Locality-sensitive hashing

Locality Sensitive Hashing (LSH) is a method of performing probabilistic dimension reduction of high-dimensional data. The basic idea is to hash the input items so that similar items are mapped to the same buckets with high probability (the number of buckets being much smaller than the universe of possible input items).

Definition

An LSH family ${\displaystyle {\mathcal {F}}}$  is defined for a metric space ${\displaystyle {\mathcal {M}}=(M,d)}$ , a threshold ${\displaystyle R>0}$  and an approximation factor ${\displaystyle c>1}$ . An LSH family ^{[1] [2]} ${\displaystyle {\mathcal {F}}}$  is a family of functions ${\displaystyle h:{\mathcal {M}}\to S}$  satisfying the following conditions for any two points ${\displaystyle p,q\in {\mathcal {M}}}$ , and a function ${\displaystyle h}$  chosen uniformly at random from ${\displaystyle {\mathcal {F}}}$ :

• if ${\displaystyle d(p,q)\leq R}$ , then ${\displaystyle h(p)=h(q)}$  (i.e.,${\displaystyle p}$  and ${\displaystyle q}$  collide) with probability at least ${\displaystyle P_{1}}$ ,
• if ${\displaystyle d(p,q)\geq cR}$ , then ${\displaystyle h(p)=h(q)}$  with probability at most ${\displaystyle P_{2}}$ .

A family is interesting when ${\displaystyle P_{1}>P_{2}}$ . Such a family ${\displaystyle {\mathcal {F}}}$  is called ${\displaystyle (R,cR,P_{1},P_{2})}$ -sensitive.

An alternative definition^[3] is defined with respect to a universe of items ${\displaystyle U}$  that have a similarity function ${\displaystyle \phi :U\times U\to [0,1]}$ . An LSH scheme is a family of hash functions ${\displaystyle H}$  coupled with a probability distribution ${\displaystyle D}$  over the functions such that a function ${\displaystyle h\in H}$  chosen according to ${\displaystyle D}$  satisfies the property that ${\displaystyle Pr_{h\in H}[h(a)=h(b)]=\phi (a,b)}$  for any ${\displaystyle a,b\in U}$ .

Applications

LSH has been applied to several problem domains including

• Near Duplicate Detection
• Image Similarity Identification
• Gene Expression Similarity Identification
• Audio Similarity Identification

Methods

Bit Sampling For Hamming Distance

One of the easiest ways to construct an LSH family is by bit sampling ^[2]. This approach works for the Hamming distance over d-dimensional vectors ${\displaystyle \{0,1\}^{d}}$ . Here, the family ${\displaystyle {\mathcal {F}}}$  of hash functions is simply the family of all the projections of points on one of the ${\displaystyle d}$  coordinates, i.e., ${\displaystyle {\mathcal {F}}=\{h:\{0,1\}^{d}\to \{0,1\}\mid h(x)=x_{i},i=1...d\}}$ , where ${\displaystyle x_{i}}$  is the ${\displaystyle i^{th}}$  coordinate of ${\displaystyle x}$ . A random function ${\displaystyle h}$  from ${\displaystyle {\mathcal {F}}}$  simply selects a random bit from input point. This family has the following parameters: ${\displaystyle P_{1}=1-R/d}$ , ${\displaystyle P_{2}=1-cR/d}$ .

Min-Wise Independent Permutations

Suppose ${\displaystyle U}$  is composed of subsets of some ground set of enumerable items ${\displaystyle S}$  and the similarity function of interest is the Jaccard index ${\displaystyle J}$ . If ${\displaystyle \pi }$  is a permutation on the indices of ${\displaystyle S}$ , for ${\displaystyle A\subseteq S}$  let ${\displaystyle h(A)=\min _{a\in A}\{\pi (a)\}}$ . Each possible choice of ${\displaystyle \pi }$  defines a single hash function ${\displaystyle h}$  mapping input sets to integers.

Define the function family ${\displaystyle H}$  to be the set of all such functions and let ${\displaystyle D}$  be the uniform distribution. Given two sets ${\displaystyle A,B\subseteq S}$  the event that ${\displaystyle h(A)=h(B)}$  corresponds exactly to the event that the minimizer of ${\displaystyle \pi }$  lies inside ${\displaystyle A\bigcap B}$ . As ${\displaystyle h}$  was chosen uniformly at random, ${\displaystyle Pr[h(A)=h(B)]=J(A,B)\,}$  and ${\displaystyle (H,D)\,}$  define an LSH scheme for the Jaccard index.

Because the symmetric group on n elements has size n!, choosing a truly random permutation from the full symmetric group is infeasible for even moderately sized n. Because of this fact, there has been significant work on finding a family of permutations that is "min-wise independent" - a permutation family for which each element of the ___domain has equal probability of being the minimum under a randomly chosen ${\displaystyle \pi }$ . It has been established that a min-wise independent family of permutations is at least of size ${\displaystyle lcm(1,2,...,n)\geq e^{n-o(n)}}$ ^[4] and that this boundary is tight^[5].

Because min-wise independent families are too big for practical applications, two variant notions of min-wise independence are introduced: restricted min-wise independent permutations families, and approximate min-wise independent families. Restricted min-wise independence is the min-wise independence property restricted to certain sets of cardinality at most k ^[6]. Approximate min-wise independence differs from the property by at most a fixed ${\displaystyle \epsilon }$ ^[7].

Random Projection

The random projection method of LSH is designed to approximate the cosine distance between vectors. The basic idea of this technique is to choose a random hyperplane (defined by a normal unit vector ${\displaystyle r}$ ) at the outset and use the hyperplane to hash input vectors.

Given an input vector ${\displaystyle v}$  and a hyperplane defined by ${\displaystyle r}$ , we let ${\displaystyle h(v)=sgn(v\cdot r)}$ . That is, ${\displaystyle h(v)=\pm 1}$  depending on which side of the hyperplane ${\displaystyle v}$  lies.

Each possible choice of ${\displaystyle r}$  defines a single function. Let ${\displaystyle H}$  be the set of all such functions and let ${\displaystyle D}$  be the uniform distribution once again. It is not difficult to prove that, for two vectors ${\displaystyle u,v}$ , ${\displaystyle Pr[h(u)=h(v)]=1-{\frac {\theta (u,v)}{\pi }}}$ , where ${\displaystyle \theta (u,v)}$  is the angle between ${\displaystyle u}$  and ${\displaystyle v}$ . ${\displaystyle 1-{\frac {\theta (u,v)}{\pi }}}$  is closely related to ${\displaystyle \cos(\theta (u,v))}$ .

In this instance hashing produces only a single bit. Two vectors' bits match with probability proportional to the cosine of the angle between them.

Stable Distributions

The hash function ^[8] ${\displaystyle h_{\mathbf {a} ,b}({\boldsymbol {\upsilon }}):{\mathcal {R}}^{d}\to {\mathcal {N}}}$  maps a d dimensional vector ${\displaystyle {\boldsymbol {\upsilon }}}$  onto a set of integers. Each hash function in the family is indexed by a choice of random ${\displaystyle \mathbf {a} }$  and ${\displaystyle b}$  where ${\displaystyle \mathbf {a} }$  is a d dimensional vector with entries chosen independently from a stable distribution and ${\displaystyle b}$  is a real number chosen uniformly from the range [0,r]. For a fixed ${\displaystyle \mathbf {a} ,b}$  the hash function ${\displaystyle h_{\mathbf {a} ,b}}$  is given by ${\displaystyle h_{\mathbf {a} ,b}({\boldsymbol {\upsilon }})=\left\lfloor {\frac {\mathbf {a} \cdot {\boldsymbol {\upsilon }}+b}{r}}\right\rfloor }$ .

LSH Algorithm for the Nearest Neighbor Search

One of the main application of LSH is to provide efficient nearest neighbor search algorithms. Consider any LSH family ${\displaystyle {\mathcal {F}}}$ . The algorithm has two main parameters: the width parameter ${\displaystyle k}$  and the number of hash tables ${\displaystyle L}$ .

In the first step, we define a new family ${\displaystyle {\mathcal {G}}}$  of hash functions ${\displaystyle g}$ , where each function ${\displaystyle g}$  is obtained by concatenating ${\displaystyle k}$  functions ${\displaystyle h_{1},...,h_{k}}$  from ${\displaystyle {\mathcal {F}}}$ , i.e., ${\displaystyle g(p)=[h_{1}(p),...,h_{k}(p)]}$ . In other words, a random hash function ${\displaystyle g}$  is obtained by concatenating ${\displaystyle k}$  randomly chosen hash functions from ${\displaystyle {\mathcal {H}}}$ . The algorithm then constructs ${\displaystyle L}$  hash tables, each corresponding to a different randomly chosen hash function ${\displaystyle g}$ .

In the preprocessing step we hash all ${\displaystyle n}$  points from the data set ${\displaystyle S}$  into each of the ${\displaystyle L}$  hash tables. Given that the resulting hash tables have only ${\displaystyle n}$  non-zero entries, one can reduce the amount of memory used per each hash table to ${\displaystyle O(n)}$  using standard hash functions.

Given a query point ${\displaystyle q}$ , the algorithm iterates over the ${\displaystyle L}$  hash functions ${\displaystyle g}$ . For each ${\displaystyle g}$  considered, it retrieves the data points that are hashed into the same bucket as ${\displaystyle q}$ . The process is stopped as soon as a point within distance ${\displaystyle cR}$  from ${\displaystyle q}$  is found.

Given the parameters ${\displaystyle k}$  and ${\displaystyle L}$ , the algorithm has the following performance guarantees:

• preprocessing time: ${\displaystyle O(nLkt)}$ , where ${\displaystyle t}$  is the time to evaluate a function ${\displaystyle h\in F}$  on an input point ${\displaystyle p}$ ;
• space: ${\displaystyle O(nL)}$ , plus the space for storing data points;
• query time: ${\displaystyle O(L(kt+dnP_{2}^{k}))}$ ;
• the algorithm succeeds in finding a point within distance ${\displaystyle cR}$  from ${\displaystyle q}$  (if it exists) with probability at least ${\displaystyle \Omega (\min\{1,LP_{1}^{k}\})}$ .

For a fixed approximation ratio ${\displaystyle c=1+\epsilon }$  and probabilities ${\displaystyle P_{1}}$  and ${\displaystyle P_{2}}$ , one can set ${\displaystyle k={\log n \over \log 1/P_{2}}}$  and ${\displaystyle L=n^{\rho }}$ , where ${\displaystyle \rho ={\log P_{1} \over \log P_{2}}}$ . Then one obtains the following performance guarantees:

• preprocessing time: ${\displaystyle O(n^{1+\rho }kt)}$ ;
• space: ${\displaystyle O(n^{1+\rho })}$ , plus the space for storing data points;
• query time: ${\displaystyle O(n^{\rho }(kt+d))}$ ;

See also

• Nearest neighbor search

Related Papers

1. Gionis, A. (1999). , "Similarity Search in High Dimensions via Hashing". Proceedings of the 25th Very Large Database (VLDB) Conference. {{cite journal}}: Check |url= value (help); Unknown parameter |coauthors= ignored (|author= suggested) (help)
2. Indyk, Piotr. (1998). , "Approximate Nearest Neighbors: Towards Removing the Curse of Dimensionality". Proceedings of 30th Symposium on Theory of Computing. {{cite journal}}: Check |url= value (help); Unknown parameter |coauthors= ignored (|author= suggested) (help)
3. Charikar, Moses S.. (2002). "Similarity Estimation Techniques from Rounding Algorithms". Proceedings of the 34th Annual ACM Symposium on Theory of Computing 2002: (ACM 1–58113–495–9/02/0005)…. doi:10.1145/509907.509965. Retrieved 2007-12-21. {{cite journal}}: Cite has empty unknown parameter: |coauthors= (help)
4. Broder, A.Z. (1998). "Min-wise independent permutations". Proceedings of the thirtieth annual ACM symposium on Theory of computing: 327–336. doi:10.1145/276698.276781. Retrieved 2007-11-14. {{cite journal}}: Unknown parameter |coauthors= ignored (|author= suggested) (help)
5. "An optimal construction of exactly min-wise independent permutations". Technical Report COMP98-62, IEICE, 1998. {{cite journal}}: Unknown parameter |coauthors= ignored (|author= suggested) (help)
6. Matoušek, J. (2002). "On Restricted Min-Wise Independence of Permutations". Preprint. Retrieved 2007-11-14. {{cite journal}}: Unknown parameter |coauthors= ignored (|author= suggested) (help)
7. Saks, M. (2000). "Low discrepancy sets yield approximate min-wise independent permutation families". Information Processing Letters. 73 (1–2): 29–32. doi:10.1016/S0020-0190(99)00163-5. Retrieved 2007-11-14. {{cite journal}}: Unknown parameter |coauthors= ignored (|author= suggested) (help)
8. Datar, M. (2004). "Locality-Sensitive Hashing Scheme Based on p-Stable Distributions". Proceedings of the Symposium on Computational Geometry. {{cite journal}}: Unknown parameter |coauthors= ignored (|author= suggested) (help)

• Alex Andoni's LSH homepage
• LSHKIT: A C++ Locality Sensitive Hashing Library