Efficient Cartesian Product algorithm_问答_开发者

Can somebody please demonstrate for me a more efficient Cartesian product algorithm than the one I am using currently (assuming there is one). I've looked around SO and googled a bit but can't see anything obvious so I could be missing something.

foreach (int i in is) {
   foreach (int j in js) {
      //Pair i and j
   }
}

This is a highly simplified version of what I do in my code. The two integers are lookup keys which are used to retrieve one/more objects and all the objects from the two lookups are paired together into new objects.

This small block of code in a much larger more complex system becomes a major performance bottleneck as the dataset it's operating over scales. Some of this could likely be mitigated by improving the data structures used to store the objects and the lookups involved but the main issue I feel is still the computation of the Cartesian product itself.

Edit

So some more background on my specific usage of the algorithm to see if there may be any tricks that I can use in response to Marc's comment. The overall system is a SPARQL query engine which processes SPARQL queries over sets of Graph data, SPARQL is a pattern based language so each query consists of a series of patterns which are matched against the Graph(s). In the case where two subsequen开发者_C百科t patterns have no common variables (they are disjoint) it is necessary to compute the Cartesian product of the solutions produced by the two patterns to get the set of possible solutions for the overall query. There may be any number of patterns and I may have to compute Cartesian products multiple times which can lead to a fairly exponential expansion in possible solutions if the query is composed of a series of disjoint patterns.

Somehow from the existing answers I doubt whether there any tricks that could apply

Update

So I thought I would post an update on what I implemented in order to minimise the need to do Cartesian products and thus optimise the query engine generally. Note that it's not always possible to completely eliminate the need for products but it's nearly always possible to optimise so the size of the two sets being joined is much smaller.

Since each BGP (Basic Graph Pattern) which is a set of Triple Patterns is executed as a block (in essence) the engine is free to reorder patterns within a BGP for optimal performance. For example consider the following BGP:

?a :someProperty ?b .
?c :anotherProperty ?d .
?b a :Class .

Executed as is the query requires a cartesian product since the results of the first pattern are disjoint from the second pattern so the results of the first two patterns is a cartesian product of their individual results. This result will contain far more results than we actually need since the third pattern restricts the possible results of the first pattern but we don't apply this restriction till afterwards. But if we reorder like so:

?b a :Class .
?a :someProperty ?b .
?c :anotherProperty ?d .

We'll still need a cartesian product to get the final results since the 2nd and 3rd patterns are still disjoint but by reordering we restrict the size of the results of the 2nd pattern meaning the size of our cartesian product will be much smaller.

There are some various other optimisations we make but I'm not going to post them here as it starts to get into fairly detailed discussion of SPARQL engine internals. If anyone is interested in further details just leave a comment or send me a tweet @RobVesse

The complexity of cartesian product is O(n²), there is no shortcut.

In specific cases, the order you iterate your two axis is important. For example, if your code is visiting every slot in an array — or every pixel in an in image — then you should try to visit the slots in natural order. An image is typically laid out in ‘scanlines’, so the pixels on any Y are adjacent. Therefore, you should iterate over the Y on the outer loop and the X on the inner.

Whether you need the cartesian product or wherther is a more efficient algorithm depends on the problem you are solving.

You can't really change the performance of a nested loop without some additional knowledge, but that would be use-specific. If you have got n items in is and m items in js, it is always going to be O(n*m).

You can change the look of it though:

var qry = from i in is
          from j in js
          select /*something involving i/j */;

This is still O(n*m), but has nominal extra overhead of LINQ (you won't notice it in normal usage, though).

What are you doing in your case? There may be tricks...

One thing to definitely avoid is anything that forces a cross-join to buffer - the foreach approach is fine and doesn't buffer - but if you add each item to a List<>, then beware the memory implications. Ditto OrderBy etc (if used inappropriately).

I can't propose anything better, than O(n^2), because that's the size of the output, and the algorithm hence can't be any faster.

What I can suggest is using another approach to whether you need to compute product. For example, you may not even need the product set P if only you are going to query whether certain elements belong to it. You only need the information about the sets it's composed of.

Indeed (pseudocode)

bool IsInSet(pair (x,y), CartesianProductSet P)
{
   return IsInHash(x,P.set[1]) && IsInHash(y,P.set[2])
}

where Cartesian product is calculated like this:

// Cartesian product of A and B is
P.set[1]=A; P.set[2]=B;

If you implement sets as hashes, then lookup in a cartesian product of m sets is just a lookup in m hashes you get for free. Construction of the cartesian product and IsInSet lookup each take O(m) time, where m is a number of sets you multiply, and it's much less than n--size of each set.

Additional information has been added to the question.

The duplicates can be avoided if you record those you've already computed so as to avoid duplicating them again - it's assumed that the cost of such bookkeeping - a hashmap or a simple list - is less than the cost of computing a duplicate.

The C# runtime is really very fast, but for extreme heavy lifting, you might want to drop into native code.

You might also notice the essential parallelness of this problem. If the computation of a product does not impact the computation of any other product, you could straightforwardly use a multiple processor cores for doing the work in parallel. Look at ThreadPool.QueueUserWorkItem.

If cache locality (or local memory required to maintain the j's) is a problem, you can make your algorithm more cache-friendly by bisecting the input arrays recursively. Something like:

cartprod(is,istart,ilen, js,jstart,jlen) {
  if(ilen <= IMIN && jlen <= JMIN) { // base case
    for(int i in is) {
      for(int j in js) {
        // pair i and j
      }
    }
    return;
  }
  if(ilen > IMIN && jlen > JMIN) { // divide in 4
    ilen2= ilen>>1;
    jlen2= jlen>>1;
    cartprod(is,istart,ilen2,            js,jstart,jlen2);
    cartprod(is,istart+ilen2,ilen-ilen2, js,jstart,jlen2);
    cartprod(is,istart+ilen2,ilen-ilen2, js,jstart+jlen2,jlen-jlen2);
    cartprod(is,istart,ilen2,            js,jstart+jlen2,jlen-jlen2);
    return;
  }
  // handle other cases...
}

Note that this access pattern will automatically take fairly good advantage of all levels of automatic cache; this kind of technique is called cache-oblivious algorithm design.

I don't know how to write Java-like-Iterators in C#, but maybe you know and can transfer my solution from here to C# yourself.

It can be interesting if your combinations are too big to keep them completely in memory.

However, if you filter by attribute over the collection, you should filter before building the combination. Example:

If you have numbers from 1 to 1000 and random words and combine them, and then filter those combinations, where the number is divisible by 20 and the word starts with 'd', you could have 1000*(26*x)=26000*x combinations to search.

Or you filter the numbers first, which gives you 50 numbers, and (if equally distributed) 1 character, which are only 50*x elements in the end.