p-stable-lsh-python 0.0.3

A 1&2-stable Locality-Sensitive Hashing implementation in Python

P-stable-LSH

The package is one implementation of paper Locality-Sensitive Hashing Scheme Based on p-Stable Distributions in SCG’2014. P-stable-lsh a novel Locality-Sensitive Hashing scheme for the Approximate Nearest Neighbor Problem under $L_p$ norm, based on p-stable distributions.

Note: This code is used as the practice of the paper, and there are few optimizations. Sharing is for communication and learning. If it is a high-performance scenario, please optimize as appropriate.

Installation

pip install p-stable-lsh-python

Usage

Example

The following example shows all features of the package, If you want to know the details, please refer to the source code and comments.

import numpy as np
import p_stable_lsh.pstable as psl

dim = 100 # vector dimension
data = [np.random.random(dim) for _ in range(2)] # generate two vectors

r = 50.0 # the parameter $r$ in paper

m1 = psl.pstable(r, dim, metric_dim=1)
m1.lsh(data[0])
m2 = psl.pstable(r, dim, metric_dim=1)
m2.lsh(data[1])
print(m1.md(m2)) # estimate value
print(m1.p(np.average(sum(np.abs(data[0]-data[1]))))) # theoretical(true) value

m1 = psl.pstable(r, dim, metric_dim=2)
m1.lsh(data[0])
m2 = psl.pstable(r, dim, metric_dim=2)
m2.lsh(data[1])
print(m1.md(m2)) # estimate value
print(m1.p(np.sqrt(sum([i**2 for i in data[0]-data[1]])))) # theoretical(true) value

Use case

Define the parameter $r$ and vector dimension for preparing the test data:

import numpy as np
import p_stable_lsh.pstable as psl
r = 50.0 # the parameter $r$ in paper
dim = 100 # vector dimension
data = [np.random.random(dim) for _ in range(3)] # generate two vectors

Instantiate pstable object with specific dimension space. ($L_1$ for example)

m1 = psl.pstable(r, dim, metric_dim=1)
m2 = psl.pstable(r, dim, metric_dim=1)

Hash vectors with p-stable LSH function.

m1.lsh(data[0])
m2.lsh(data[1])

Estimate distance between two object.

m1.md(m2) # estimate value

Show the ground truth distance probability using integration shown in paper.

l1_distance = np.average(sum(np.abs(data[0]-data[1])))
m1.p(l1_distance)

Another way to instantiate pstable object with hash values.

m3 = psl.pstable(r, dim, metric_dim=1, hashvalues=m1.hashvalues)
m3.md(m2)

Update object hash values with different vector.

m2.lsh(data[2])
m3.md(m2)

Parameters

The parameter $r$ in paper determines the relationship between the distance (in metric space) between vectors and the probability of collision (the .md method). The following shows different probability results varied by $r$.

L1 metric space

Data points code

import p_stable_lsh.pstable as psl
r_list = [10, 30, 50, 100, 300, ]
result = []
for r in r_list:
    tmp = []
    m = psl.pstable(r, dim, metric_dim=1)
    for i in range(1, 800):
        tmp.append(m.p(i))
    result.append(tmp)

L2 metric space