=2,fp_rate=100.000000,time_taken_ms=92
2022/11/08 14:11:15 size=5120,size_frac=0.132477,num_hash=2,fp_rate=100.000000,time_taken_ms=101
2022/11/08 14:11:15 size=10240,size_frac=0.264955,num_hash=2,fp_rate=99.959469,time_taken_ms=80
2022/11/08 14:11:15 size=20480,size_frac=0.529909,num_hash=2,fp_rate=96.619741,time_taken_ms=116
2022/11/08 14:11:15 size=30720,size_frac=0.794864,num_hash=2,fp_rate=87.203113,time_taken_ms=97
2022/11/08 14:11:16 size=40960,size_frac=1.059819,num_hash=2,fp_rate=75.349239,time_taken_ms=89
2022/11/08 14:11:16 size=51200,size_frac=1.324773,num_hash=2,fp_rate=64.595098,time_taken_ms=103
2022/11/08 14:11:16 size=102400,size_frac=2.649547,num_hash=2,fp_rate=30.957335,time_taken_ms=100
2022/11/08 14:11:16 size=153600,size_frac=3.974320,num_hash=2,fp_rate=17.720014,time_taken_ms=89
2022/11/08 14:11:16 size=204800,size_frac=5.299093,num_hash=2,fp_rate=11.489097,time_taken_ms=106
2022/11/08 14:11:16 size=256000,size_frac=6.623867,num_hash=2,fp_rate=7.844038,time_taken_ms=104
2022/11/08 14:11:16 size=307200,size_frac=7.948640,num_hash=2,fp_rate=5.652679,time_taken_ms=92
2022/11/08 14:11:16 size=358400,size_frac=9.273414,num_hash=2,fp_rate=4.436759,time_taken_ms=99
2022/11/08 14:11:16 size=409600,size_frac=10.598187,num_hash=2,fp_rate=3.528871,time_taken_ms=110
2022/11/08 14:11:16 size=358400,size_frac=9.273414,num_hash=2,fp_rate=4.436759,time_taken_ms=88
2022/11/08 14:11:17 size=512000,size_frac=13.247734,num_hash=2,fp_rate=2.437245,time_taken_ms=87

Number of Hash Functions vs False Positive Rate and Time

Here we see how the number of hash functions used in a bloom filter affects the false positivity rate and also time. The FP rate decreases initially but after a certain stage it saturates and then increases.

The time takes for the operation increases almost linearly with the number of hash functions given how expensive hash computation can get.

2022/11/08 14:11:15 total length of data 3864812 bytes
2022/11/08 14:11:15 number of words in corpus 333097
2022/11/08 14:11:15 number of words in test 37009
2022/11/08 14:11:17 size=512000,size_frac=13.247734,num_hash=1,fp_rate=8.119647,time_taken_ms=97
2022/11/08 14:11:17 size=512000,size_frac=13.247734,num_hash=2,fp_rate=2.437245,time_taken_ms=120
2022/11/08 14:11:17 size=512000,size_frac=13.247734,num_hash=3,fp_rate=1.075414,time_taken_ms=106
2022/11/08 14:11:17 size=512000,size_frac=13.247734,num_hash=4,fp_rate=0.605258,time_taken_ms=114
2022/11/08 14:11:17 size=512000,size_frac=13.247734,num_hash=5,fp_rate=0.413413,time_taken_ms=142
2022/11/08 14:11:17 size=512000,size_frac=13.247734,num_hash=6,fp_rate=0.337756,time_taken_ms=158
2022/11/08 14:11:17 size=512000,size_frac=13.247734,num_hash=7,fp_rate=0.289119,time_taken_ms=214
2022/11/08 14:11:18 size=512000,size_frac=13.247734,num_hash=8,fp_rate=0.259396,time_taken_ms=235
2022/11/08 14:11:18 size=512000,size_frac=13.247734,num_hash=9,fp_rate=0.262098,time_taken_ms=192
2022/11/08 14:11:18 size=512000,size_frac=13.247734,num_hash=10,fp_rate=0.272907,time_taken_ms=242
2022/11/08 14:11:18 size=512000,size_frac=13.247734,num_hash=11,fp_rate=0.286417,time_taken_ms=237
2022/11/08 14:11:19 size=512000,size_frac=13.247734,num_hash=12,fp_rate=0.329650,time_taken_ms=236
2022/11/08 14:11:19 size=512000,size_frac=13.247734,num_hash=13,fp_rate=0.389095,time_taken_ms=286
2022/11/08 14:11:19 size=512000,size_frac=13.247734,num_hash=14,fp_rate=0.478262,time_taken_ms=318
2022/11/08 14:11:19 size=512000,size_frac=13.247734,num_hash=15,fp_rate=0.491772,time_taken_ms=288