HyperMinSketch

Besides being a compact and pretty speedy HyperLogLog implementation for cardinality counting, this modified HyperLogLog allows intersection and similarity estimation of different HyperLogLogs.

Details

A simple implementation of HyperLogLog (LogLog-Beta to be specific):

16 bit registers instead of 6 bit, the new 10 bit are for b-bit signatures
Similarity function estimates Jaccard indices (a number between 0-1) of 0.01 for set cardinalities on the order of 1e9 with accuracy around 5%
Intersection applies the Jaccard index on the union of the sets to return the intersecting set cardinality

The work is based on "HyperMinHash: Jaccard index sketching in LogLog space - Yun William Yu, Griffin M. Weber"

Example Usage

sk1 := hyperminhash.New()
sk2 := hyperminhash.New()

for i := 0; i < 10000; i++ {
    sk1.Add([]byte(strconv.Itoa(i)))
}

sk1.Cardinality() // 10001 (should be 10000)

for i := 3333; i < 23333; i++ {
    sk2.Add([]byte(strconv.Itoa(i)))
}

sk2.Cardinality()     // 19977 (should be 20000)
sk1.Similarity(sk2)   // 0.284589082 (should be 0.2857326533)
sk1.Intersection(sk2) // 6623 (should be 6667)

sk1.Merge(sk2)
sk1.Cardinality() // 23271 (should be 23333)

Results

Max Cardinality 1e3

Set1	HLL1	Set2	HLL2	S1 ∪ S2	HLL1 ∪ HLL2	S1 ∩ S2	HLL1 ∩ HLL2
350	352	752	752	831	832	271 (32.611312%)	274 (32.932692%)
746	748	591	590	834	835	503 (60.311751%)	501 (60.000000%)
248	248	789	791	897	899	140 (15.607581%)	144 (16.017798%)
9	9	818	818	824	825	3 (0.364078%)	3 (0.363636%)
408	411	412	408	771	771	49 (6.355383%)	47 (6.095979%)

Max Cardinality 1e4

Set1	HLL1	Set2	HLL2	S1 ∪ S2	HLL1 ∪ HLL2	S1 ∩ S2	HLL1 ∩ HLL2
2126	2138	1162	1158	3063	3060	225 (7.345739%)	223 (7.287582%)
7767	7706	7054	7064	8889	8887	5932 (66.734166%)	5888 (66.254079%)
842	844	5183	5135	5880	5842	145 (2.465986%)	135 (2.310852%)
6833	6791	664	666	7410	7345	87 (1.174089%)	89 (1.211709%)
1814	1820	6214	6169	7697	7639	331 (4.300377%)	320 (4.189030%)

Max Cardinality 1e5

Set1	HLL1	Set2	HLL2	S1 ∪ S2	HLL1 ∪ HLL2	S1 ∩ S2	HLL1 ∩ HLL2
29667	29540	88700	88167	92444	91667	25923 (28.041842%)	25036 (27.311901%)
79242	78731	30216	30137	83502	82953	25956 (31.084285%)	25995 (31.337022%)
57830	57223	79550	79194	82112	81595	55268 (67.308067%)	54684 (67.018812%)
64610	63501	21696	21729	75895	74816	10411 (13.717636%)	10083 (13.477064%)
92204	91453	96417	95556	165025	163370	23596 (14.298440%)	24130 (14.770154%)

Max Cardinality 1e6

Set1	HLL1	Set2	HLL2	S1 ∪ S2	HLL1 ∪ HLL2	S1 ∩ S2	HLL1 ∩ HLL2
150443	149810	974366	979514	1088517	1096991	36292 (3.334077%)	37417 (3.410876%)
156337	155347	19083	19070	167353	165433	8067 (4.820350%)	8017 (4.846071%)
800969	802044	51053	51244	851388	853396	634 (0.074467%)	511 (0.059878%)
176155	174707	520111	516822	570092	569289	126174 (22.132217%)	123766 (21.740452%)
485954	481362	967341	972651	1081990	1091296	371305 (34.316861%)	376007 (34.455088%)

Max Cardinality 1e7

Set1	HLL1	Set2	HLL2	S1 ∪ S2	HLL1 ∪ HLL2	S1 ∩ S2	HLL1 ∩ HLL2
7132942	7150720	122116	121539	7243153	7261709	11905 (0.164362%)	12550 (0.172824%)
8646240	8649049	1277784	1295017	9821480	9854242	102544 (1.044079%)	99163 (1.006298%)
4192390	4164637	2788913	2779975	4526476	4499897	2454827 (54.232630%)	2454356 (54.542493%)
9803344	9826412	1705700	1715798	10255010	10262719	1254034 (12.228501%)	1273821 (12.412120%)
1308849	1322604	9940327	9971519	11179030	11201850	70146 (0.627478%)	80717 (0.720568%)

Max Cardinality 1e8

Set1	HLL1	Set2	HLL2	S1 ∪ S2	HLL1 ∪ HLL2	S1 ∩ S2	HLL1 ∩ HLL2
13237748	13298469	57073758	57124720	59474437	59394847	10837069 (18.221390%)	11143669 (18.762013%)
90757994	88576114	5717797	5701796	95061178	93016636	1414613 (1.488108%)	1350058 (1.451416%)
60150663	60033013	79238333	77672994	110438475	108311818	28950521 (26.214162%)	27666946 (25.543792%)
30187492	30718889	37756209	37153655	67443566	66938074	500135 (0.741561%)	447406 (0.668388%)
53196095	53461989	48344583	47535284	93284291	91321031	8256387 (8.850780%)	8036467 (8.800237%)

massyah/hyperminhash