I have word vectors split out into N groups, where each group contains M vectors. The problem is to find an optimal clustering of vectors where one and only one vector appears from each of the N groups.
As an example, say we had 3 groups of vectors as so:
{"Hot Dog", "Hot sauce", "Hotshot"}
{"Hamburger", "Hamburg", "Hamburgler"}
{"Pizza", "Pisa", "Piazza"}
The optimal cluster would be {"Hot Dog", "Hamburger", "Pizza"} because according to some function I have F(), these vectors are clustered closely to each other within the vector space I have defined.
I can arrive at this result in a brute force way by merely trying every combination. But as N and M grow, this becomes unfeasible. Is there a dynamic programming approach I could use? Any reference algorithm I can look up?
Thanks.
Edit:
To clarify my example above, each of those strings is like an ID for a vector so to rephrase it, Group 1 is {v1, v2, v3} Group 2 is {v4, v5, v6}, Group 3 is {v7, v8, v9}.
My desired output is {v1, v4, v7} but in a non-brute force way.
#m69's comment below correctly describes what I mean by a cluster - a group of vectors whose distance to one another as computed by some function F() is all within some threshold t.
Related
I studied on a fundamental SOM initialization and was looking to understand exactly how this process, PCI, works for initializing weight vectors on the map. My understanding is that for a two dimensional Map, this initialization method looks at the eigenvectors for the two largest eigenvalues of the data matrix and then uses the subspace spanned by these eigenvectors to initialize the map. Does that mean that in order to get the initial map weights, does this method take random linear combinations of the largest two eigenvectors in order to generate the map weights? Is there a patten?
For example, for 40 input data vectors on the map, does the lininit initialization method take combinations a1*[e1] + a2*[e2] where [e1] and [e2] are the two largest eigenvectors and a1 and a2 are random integers ranging from -3 to 3? Or is there a different mechanism? I was looking to make sure I knew exactly how lininit takes the two largest eigenvectors of the input data matrix and uses them to construct the initial weight vectors for the map.
The SOM creates a map that has the neighbourhood relationship between nearby nodes. Random initialisation does not help this process, since the nodes start randomly. Therefore, the idea of using the PCA initialisation is just a shortcut to get the map closer to the final state. This saves a lot of computation.
So how does this work? The first two principal components (PCs) are used. Set the initial weights as linear combination of the PCs. Rather than using random a1 and a2, the weights are set in a range that corresponds to the scale of the principal components.
For example, for a 5x3 map, a1 and a2 can both be in the range (-1, 1) with the relevant number of elements. In other words, for the 5x3 map, a1 = [-1.0 -0.5 0.0 0.5 1.0] and a2 = [-1.0 0.0 1.0], with 5 nodes and 3 nodes, respectively.
Then set each of the weights of nodes. For a rectangular SOM, each node has indices [m, n]. Use the values of a1[m] and a2[n]. Thus, for all m = [1 2 3 4 5] and n = [1 2 3]:
weight[m, n] = a1[m] * e1 + a2[n] * e2
That is how to initialize the weights using the principal components. This makes the initial state globally ordered, so now the SOM algorithm is used to create the local ordering.
The Principal Component part of the name is a reference to https://en.wikipedia.org/wiki/Principal_component_analysis.
Here is the idea. You start with data points placed at vectors of many underlying factors. But they may be correlated in your data. So, for example, if you're measuring height, weight, blood pressure, etc, you expect that tall people will weigh more. But what you want to do is replace this with vectors of factors that are not correlated with each other in your data.
So your principal component is a vector of length 1 which is as strongly correlated as possible with the variation in your dataset.
Your secondary component is the vector of length 1 at right angles to the first which is as strongly correlated as possible with the rest of the variation in your data set.
Your tertiary component is the vector of length 1 at right angles to the first two which is as strongly correlated as possible with the rest of the variation in your data set.
And so on.
In practice you may start with many factors, but most of the information is captured in just the first few. For example in the results of intelligence testing the first component is IQ and the second is the difference between how you are at verbal and quantitative reasoning.
How this applies to SOM initialization is that a simple linear model built off of PCA analysis is a pretty good guess for the answer that you're looking for, so starting there reduces how much work you have to do to finish getting the answer.
I'm working on a problem that requires an array (dA[j], j=-N..N) to be calculated from the values of another array (A[i], i=-N..N) based on a conservation of momentum rule (x+y=z+j). This means that for a given index j for all the valid combinations of (x,y,z) I calculate A[x]A[y]A[z]. dA[j] is equal to the sum of these values.
I'm currently precomputing the valid indices for each dA[j] by looping x=-N...+N,y=-N...+N and calculating z=x+y-j and storing the indices if abs(z) <= N.
Is there a more efficient method of computing this?
The reason I ask is that in future I'd like to also be able to efficiently find for each dA[j] all the terms that have a specific A[i]. Essentially to be able to compute the Jacobian of dA[j] with respect to dA[i].
Update
For the sake of completeness I figured out a way of doing this without any if statements: if you parametrize the equation x+y=z+j given that j is a constant you get the equation for a plane. The constraint that x,y,z need to be integers between -N..N create boundaries on this plane. The points that define this boundary are functions of N and j. So all you have to do is loop over your parametrized variables (s,t) within these boundaries and you'll generate all the valid points by using the vectors defined by the plane (s*u + t*v + j*[0,0,1]).
For example, if you choose u=[1,0,-1] and v=[0,1,1] all the valid solutions for every value of j are bounded by a 6 sided polygon with points (-N,-N),(-N,-j),(j,N),(N,N),(N,-j), and (j,-N).
So for each j, you go through all (2N)^2 combinations to find the correct x's and y's such that x+y= z+j; the running time of your application (per j) is O(N^2). I don't think your current idea is bad (and after playing with some pseudocode for this, I couldn't improve it significantly). I would like to note that once you've picked a j and a z, there is at most 2N choices for x's and y's. So overall, the best algorithm would still complete in O(N^2).
But consider the following improvement by a factor of 2 (for the overall program, not per j): if z+j= x+y, then (-z)+(-j)= (-x)+(-y) also.
Say I have a Group data structure which contains a list of Element objects, such that each group has a unique set of elements.:
public class Group
{
public List<Element> Elements;
}
and say I have a list of populations who require certain elements, in such a way that each population has a unique set of required elements:
public class Population
{
public List<Element> RequiredElements;
}
I have an unlimited quantity of each defined Group, i.e. they are not consumed by populations.
Say I am looking at a particular Population. I want to find the best possible match of groups such that there is minimum excess elements, and no unmatched elements.
For example: I have a population which needs wood, steel, grain, and coal. The only groups available are {wood, herbs}, {steel, coal, oil}, {grain, steel}, and {herbs, meat}.
The last group - {herbs, meat} isn't required at all by my population so it isn't used. All others are needed, but herbs and oil are not required so it is wasted. Furthermore, steel exists twice in the minimum set, so one lot of steel is also wasted. The best match in this example has a wastage of 3.
So for a few hundred Population objects, I need to find the minimum wastage best match and compute how many elements are wasted.
How do I even begin to solve this? Once I have found a match, counting the wastage is trivial. Finding the match in the first place is hard. I could enumerate all possibilities but with a few thousand populations and many hundreds of groups, it's quite a task. Especially considering this whole thing sits inside each iteration of a simulated annealing algorithm.
I'm wondering whether I can formulate the whole thing as a mixed-integer program and call a solver like GLPK at each iteration.
I hope I have explained the problem correctly. I can clarify anything that's unclear.
Here's my binary program, for those of you interested...
x is the decision vector, an element of {0,1}, which says that the population in question does/doesn't receive from group i. There is an entry for each group.
b is the column vector, an element of {0,1}, which says which resources the population in question does/doesn't need. There is an entry for each resource.
A is a matrix, an element of {0,1}, which says what resources are in what groups.
The program is:
Minimise: ((Ax - b)' * 1-vector) + (x' * 1-vector);
Subject to: Ax >= b;
The constraint just says that all required resources must be satisfied. The objective is to minimise all excess and the total number of groups used. (i.e. 0 excess with 1 group used is better than 0 excess with 5 groups used).
You can formulate an integer program for each population P as follows. Use a binary variable xj to denote whether group j is chosen or not. Let A be a binary matrix, such that Aij is 1 if and only if item i is present in group j. Then the integer program is:
min Ei,j (xjAij)
s.t. Ej xjAij >= 1 for all i in P.
xj = 0, 1 for all j.
Note that you can obtain the minimum wastage by subtracting |P| from the optimal solution of the above IP.
Do you mean the Maximum matching problem?
You need to build a bipartite graph, where one of the sides is your populations and the other is groups, and edge exists between group A and population B if it have it in its set.
To find maximum edge matching you can easily use Kuhn algorithm, which is greatly described here on TopCoder.
But, if you want to find mimimum edge dominating set (the set of minimum edges that is covering all the vertexes), the problem becomes NP-hard and can't be solved in polynomial time.
Take a look at the weighted set cover problem, I think this is exactly what you described above. A basic description of the (unweighted) problem can be found here.
Finding the minimal waste as you defined above is equivalent to finding a set cover such that the sum of the cardinalities of the covering sets is minimal. Hence, the weight of each set (=a group of elements) has to be defined equal to its cardinality.
Since even the unweighted the set cover problem is NP-complete, it is not likely that an efficient algorithm for your problem instances exist. Maybe a good greedy approximation algorithm will be sufficient or your purpose? Googling weighted set cover provides several promising results, e.g. this script.
I have a graph-theoretic (which is also related to combinatorics) problem that is illustrated below, and wonder what is the best approach to design an algorithm to solve it.
Given 4 different graphs of 6 nodes (by different, I mean different structures, e.g. STAR, LINE, COMPLETE, etc), and 24 unique objects, design an algorithm to assign these objects to these 4 graphs 4 times, so that the number of repeating neighbors on the graphs over the 4 assignments is minimized. For example, if object A and B are neighbors on 1 of the 4 graphs in one assignment, then in the best case, A and B will not be neighbors again in the other 3 assignments.
Obviously, the degree to which such minimization can go is dependent on the specific graph structures given. But I am more interested in a general solution here so that given any 4 graph structures, such minimization is guaranteed as the result of the algorithm.
Any suggestion/idea of solving this problem is welcome, and some pseudo-code may well be sufficient to illustrate the design. Thank you.
Representation:
You have 24 elements, I will name this elements from A to X (24 first letters).
Each of these elements will have a place in one of the 4 graphs. I will assign a number to the 24 nodes of the 4 graphs from 1 to 24.
I will identify the position of A by a 24-uple =(xA1,xA2...,xA24), and if I want to assign A to the node number 8 for exemple, I will write (xa1,Xa2..xa24) = (0,0,0,0,0,0,0,1,0,0...0), where 1 is on position 8.
We can say that A =(xa1,...xa24)
e1...e24 are the unit vectors (1,0...0) to (0,0...1)
note about the operator '.':
A.e1=xa1
...
X.e24=Xx24
There are some constraints on A,...X with these notations :
Xii is in {0,1}
and
Sum(Xai)=1 ... Sum(Xxi)=1
Sum(Xa1,xb1,...Xx1)=1 ... Sum(Xa24,Xb24,... Xx24)=1
Since one element can be assign to only one node.
I will define a graph by defining the neighbors relation of each node, lets say node 8 has neighbors node 7 and node 10
to check that A and B are neighbors on node 8 for exemple I nedd:
A.e8=1 and B.e7 or B.e10 =1 then I just need A.e8*(B.e7+B.e10)==1
in the function isNeighborInGraphs(A,B) I test that for every nodes and I get one or zero depending on the neighborhood.
Notations:
4 graphs of 6 nodes, the position of each element is defined by an integer from 1 to 24.
(1 to 6 for first graph, etc...)
e1... e24 are the unit vectors (1,0,0...0) to (0,0...1)
Let A, B ...X be the N elements.
A=(0,0...,1,...,0)=(xa1,xa2...xa24)
B=...
...
X=(0,0...,1,...,0)
Graph descriptions:
IsNeigborInGraphs(A,B)=A.e1*B.e2+...
//if 1 and 2 are neigbors in one graph
for exemple
State of the system:
L(A)=[B,B,C,E,G...] // list of
neigbors of A (can repeat)
actualise(L(A)):
for element in [B,X]
if IsNeigbotInGraphs(A,Element)
L(A).append(Element)
endIf
endfor
Objective functions
N(A)=len(L(A))+Sum(IsneigborInGraph(A,i),i in L(A))
...
N(X)= ...
Description of the algorithm
start with an initial position
A=e1... X=e24
Actualize L(A),L(B)... L(X)
Solve this (with a solveur, ampl for
exemple will work I guess since it's
a nonlinear optimization
problem):
Objective function
min(Sum(N(Z),Z=A to X)
Constraints:
Sum(Xai)=1 ... Sum(Xxi)=1
Sum(Xa1,xb1,...Xx1)=1 ...
Sum(Xa24,Xb24,... Xx24)=1
You get the best solution
4.Repeat step 2 and 3, 3 more times.
If all four graphs are K_6, then the best you can do is choose 4 set partitions of your 24 objects into 4 sets each of cardinality 6 so that the pairwise intersection of any two sets has cardinality at most 2. You can do this by choosing set partitions that are maximally far apart in the Hasse diagram of set partitions with partial order given by refinement. The general case is much harder, but perhaps you can still begin with this crude approximation of a solution and then be clever with which vertex is assigned which object in the four assignments.
Assuming you don't want to cycle all combinations and calculate the sum every time and choose the lowest, you can implement a minimum problem (solved depending on your constraints using either a linear programming solver i.e. symplex algorithm engines or a non-linear solver, much harder talking in terms of time) with constraints on your variables (24) depending on the shape of your path. You can also use free software like LINGO/LINDO to create rapidly a decision theory model and test its correctness (you need decision theory notions though)
If this has anything to do with the real world, then it's unlikely that you absolutely must have a solution that is the true minimum. Close to the minimum should be good enough, right? If so, you could repeatedly randomly make the 4 assignments and check the results until you either run out of time or have a good-enough solution or appear to have stopped improving your best solution.
This is intended to be a more concrete, easily expressable form of my earlier question.
Take a list of words from a dictionary with common letter length.
How to reorder this list tto keep as many letters as possible common between adjacent words?
Example 1:
AGNI, CIVA, DEVA, DEWA, KAMA, RAMA, SIVA, VAYU
reorders to:
AGNI, CIVA, SIVA, DEVA, DEWA, KAMA, RAMA, VAYU
Example 2:
DEVI, KALI, SHRI, VACH
reorders to:
DEVI, SHRI, KALI, VACH
The simplest algorithm seems to be: Pick anything, then search for the shortest distance?
However, DEVI->KALI (1 common) is equivalent to DEVI->SHRI (1 common)
Choosing the first match would result in fewer common pairs in the entire list (4 versus 5).
This seems that it should be simpler than full TSP?
What you're trying to do, is calculate the shortest hamiltonian path in a complete weighted graph, where each word is a vertex, and the weight of each edge is the number of letters that are differenct between those two words.
For your example, the graph would have edges weighted as so:
DEVI KALI SHRI VACH
DEVI X 3 3 4
KALI 3 X 3 3
SHRI 3 3 X 4
VACH 4 3 4 X
Then it's just a simple matter of picking your favorite TSP solving algorithm, and you're good to go.
My pseudo code:
Create a graph of nodes where each node represents a word
Create connections between all the nodes (every node connects to every other node). Each connection has a "value" which is the number of common characters.
Drop connections where the "value" is 0.
Walk the graph by preferring connections with the highest values. If you have two connections with the same value, try both recursively.
Store the output of a walk in a list along with the sum of the distance between the words in this particular result. I'm not 100% sure ATM if you can simply sum the connections you used. See for yourself.
From all outputs, chose the one with the highest value.
This problem is probably NP complete which means that the runtime of the algorithm will become unbearable as the dictionaries grow. Right now, I see only one way to optimize it: Cut the graph into several smaller graphs, run the code on each and then join the lists. The result won't be as perfect as when you try every permutation but the runtime will be much better and the final result might be "good enough".
[EDIT] Since this algorithm doesn't try every possible combination, it's quite possible to miss the perfect result. It's even possible to get caught in a local maximum. Say, you have a pair with a value of 7 but if you chose this pair, all other values drop to 1; if you didn't take this pair, most other values would be 2, giving a much better overall final result.
This algorithm trades perfection for speed. When trying every possible combination would take years, even with the fastest computer in the world, you must find some way to bound the runtime.
If the dictionaries are small, you can simply create every permutation and then select the best result. If they grow beyond a certain bound, you're doomed.
Another solution is to mix the two. Use the greedy algorithm to find "islands" which are probably pretty good and then use the "complete search" to sort the small islands.
This can be done with a recursive approach. Pseudo-code:
Start with one of the words, call it w
FindNext(w, l) // l = list of words without w
Get a list l of the words near to w
If only one word in list
Return that word
Else
For every word w' in l do FindNext(w', l') //l' = l without w'
You can add some score to count common pairs and to prefer "better" lists.
You may want to take a look at BK-Trees, which make finding words with a given distance to each other efficient. Not a total solution, but possibly a component of one.
This problem has a name: n-ary Gray code. Since you're using English letters, n = 26. The Wikipedia article on Gray code describes the problem and includes some sample code.