MinHashing vs SimHashing - cluster-analysis

Suppose I have five sets I'd like to cluster. I understand that the SimHashing technique described here:
https://moultano.wordpress.com/2010/01/21/simple-simhashing-3kbzhsxyg4467-6/
could yield three clusters ({A}, {B,C,D} and {E}), for instance, if its results were:
A -> h01
B -> h02
C -> h02
D -> h02
E -> h03
Similarly, the MinHashing technique described in the Chapter 3 of the MMDS book:
http://infolab.stanford.edu/~ullman/mmds/ch3.pdf
could also yield the same three clusters if its results were:
A -> h01 - h02 - h03
B -> h04 - h05 - h06
|
C -> h04 - h07 - h08
|
D -> h09 - h10 - h08
E -> h11 - h12 - h13
(Each set corresponds to a MH signature composed of three "bands", and two sets are grouped if at least one of their signature bands is matching. More bands would mean more matching chances.)
However I have several questions related to these:
(1) Can SH be understood as a single band version of MH?
(2) Does MH necessarily imply the use of a data structure like Union-Find to build the clusters?
(3) Am I right in thinking that the clusters, in both techniques, are actually "pre-clusters", in the sense that they are just sets of "candidate pairs"?
(4) If (3) is true, does it imply that I still have to do an O(n^2) search inside each "pre-cluster", to partition them further into "real" clusters? (which might be reasonable if I have a lot of small and fairly balanced pre-clusters, not so much otherwise)

SimHash and MinHash are both hashing algorithms that are able to map a set to a list of values which corresponds to the signature of the set.
In case of SimHash the list of values is just a list of bits (the values are either 0 or 1). In case of MinHash a value in the list represents the minimum hash value of all set elements relative to a given hash function which is typically a 32bit or 64bit value.
A major difference of both algorithms is the probability of hash collisions. In case of SimHash it is equal to the cosine similarity and in case of MinHash it is equal to the Jaccard similarity. Depending how you define the similarity between sets the one or the other algorithm could be more appropriate.
Regardless of the chosen hashing algorithm, the values of the calculated signature are equally partitioned over a certain number of bands. If the signatures of any two sets are identical within at least one band, the corresponding pair of sets is selected as candidate for similarity. (This means if n sets have the same signature within a band, there are O(n^2) candidate pairs just from this band.) Estimating the similarity of each candidate pair using the complete signature (including the values from other bands) and keeping only those pairs with an estimated similarity above a given threshold gives you all similar pairs of sets which finally define the final clustering.

Related

what is the difference between defining a vector using linspace and defining a vector using steps?

i am trying to learn the basics of matlab ,
i wanted to write a mattlab script ,
in this script i defined a vector x with a "d" step that it's length is (2*pi/1000)
and i wanted to plot two sin function according to x :
the first sin is with a frequency of 1, and the second sin frequency 10.3 ..
this is what i did:
d=(2*pi/1000);
x=-pi:d:pi;
first=sin(x);
second=sin(10.3*x);
plot(x,first,x,second);
my question:
what is the different between :
x=linspace(-pi,pi,1000);
and ..
d=(2*pi/1000);
x=-pi:d:pi;
? i am asking because i got confused since i think they both are the same but i think there is something wrong with my assumption ..
also is there is a more sufficient way to write sin function with a giveng frequency ?
The main difference can be summarizes as predefined size vs predefined step. And your example highlights it very well, indeed (1000 elements vs 1001 elements).
The linspace function produces a fixed-length vector (the length being defined by the third input argument, which defaults to 100) whose lower and upper limits are set, respectively, by the first and the second input arguments. The correct step to use is internally computed by the function itself (step = (x2 - x1) / n).
The colon operator defines a vector of elements whose values range between the specified lower and upper limits. The step, which is an optional parameter that defaults to 1, is the discriminant of the vector length. This means that the length of the result is determined by the number of steps that must be accomplished in order to reach the upper limit, starting from the lower one. On an side note, on this MathWorks thread you can find a very interesting discussion concerning the behavior of the colon operator in respect of floating-point management.
Another difference, related to the first one, is that linspace always includes the upper limit value while the colon operator only contains it if the specified step allows it (0:5:14 = [0 5 10]).
As a general rule, I prefer to use the former when I want to produce a vector of a predefined length (pretty obvious, isn't it?), and the latter when I need to create a sequence whose length has only a marginal relevance (or no relevance at all)

Matlab : Conceptual difficulty in How to create multiple hash tables in Locality sensitive Hashing

The key idea of Locality sensitive hashing (LSH) is that neighbor points, v are more likely
mapped to the same bucket but points far from each other are more likely mapped to different buckets. In using Random projection, if the the database contains N samples each of higher dimension d, then theory says that we must create k randomly generated hash functions, where k is the targeted reduced dimension denoted as g(**v**) = (h_1(v),h_2(v),...,h_k(v)). So, for any vector point v, the point is mapped to a k-dimensional vector with a g-function. Then the hash code is the vector of reduced length /dimension k and is regarded as a bucket. Now, to increase probability of collision, theory says that we should have L such g-functions g_1, g_2,...,g_L at random. This is the part that I do not understand.
Question : How to create multiple hash tables? How many buckets are contained in a hash table?
I am following the code given in the paper Sparse Projections for High-Dimensional Binary Codes by Yan Xia et. al Link to Code
In the file Coding.m
dim = size(X_train, 2);
R = randn(dim, bit);
% coding
B_query = (X_query*R >= 0);
B_base = (X_base*R >=0);
X_query is the set of query data each of dimension d and there are 1000 query samples; R is the random projection and bit is the target reduced dimensionality. The output of B_query and B_base are N strings of length k taking 0/1 values.
Does this way create multiple hash tables i.e. N is the number of hash tables? I am confused as to how. A detailed explanation will be very helpful.
How to create multiple hash tables?
LSH creates hash-table using (amplified) hash functions by concatenation:
g(p) = [h1(p), h2(p), · · · , hk (p)], hi ∈R H
g() is a hash function and it corresponds to one hashtable. So we map the data, via g() to that hashtable and with probability, the close ones will fall into the same bucket and the non-close ones will fall into different buckets.
We do that L times, thus we create L hashtables. Note that every g() is/should most likely to be different that the other g() hash functions.
Note: Large k ⇒ larger gap between P1, P2. Small P1 ⇒ larer L so as to find neighbors. A practical choice is L = 5 (or 6). P1 and P2 are defined in the image below:
How many buckets are contained in a hash table?
Wish I knew! That's a difficult question, how about sqrt(N) where N is the number of points in the dataset. Check this: Number of buckets in LSH
The code of Yan Xia
I am not familiar with that, but from what you said, I believe that the query data you see are 1000 in number, because we wish to pose 1000 queries.
k is the length of the strings, because we have to hash the query to see in which bucket of a hashtable it will be mapped. The points inside that bucket are potential (approximate) Nearest Neighbors.

Matlab non-linear binary Minimisation

I have to set up a phoneme table with a specific probability distribution for encoding things.
Now there are 22 base elements (each with an assigned probability, sum 100%), which shall be mapped on a 12 element table, which has desired element probabilities (sum 100%).
So part of the minimisation is to merge several base elements to get 12 table elements. Each base element must occur exactly once.
In addition, the table has 3 rows. So the same 12 element composition of the 22 base elements must minimise the error for 3 target vectors. Let's say the given target vectors are b1,b2,b3 (dimension 12x1), the given base vector is x (dimension 22x1) and they are connected by the unknown matrix A (12x22) by:
b1+err1=Ax
b2+err2=Ax
b3+err3=Ax
To sum it up: A is to be found so that dot_prod(err1+err2+err3, err1+err2+err3)=min (least squares). And - according to the above explanation - A must contain only 1's and 0's, while having exactly one 1 per column.
Unfortunately I have no idea how to approach this problem. Can it be expressed in a way different from the matrix-vector form?
Which tools in matlab could do it?
I think I found the answer while parsing some sections of the Matlab documentation.
First of all, the problem can be rewritten as:
errSum=err1+err2+err3=3Ax-b1-b2-b3
=> dot_prod(errSum, errSum) = min(A)
Applying the dot product (least squares) yields a quadratic scalar expression.
Syntax-wise, the fmincon tool within the optimization box could do the job. It has constraints parameters, which allow to force Aij to be binary and each column to be 1 in sum.
But apparently fmincon is not ideal for binary problems algorithm-wise and the ga tool should be used instead, which can be called in a similar way.
Since the equation would be very long in my case and needs to be written out, I haven't tried yet. Please correct me, if I'm wrong. Or add further solution-methods, if available.

Difference bloom filters and FM-sketches

What is the difference between bloom filters and hash sketches (also FM-sketches) and what is their use?
Hash sketches/Flajolet-Martin Sketches
Flajolet, P./Martin, G. (1985): Probabilistic counting algorithms for data base applications, in: Journal of Computer and System Sciences, Vol. 31, No. 2 (September 1985), pp. 182-209.
Durand, M./Flajolet, P. (2003): Loglog Counting of Large Cardinalities, in: Springer LNCS 2832, Algorithms ESA 2003, pp. 605–617.
Hash sketches are used to count the number of distinct elements in a set.
given:
a bit array B[] of length l
a (single) hash function h() that maps to [0,1,...2^l)
a function r() that gives the position of the least-significant 1-bit in the binary representation of its input (e.g. 000101 returns 1, 001000 returns 4)
insertion of element x:
pn := h(x) returns a pseudo-random number
apply r(pn) to get the position of the bit array to set to 1
since output of h() is pseudo-random every bit i is set to 1 ~n/(2^(i+1)) times
number of distinct elements in the set:
find the position p of the right-most 0 in the bit array
p = log2(n), solve for n to get the number of distinct element in the set;
the result might be up to 1.83 magnitudes off
usage:
in Data Mining, P2P/distributed applications, estimation of the document frequency, etc.
Bloom filters
Bloom, H. (1970): Space/time trade-offs in hash coding with allowable errors, in: Communications of the ACM, Vol. 13, No. 7 (July 1970), pp. 422-426.
Bloom filters are used to test whether an element is a member of a set.
given:
a bit array B[] of length m
k different hash functions h_k() that map to [0,...,m-1], i.e. to one of the position of the m-bit array
insertion of element x:
apply h_k to x (h_k(x)), for all k, i.e. you get k values
set the resulting bits in the array B to 1 (if already set to 1, don't change anything)
check if y is already in the set:
get the positions p_k to check using all the hash functions h_k (h_k(y)), i.e. for each function h_k you get a position p_k
if one of the positions p_k is set to 0 in the array B, the element y is definitively not in the set
if all positions given by p_k are 1, the element y might (!) be in the set
false positive rate is approximately (1 - e^(-kn/m))^k, no false negatives are possible!
by increasing the number of hashing functions, the false positive rate can be decreased; however, at the same time your bloom filter gets slower; the optimal value of k is k = (m/n)ln(2)
usage:
in the beginning used as a cheap filter in databases to filter out elements that do not match a query
various applications today, e.g. in Google BigTable, but also in networking for IP lookups, etc.
The Bloom Filter is a data structure used for Membership lookup while FM Sketch is primarily used for counting of elements. These two data structures provide the respective solutions optimizing over the space required to perform the lookup/computation and the trade off is the accuracy of the result.

Pearson perfect hashing

I'm trying to write a generator that produces Pearson perfect hashes. Note that I don't need a minimal perfect hash. Wikipedia says that a Pearson perfect hash can be found in O(|S|) time using a randomized algorithm (where S is the set of keys). However, I haven't been able to find such an algorithm online. Is this even possible?
Note: I don't want to use gperf/cmph/etc., I'd rather write my own implementation.
Pearson's original paper outlines an algorithm to construct a permutation table T for perfect hashing:
The table T at the heart of this new hashing function can sometimes be modified to produce a minimal, perfect hashing function over a modest list of words. In fact, one can usually choose the exact value of the function for a particular word. For example, Knuth [3] illustrates perfect hashing with an algorithm that maps a list of 31 common English words onto unique integers between −10 and 30. The table T presented in Table II maps these same 31 words onto the integers from 1 to 31 in alphabetic order.
Although the procedure for constructing the table in Table II is too involved to be detailed here, the following highlights will enable the interested reader to repeat the process:
A table T was constructed by pseudorandom permutation of the integers (0 ... 255).
One by one, the desired values were assigned to the words in the list. Each assignment was effected by exchanging two elements in the table.
For each word, the first candidate considered for exchange was T[h[n − 1] ⊕ C[n]], the last table element referenced in the computation of the hash function for that word.
A table element could not be exchanged if it was referenced during the hashing of a previously assigned word or if it was referenced earlier in the hashing of the same word.
If the necessary exchange was forbidden by Rule 4, attention was shifted to the previously referenced table element, T[h[n − 2] ⊕ C[n − 1]].
The procedure is not always successful. For example, using the ASCII character codes, if the word “a” hashes to 0 and the word “i” hashes to 15, it turns out that the word “in” must hash to 0. Initial attempts to map Knuth's 31 words onto the integers (0 ... 30) failed for exactly this reason. The shift to the range (1 ... 31) was an ad hoc tactic to circumvent this problem.
Does this tampering with T damage the statistical behavior of the hashing function? Not seriously. When the 26,662 dictionary entries are hashed into 256 bins, the resulting distribution is still not significantly different from uniform (χ² = 266.03, 255 d.f., p = 0.30). Hashing the 128 randomly selected dictionary words resulted in an average of 27.5 collisions versus 26.8 with the unmodified T. When this function is extended as described above to produce 16-bit hash indices, the same test produces a substantially greater number of collisions (4,870 versus 4,721 with the unmodified T), although the distribution still is not significantly different from uniform (χ² = 565.2, 532 d.f., p = 0.154).