Latest revision as of 23:05, 26 April 2017
Bit Sampling LSH
- LSH for the Hamming Distance
- can convert $L_1$ to Hamming distance
- NNs are usually the same for $L_1$ and $L_2$
- see Figiel et al. "The dimension of almost spherical sections of convex bodies." 1977. [1]
- so it can be ("sort of") used for Euclidean spaces
Hamming LSH
Suppose we have $P = \{ \mathbf p_i \}$ where $\mathbf p_i \in H^{d} = \{0, 1\}^{d}$ - i.t. points are in the (binary) Hamming Space of dimensionality $d$ (or $Cd$, in $C$ chunks of size $d$)
Hash Functions
- sample $L$ subsets $I_1, \ ... \ , I_L$ of size $k$ uniformly (with replacement) from \{1 \ ... \ d \}$
- let $g_j(\mathbf p)$ be the projection of $\mathbf p$ on $I_j$: it selects coordinate positions per $I_j$ and concatenates bits on these positions
Indexing
Preprocessing (indexing):
- store each $\mathbf p \in P$ in the bucket $g_j(\mathbf p)$ for all $j = 1 \, .. \, L$
- total number of resulting buckets may be large ($g_j(\mathbf p)$'s are sparse), so reduce the desparsify and reduce dimensionality using usual hashing
So have two levels of hashing:
- LSH Hash to map $\mathbf p$ to bucket $g_j(\mathbf p)$
- standard hash function to map $g_j(\mathbf p)$ to a hash table of size $M$
Pseudocode:
- Input: database $P$, number of hash tables $L$
- Output: $L$ hash tables $\tau_j$
- generate $L$ random hash functions $g_j(\cdot)$ - for each $\tau_j$
- for each $\mathbf p_i \in P$ and for each $(g_j, \tau_j)$:
- $\tau_j\big[ g_j(\mathbf p_i) \big] = \mathbf p_i$
Querying
Querying:
- given $\mathbf q$
- search all buckets $g_j(\mathbf q)$ until:
- either encounter $c \cdot L$ points ($c$ to be specified)
- or checked all $L$ indexes
- then for all candidates keep only $K$ closest
Pseudocode:
- input:
- a query point $\mathbf q$,
- number of nearest neighbors $K$,
- $L$ tables $\tau_i$
- output: $K$ (or less) NNs
- let $S \leftarrow \varnothing$ be a candidate list
- for each $(g_j, \tau_j)$
- $S \leftarrow S \cup \tau_j \big[ g_j(\mathbf q) \big]$
- return $K$NNs from $S$ (can be found using linear search)
Parameters
- $L$: number of subsets $I_j$'s and hence the # of hash functions
- $k$: size of $I_j$'s: $k$ is chosen s.t.
- it maximizes the probability that if $\mathbf p$ is close to $\mathbf q$, then they must end up in the same bucket
- it minimizes the prob that if $\mathbf p$ and $\mathbf q$ ending up in the same bucket when $\mathbf p$ is not close to $\mathbf q$
- $k \approx 700$ is a good value for $d \approx 64$
Embedding $L_1$ to Hamming
For converting from $L_1$ to Hamming:
- All $\mathbf p \in P$ are positive integers
- coordinates may be made all-positive by translating the origin
Let $C$ be the largest coordinate in all points in $P$
- then we can embed $P$ into a Hamming cube $\{0, 1\}^{Cd}$ where $d$ is the dimensionality of the original space
- transform each $\mathbf p = (p_1, \ ... \ , p_d$ into a binary vector:
- let $v(\mathbf p) = \big( \text{unary}_C(p_1), \ ... \ , \text{unary}_C(p_d) \big)$
- $\text{unary}_C(p)$ denotes the binary representation of $p$: a sequence of $p$ ones followed by $C - p$ zeros
- e.g. for $C = 10$ and $p = 4$, we'd have $\text{unary}_{10}(4) = 11110 \, 00000$
- for a vector $\mathbf p = (3, 4, 5)$ we would have $v(\mathbf p) = ( 11100 \, 00000 \, 11110 \, 00000 \, 11111 \, 00000 )$
Now,
- if $\mathbf p, \mathbf q \in \{1 \, .. \, C \}$
- then $d_1 \big(\mathbf p, \mathbf q \big) = d_H \big(v(\mathbf p), v(\mathbf q) \big)$
- i.e. this embedding preserves the $L_1$ distance between points
Mapping $\mathbf p \in \mathbb R^d$ to $\mathbf p' \in H^{Cd}$: (effective hashing calculation)
- choose $I$: a $k$-vector of indexes sampled from $\{0, \, .. \, C \}$
- we need to compute a projection on $I$
- for each component $p_i$ of $\mathbf p$ do:
- let $I^{(i)}$ denote coordinates of $I$ that correspond to $p_i$
- these are some of the $C$ coordinates of $p_i$ in the Hamming embedding that got selected in $I$
- order indexes inside $I^{(i)}$
- when we project $\mathbf p$ on $I^{(i)}$, the result is a monotone sequence of bits: there are 1's followed by 0's
- let $o_i$ be the number of 1's of $p_i$
- let $\mathbf p'_i$ denote the part of $\mathbf p'$ that corresponds to coordinate $p_i$
- so, to represent $\mathbf p'_i$ it's enough to know only $o_i$
- thus $\mathbf p'$ can be represented by $\{o_1, o_2, \ ... \ , o_d \}$
Computing $o_i$ fast:
- there's a way to compute $o_i$ fast
- finding $o_i$ is equivalent to finding the number of elements in sorted array $I^{(i)}$ which are smaller than $p_i$
- can be done via Binary Search in $O(\log C)$
- or even in constant time if we use a precomputed array of $C$ bits
Total projection time
- then total time to project $\mathbf p$ to $\mathbf p'$ is
- $O(d \, \log C)$ for Binary Search or
- $O(d)$ for precomputed arrays
Sources
- Gionis, Aristides, Piotr Indyk, and Rajeev Motwani. "Similarity search in high dimensions via hashing." 1999. [2]