This is kind of a long shot, but I am hoping that someone has been in a similar situation as I am looking for some advice how to efficiently bring a set of large word2vec models into a production environment.
We have a range of trained w2v models with a dimensionality of 300. Due to the underlying data - huge corpus with POS tagged words; specialized vocabularies with up to 1 mio words - these models became quite large and we are currently looking into effective ways how to expose these to our users w/o paying a too high price in infrastructure.
Besides trying to better control the vocabulary size, obviously, dimensionality reduction on the feature vectors would be an option. Is anyone aware of publications around that, particularly on how this would affect model quality, and how to best measure this?
Another option is to pre-calculate the top X most similar words to each vocabulary word and to provide a lookup table. With the model size being that big, this is currently also very inefficient. Are there any heuristics known that could be used reduce the number of necessary distance calculations from n x n-1 to a lower number?
Thank you very much!
There are pre-indexing techniques for similarity-search in high-dimensional spaces which can speed nearest-neighbor discovery, but usually at a cost of absolute accuracy. (They also need more memory for the index.)
An example is the ANNOY library. The gensim project includes a demo notebook showing its use with Word2Vec.
I once did some experiments using just 16-bit (rather than 32-bit) floats in a Word2Vec model. It saved memory in the idle state, and nearest-neighbor top-N results were nearly unchanged. But, perhaps because some behind-the-scenes up-conversion to 32-bit floats was still occurring during the one-against-all distance-calculations, speed of operations was actually reduced. (And this suggests that each distance-calculation may have caused a temporary memory expansion offsetting any idle-state savings.) So it's not a quick fix, but further research here – perhaps involving finding/implementing the right routines for float16 array operations – could maybe mean 50% model-size savings and equivalent or even better speed.
For many applications, discarding the least-frequent words doesn't hurt much – or even, when done before training, can improve the quality of the remaining vectors. As many implementations, including gensim, sort the word-vector array in most-to-least-frequent order, you can discard the tail-end of the array to save memory, or limit most_similar() searches to the first-N entries to speed calculations.
Once you've minimized the vocabulary size, you want to be sure the full set is in RAM, and no swapping is triggered during the (typical) full-sweep distance-calculations. If you need multiple processes to serve answers from the same vector set, as in a web service on a multicore machine, gensim's memory-mapping operations can prevent each process from loading its own redundant copy of the vectors. You can see a discussion of this technique in this answer about speeding gensim Word2Vec loading time.
Finally, while precomputing top-N neighbors for a larger vocabulary is both time-consuming and memory-intensive, if your pattern of access is such that some tokens are checked far more than others, a cache of the N most-recently or M most-frequently requested top-N could improve perceived performance a lot – making only less-frequently-requested neighbor-lists require the full distance calculations to every other token.
Related
I want to know if there is spatial partition data structure that is better suited for a placement system than a quadtree. By better suited I mean for the data structure to have a O(logn) time complexity or less when search querying it and using less memory. I want to know what data structure can organize my data in such a way that querying it is faster than a quadtree. Its all 2D and its all rectangles which should never overlap. I currently have a quadtree done and it works great and its fast, I am just curious to know if there is a data structure that uses less resources and its faster than a quadtree for this case.
The fastest is probably brute forcing it on a GPU.
Also, it is really worth trying out different implementations, I found performance differences between implementations to be absolutely wild.
Another tip: measure performance with realistic data (potentially multiple scenarios), data and usage characteristics can have enormous influence on index performance.
Some of these characteristics are (you already mentioned "rectangle data" and "2D"):
How big is your dataset
How much overlap do you have between rectangles?
Do you need to update data often?
Do you have a large variance between small and large rectangles?
Do you have dense cluster of rectangles?
How large is the are you cover?
Are your coordinates integers or floats?
Is it okay if the execution time of operations varies or should it be consistent?
Can you pre-load data? Do you need to update the index?
Quadtrees can be a good initial choice. However they have some problems, e.g.:
They can get very deep (and inefficient) with dense clusters
They don't work very well when there is a lot of overlap between rectangles
Update operations may take longer if nodes are merged or split.
Another popular choice are R-Trees (I found R-star-Trees to be the best). Some properties:
Balanced (good for predictable search time but bad because update times can be very unpredictable due to rebalancing)
Quite complex to implement.
R-Trees can also be preloaded (takes longer but allows queries to be faster), this is called STR-Tree (Sort-tile-recurse-R-Tree)
It may be worth looking at the PH-Tree (disclaimer: self advertisement):
Similar to a quadtree but depth is limited to the bit-width of the data (usually 32 or 64 (bits)).
No rebalancing. Merging or splitting is guaranteed to move only one entry (=cheap)
Prefers integer coordinates but works reasonably well with floating point data as well.
Implementations can be quite space efficient (they don't need to store all bit of coordinates). However, not all implementations support that. Also, the effect varies and is strongest with integer coordinates.
I made some measurements here. The measurements include a 2D dataset where I store line segments from OpenStreetMap as boxes, the relevant diagrams are labeled with "OSM-R" (R for rectangles).
Fig. 3a shows timings for inserting a given amount of data into a tree
Fig. 9a shows memory usage
Fig. 15a shows query times for queries that return on average 1000 entries
Fig. 17a shows how query performance changes when varying the query window size (on an index with 1M entries)
Fig. 41a shows average times for updating an index with 1M entries
PH/PHM is the PH-Tree, PHM has coordinates converted to integer before storing them
RSZ/RSS are two different R-Tree implementations
STR is an STR-Tree
Q(T)Z is a quadtree
In case you are using Java, have a look at my spatial index collection.
Similar collections exist for other programming languages.
I was reading this answer
Efficient (and well explained) implementation of a Quadtree for 2D collision detection
and encountered this paragraph
All right, so actually quadtrees are not my favorite data structure for this purpose. I tend to prefer a grid hierarchy, like a coarse grid for the world, a finer grid for a region, and an even finer grid for a sub-region (3 fixed levels of dense grids, and no trees involved), with row-based optimizations so that a row that has no entities in it will be deallocated and turned into a null pointer, and likewise completely empty regions or sub-regions turned into nulls. While this simple implementation of the quadtree running in one thread can handle 100k agents on my i7 at 60+ FPS, I've implemented grids that can handle a couple million agents bouncing off each other every frame on older hardware (an i3). Also I always liked how grids made it very easy to predict how much memory they'll require, since they don't subdivide cells. But I'll try to cover how to implement a reasonably efficient quadtree.
This type of grid seems intuitive, it sort of sounds like a "N-order" grid, where instead of 4 child nodes, you have N child nodes per parent. N^3 can go much further than 4^3, which allows better precision with potentially (I guess) less branching (since there are many less nodes to branch).
I'm a little puzzled because I would intuitively use a single, or maybe 3 std::map with the proper < operator(), to reduce its memory footprint, but I'm not sure it would be so fast, since querying an AABB would mean stacking several accesses that are O(log n).
What exactly are those row-based optimizations he is talking about? Is this type of grid search common?
I have some understanding of a z order curve, and I'm not entirely satisfied with a quadtree.
It's my own quote. But that's based on a common pattern I encountered in my personal experience. Also, terms like "parent" and "child" are ones I'd largely discard when talking about grids. You just got a big 2-dimensional or N-dimensional table/matrix storing stuff. There's not really a hierarchy involved whatsoever -- these data structures are more comparable to arrays than trees.
"Coarse" and "Fine"
On "coarse" and "fine", what I meant there is that "coarse" search queries tend to be cheaper but give more false positives. A coarser grid would be one that is lower in grid resolution (fewer, larger cells). Coarse searches may involve traversing/searching fewer and larger grid cells. For example, say we want to see if an element intersects a point/dot in a gigantic cell (imagine just a 1x1 grid storing everything in the simulation). If the dot intersects the cell, we may get a whole lot of elements returned in that cell but maybe only one or none of them actually intersect the dot.
So a "coarse" query is broad and simple but not very precise at narrowing down the list of candidates (or "suspects"). It may return too many results and still leave a whole lot of processing required left to do to narrow down what actually intersects the search parameter*.
It's like in those detective shows when they search a database for a
possible killer, putting in "white male" might not require much
processing to list the results but might give way too many results to
properly narrow down the suspects. "Fine" would be the opposite and might require more processing of the database but narrow down the result to just one suspect.
This is a crude and flawed analogy but I hope it helps.
Often the key to broadly optimizing spatial indices before we get into things like memory optimizations whether we're talking spatial hashes or quadtrees is to find a nice balance between "coarse" and "fine". Too "fine" and we might spend too much time traversing the data structure (searching many small cells in a uniform grid, or spending too much time in tree traversal for adaptive grids like quadtrees). Too "coarse" and the spatial index might give back too many results to significantly reduce the amount of time required for further processing. For spatial hashes (a data structure I don't personally like very much but they're very popular in gamedev), there's often a lot of thought and experimentation and measuring involved in determining an optimal cell size to use.
With uniform NxM grids, how "coarse" or "fine" they are (big or small cells and high or low grid resolution) not only impacts search times for a query but can also impact insertion and removal times if the elements are larger than a point. If the grid is too fine, a single large or medium-sized element may have to be inserted into many tiny cells and removed from many tiny cells, using lots of extra memory and processing. If it's too coarse, the element may only have to be inserted and removed to/from one large cell but at costs to the data structure's ability to narrow down the number of candidates returned from a search query to a minimum. Without care, going too "fine/granular" can become very bottlenecky in practice and a developer might find his grid structure using gigabytes of RAM for a modest input size. With tree variants like quadtrees, a similar thing can happen if the maximum allowed tree depth is too high a value causing explosive memory use and processing when the leaf nodes of the quadtree store the tiniest cell sizes (we can even start running into floating-point precision bugs that wreck performance if the cells are allowed to be subdivided to too small a size in the tree).
The essence of accelerating performance with spatial indices is often this sort of balancing act. For example, we typically don't want to apply frustum culling to individual polygons being rendered in computer graphics because that's typically not only redundant with what the hardware already does at the clipping stage, but it's also too "fine/granular" and requires too much processing on its own compared to the time required to just request to render one or more polygons. But we might net huge performance improvements with something a bit "coarser", like applying frustum culling to an entire creature or space ship (an entire model/mesh), allowing us to avoid requesting to render many polygons at once with a single test. So I often use the terms, "coarse" and "fine/granular" frequently in these sorts of discussions (until I find better terminology that more people can easily understand).
Uniform vs. Adaptive Grid
You can think of a quadtree as an "adaptive" grid with adaptive grid cell sizes arranged hierarchically (working from "coarse' to "fine" as we drill down from root to leaf in a single smart and adaptive data structure) as opposed to a simple NxM "uniform" grid.
The adaptive nature of the tree-based structures is very smart and can handle a broad range of use cases (although typically requiring some fiddling of maximum tree depth and/or minimum cell size allowed and possibly how many maximum elements are stored in a cell/node before it subdivides). However, it can be more difficult to optimize tree data structures because the hierarchical nature doesn't lend itself so easily to the kind of contiguous memory layout that our hardware and memory hierarchy is so well-suited to traverse. So very often I find data structures that don't involve trees to be easier to optimize in the same sense that optimizing a hash table might be simpler than optimizing a red-black tree, especially when we can anticipate a lot about the type of data we're going to be storing in advance.
Another reason I tend to favor simpler, more contiguous data structures in lots of contexts is that the performance requirements of a realtime simulation often want not just fast frame rates, but consistent and predictable frame rates. The consistency is important because even if a video game has very high frame rates for most of the game but some part of the game causes the frame rates to drop substantially for even a brief period of time, the player may die and game over as a result of it. It was often very important in my case to avoid these types of scenarios and have data structures largely absent pathological worst-case performance scenarios. In general, I find it easier to predict the performance characteristics of lots of simpler data structures that don't involve an adaptive hierarchy and are kind of on the "dumber" side. Very often I find the consistency and predictability of frame rates to be roughly tied to how easily I can predict the data structure's overall memory usage and how stable that is. If the memory usage is wildly unpredictable and sporadic, I often (not always, but often) find the frame rates will likewise be sporadic.
So I often find better results using grids personally, but if it's tricky to determine a single optimal cell size to use for the grid in a particular simulation context, I just use multiple instances of them: one instance with larger cell sizes for "coarse" searches (say 10x10), one with smaller ones for "finer" searches (say 100x100), and maybe even one with even smaller cells for the "finest" searches (say 1000x1000). If no results are returned in the coarse search, then I don't proceed with the finer searches. I get some balance of the benefits of quadtrees and grids this way.
What I did when I used these types of representations in the past is not to store a single element in all three grid instances. That would triple the memory use of an element entry/node into these structures. Instead, what I did was insert the indices of the occupied cells of the finer grids into the coarser grids, as there are typically far fewer occupied cells than there are a total number of elements in the simulation. Only the finest, highest-resolution grid with the smallest cell sizes would store the element. The cells in the finest grid are analogous to the leaf nodes of a quadtree.
The "loose-tight double grid" as I'm calling it in one of the answers to that question is an expansion on this multi-grid idea. The difference is that the finer grid is actually loose and has cell sizes that expand and shrink based on the elements inserted to it, always guaranteeing that a single element, no matter how large or small, needs only be inserted to one cell in the grid. The coarser grid stores the occupied cells of the finer grid leading to two constant-time queries (one in the coarser tight grid, another into the finer loose grid) to return an element list of potential candidates matching the search parameter. It also has the most stable and predictable memory use (not necessarily the minimal memory use because the fine/loose cells require storing an axis-aligned bounding box that expands/shrinks which adds another 16 bytes or so to a cell) and corresponding stable frame rates because one element is always inserted to one cell and doesn't take any additional memory required to store it besides its own element data with the exception of when its insertion causes a loose cell to expand to the point where it has to be inserted to additional cells in the coarser grid (which should be a fairly rare-case scenario).
Multiple Grids For Other Purposes
I'm a little puzzled because I would intuitively use a single, or maybe 3 std::map with the proper operator(), to reduce its memory footprint, but I'm not sure it would be so fast, since querying an AABB would mean stacking several accesses that are O(log n).
I think that's an intuition many of us have and also probably a subconscious desire to just lean on one solution for everything because programming can get quite tedious and repetitive and it'd be ideal to just implement something once and reuse it for everything: a "one-size-fits-all" t-shirt. But a one-sized-fits-all shirt can be poorly tailored to fit our very broad and muscular programmer bodies*. So sometimes it helps to use the analogy of a small, medium, and large size.
This is a very possibly poor attempt at humor on my part to make my long-winded texts less boring to read.
For example, if you are using std::map for something like a spatial hash, then there can be a lot of thought and fiddling around trying to find an optimal cell size. In a video game, one might compromise with something like making the cell size relative to the size of your average human in the game (perhaps a bit smaller or bigger), since lots of the models/sprites in the game might be designed for human use. But it might get very fiddly and be very sub-optimal for teeny things and very sub-optimal for gigantic things. In that case, we might do well to resist our intuitions and desires to just use one solution and use multiple (it could still be the same code but just multiple instances of the same class instance for the data structure constructed with varying parameters).
As for the overhead of searching multiple data structures instead of a single one, that's something best measured and it's worth remembering that the input sizes of each container will be smaller as a result, reducing the cost of each search and very possibly improve locality of reference. It might exceed the benefits in a hierarchical structure that requires logarithmic search times like std::map (or not, best to just measure and compare), but I tend to use more data structures which do this in constant-time (grids or hash tables). In my cases, I find the benefits far exceeding the additional cost of requiring multiple searches to do a single query, especially when the element sizes vary radically or I want some basic thing resembling a hierarchy with 2 or more NxM grids that range from "coarse" to "fine".
Row-Based Optimizations
As for "row-based optimizations", that's very specific to uniform fixed-sized grids and not trees. It refers to using a separate variable-sized list/container per row instead of a single one for the entire grid. Aside from potentially reducing memory use for empty rows that just turn into nulls without requiring an allocated memory block, it can save on lots of processing and improve memory access patterns.
If we store a single list for the entire grid, then we have to constantly insert and remove from that one shared list as elements move around, particles are born, etc. That could lead to more heap allocations/deallocations growing and shrinking the container but also increases the average memory stride to get from one element in that list to the next which will tend to translate to more cache misses as a result of more irrelevant data being loaded into a cache line. Also these days we have so many cores so having a single shared container for the entire grid may reduce the ability to process the grid in parallel (ex: searching one row while simultaneously inserting to another). It can also lead to more net memory use for the structure since if we use a contiguous sequence like std::vector or ArrayList, those can often store the memory capacity of as many as twice the elements required to reduce the time of insertions to amortized constant time by minimizing the need to reallocate and copy the former elements in linear-time by keeping excess capacity.
By associating a separate medium-sized container per grid row or per column instead of gigantic one for the entire grid, we can mitigate these costs in some cases*.
This is the type of thing you definitely measure before and after though to make sure it actually improves overall frame rates, and probably attempt in response to a first attempt storing a single list for the entire grid revealing many non-compulsory cache misses in the profiler.
This might beg the question of why we don't use a separate teeny list container for every single cell in the grid. It's a balancing act. If we store that many containers (ex: a million instances of std::vector for a 1000x1000 grid possibly storing very few or no elements each), it would allow maximum parallelism and minimize the stride to get from one element in a cell to the next one in the cell, but that can be quite explosive in memory use and introduce a lot of extra processing and heap overhead.
Especially in my case, my finest grids might store a million cells or more, but I only require 4 bytes per cell. A variable-sized sequence per cell would typically explode to at least something like 24 bytes or more (typically far more) per cell on 64-bit architectures to store the container data (typically a pointer and a couple of extra integers, or three pointers on top of the heap-allocated memory block), but on top of that, every single element inserted to an empty cell may require a heap allocation and compulsory cache miss and page fault and very frequently due to the lack of temporal locality. So I find the balance and sweet spot to be one list container per row typically among my best-measured implementations.
I use what I call a "singly-linked array list" to store elements in a grid row and allow constant-time insertions and removals while still allowing some degree of spatial locality with lots of elements being contiguous. It can be described like this:
struct GridRow
{
struct Node
{
// Element data
...
// Stores the index into the 'nodes' array of the next element
// in the cell or the next removed element. -1 indicates end of
// list.
int next = -1;
};
// Stores all the nodes in the row.
std::vector<Node> nodes;
// Stores the index of the first removed element or -1
// if there are no removed elements.
int free_element = -1;
};
This combines some of the benefits of a linked list using a free list allocator but without the need to manage separate allocator and data structure implementations which I find to be too generic and unwieldy for my purposes. Furthermore, doing it this way allows us to halve the size of a pointer down to a 32-bit array index on 64-bit architectures which I find to be a big measured win in my use cases when the alignment requirements of the element data combined with the 32-bit index don't require an additional 32-bits of padding for the class/struct which is frequently the case for me since I often use 32-bit or smaller integers and 32-bit single-precision floating-point or 16-bit half-floats.
Unorthodox?
On this question:
Is this type of grid search common?
I am not sure! I tend to struggle a bit with terminology and I'll have to ask people's forgiveness and patience in communication. I started programming from early childhood in the 1980s before the internet was widespread, so I came to rely on inventing a lot of my own techniques and using my own crude terminology as a result. I got my degree in computer science about a decade and a half later when I reached my 20s and corrected some of my terminology and misconceptions but I've had many years just rolling my own solutions. So I am often not sure if other people have come across some of the same solutions or not, and if there are formal names and terms for them or not.
That makes communication with other programmers difficult and very frustrating for both of us at times and I have to ask for a lot of patience to explain what I have in mind. I've made it a habit in meetings to always start off showing something with very promising results which tends to make people more patient with my crude and long-winded explanations of how they work. They tend to give my ideas much more of a chance if I start off by showing results, but I'm often very frustrated with the orthodoxy and dogmatism that can be prevalent in this industry that can sometimes prioritize concepts far more than execution and actual results. I'm a pragmatist at heart so I don't think in terms of "what is the best data structure?" I think in terms of what we can effectively implement personally given our strengths and weaknesses and what is intuitive and counter-intuitive to us and I'm willing to endlessly compromise on the purity of concepts in favor of a simpler and less problematic execution. I just like good, reliable solutions that roll naturally off our fingertips no matter how orthodox or unorthodox they may be, but a lot of my methods may be unorthodox as a result (or not and I might just have yet to find people who have done the same things). I've found this site useful at rare times in finding peers who are like, "Oh, I've done that too! I found the best results if we do this [...]" or someone pointing out like, "What you are proposing is called [...]."
In performance-critical contexts, I kind of let the profiler come up with the data structure for me, crudely speaking. That is to say, I'll come up with some quick first draft (typically very orthodox) and measure it with the profiler and let the profiler's results give me ideas for a second draft until I converge to something both simple and performant and appropriately scalable for the requirements (which may become pretty unorthodox along the way). I'm very happy to abandon lots of ideas since I figure we have to weed through a lot of bad ideas in a process of elimination to come up with a good one, so I tend to cycle through lots of implementations and ideas and have come to become a really rapid prototyper (I have a psychological tendency to stubbornly fall in love with solutions I spent lots of time on, so to counter that I've learned to spend the absolute minimal time on a solution until it's very, very promising).
You can see my exact methodology at work in the very answers to that
question where I iteratively converged through lots of profiling and
measuring over the course of a few days and prototyping from a fairly orthodox quadtree to that
unorthodox "loose-tight double grid" solution that handled the largest
number of agents at the most stable frame rates and was, for me
anyway, much faster and simpler to implement than all the structures
before it. I had to go through lots of orthodox solutions and measure them though to generate the final idea for the unusual loose-tight variant. I always start off with and favor the orthodox solutions and start off inside the box because they're well-documented and understood and just very gently and timidly venture outside, but I do often find myself a bit outside the box when the requirements are steep enough. I'm no stranger to the steepest requirements since in my industry and as a fairly low-level type working on engines, the ability to handle more data at good frame rates often translates not only to greater interactivity for the user but also allows artists to create more detailed content of higher visual quality than ever before. We're always chasing higher and higher visual quality at good frame rates, and that often boils down to a combination of both performance and getting away with crude approximations whenever possible. This inevitably leads to some degree of unorthodoxy with lots of in-house solutions very unique to a particular engine, and each engine tends to have its own very unique strengths and weaknesses as you find comparing something like CryEngine to Unreal Engine to Frostbite to Unity.
For example, I have this data structure I've been using since childhood and I don't know the name of it. It's a straightforward concept and it's just a hierarchical bitset that allows set intersections of potentially millions of elements to be found in as little as a few iterations of simple work as well as traverse millions of elements occupying the set with just a few iterations (less than linear-time requirements to traverse everything in the set just through the data structure itself which returns ranges of occupied elements/set bits instead of individual elements/bit indices). But I have no idea what the name is since it's just something I rolled and I've never encountered anyone talking about it in computer science. I tend to refer to it as a "hierarchical bitset". Originally I called it a "sparse bitset tree" but that seems a tad verbose. It's not a particularly clever concept at all and I wouldn't be surprised or disappointed (actually quite happy) to find someone else discovering the same solution before me but just one I don't find people using or talking about ever. It just expands on the strengths of a regular, flat bitset in rapidly finding set intersections with bitwise OR and rapidly traverse all set bits using FFZ/FFS but reducing the linear-time requirements of both down to logarithmic (with the logarithm base being a number much larger than 2).
Anyway, I wouldn't be surprised if some of these solutions are very unorthodox, but also wouldn't be surprised if they are reasonably orthodox and I've just failed to find the proper name and terminology for these techniques. A lot of the appeal of sites like this for me is a lonely search for someone else who has used similar techniques and to try to find proper names and terms for them often to end in frustration. I'm also hoping to improve on my ability to explain them although I've always been so bad and long-winded here. I find using pictures helps me a lot because I find human language to be incredibly riddled with ambiguities. I'm also fond of deliberately imprecise figurative language which embraces and celebrates the ambiguities such as metaphor and analogy and humorous hyperbole, but I've not found it's the type of thing programmers tend to appreciate so much due to its lack of precision. But I've never found precision to be that important so long as we can convey the meaty stuff and what is "cool" about an idea while they can draw their own interpretations of the rest. Apologies for the whole explanation but hopefully that clears some things up about my crude terminology and the overall methodology I use to arrive at these techniques. English is also not my first language so that adds another layer of convolution where I have to sort of translate my thoughts into English words and struggle a lot with that.
I am trying to find a fast algorithm for finding the (approximate, if need be) nearest neighbours of a given point in a two-dimensional space where points are frequently removed from the dataset and new points are added.
(Relatedly, there are two variants of this problem that interest me: one in which points can be thought of as being added and removed randomly and another in which all the points are in constant motion.)
Some thoughts:
kd-trees offer good performance, but are only suitable for static point sets
R*-trees seem to offer good performance for a variety of dimensions, but the generality of their design (arbitrary dimensions, general content geometries) suggests the possibility that a more specific algorithm might offer performance advantages
Algorithms with existing implementations are preferable (though this is not necessary)
What's a good choice here?
I agree with (almost) everything that #gsamaras said, just to add a few things:
In my experience (using large dataset with >= 500,000 points), kNN-performance of KD-Trees is worse than pretty much any other spatial index by a factor of 10 to 100. I tested them (2 KD-trees and various other indexes) on a large OpenStreetMap dataset. In the following diagram, the KD-Trees are called KDL and KDS, the 2D dataset is called OSM-P (left diagram): The diagram is taken from this document, see bullet points below for more information.
This research describes an indexing method for moving objects, in case you keep (re-)inserting the same points in slightly different positions.
Quadtrees are not too bad either, they can be very fast in 2D, with excellent kNN performance for datasets < 1,000,000 entries.
If you are looking for Java implementations, have a look at my index library. In has implementations of quadtrees, R-star-tree, ph-tree, and others, all with a common API that also supports kNN. The library was written for the TinSpin, which is a framework for testing multidimensional indexes. Some results can be found enter link description here (it doesn't really describe the test data, but 'OSM-P' results are based on OpenStreetMap data with up to 50,000,000 2D points.
Depending on your scenario, you may also want to consider PH-Trees. They appear to be slower for kNN-queries than R-Trees in low dimensionality (though still faster than KD-Trees), but they are faster for removal and updates than RTrees. If you have a lot of removal/insertion, this may be a better choice (see the TinSpin results, Figures 2 and 46). C++ versions are available here and here.
Check the Bkd-Tree, which is:
an I/O-efficient dynamic data structure based on the kd-tree. [..] the Bkd-tree maintains its high space utilization and excellent
query and update performance regardless of the number of updates performed on it.
However this data structure is multi dimensional, and not specialized to lower dimensions (like the kd-tree).
Play with it in bkdtree.
Dynamic Quadtrees can also be a candidate, with O(logn) query time and O(Q(n)) insertion/deletion time, where Q(n) is the time
to perform a query in the data structure used. Note that this data structure is specialized for 2D. For 3D however, we have octrees, and in a similar way the structure can be generalized for higher dimensions.
An implentation is QuadTree.
R*-tree is another choice, but I agree with you on the generality. A r-star-tree implementations exists too.
A Cover tree could be considered as well, but I am not sure if it fits your description. Read more here,and check the implementation on CoverTree.
Kd-tree should still be considered, since it's performance is remarkable on 2 dimensions, and its insertion complexity is logarithic in size.
nanoflann and CGAL are jsut two implementations of it, where the first requires no install and the second does, but may be more performant.
In any case, I would try more than one approach and benchmark (since all of them have implementations and these data structures are usually affected by the nature of your data).
I am working with sample data set to learn clustering. This data set contains number of occurrences for the keywords.
Since all are number of occurrences for the different keywords, will it be OK not to scale the values and use them as it is?
I read couple of articles on internet where its emphasized that scaling is important as it will adjust the relativity of the frequency. Since most of frequencies are 0 (95%+), z score scaling will change the shape of distribution, which I am feeling could be problem as I am changing the nature of data.
I am thinking of not changing values at all to avoid this. Will that affect the quality of results I get from the clustering?
As it was already noted, the answer heavily depends on an algorithm being used.
If you're using distance-based algorithms with (usually default) Euclidean distance (for example, k-Means or k-NN), it'll rely more on features with bigger range just because a "typical difference" of values of that feature is bigger.
Non-distance based models can be affected, too. Though one might think that linear models do not get into this category since scaling (and translating, if needed) is a linear transformation, so if it makes results better, then the model should learn it, right? Turns out, the answer is no. The reason is that no one uses vanilla linear models, they're always used with with some sort of a regularization which penalizes too big weights. This can prevent your linear model from learning scaling from data.
There are models that are independent of the feature scale. For example, tree-based algorithms (decision trees and random forests) are not affected. A node of a tree partitions your data into 2 sets by comparing a feature (which splits dataset best) to a threshold value. There's no regularization for the threshold (because one should keep height of the tree small), so it's not affected by different scales.
That being said, it's usually advised to standardize (subtract mean and divide by standard deviation) your data.
Probably it depends on the classification algorithm. I'm only familiar with SVM. Please see Ch. 2.2 for the explanation of scaling
The type of feature (count of words) doesn't matter. The feature ranges should be more or less similar. If the count of e.g. "dignity" is 10 and the count of "have" is 100000000 in your texts, then (at least on SVM) the results of such features would be less accurate as when you scaled both counts to similar range.
The cases, where no scaling is needed are those, where the data is scaled implicitly e.g. features are pixel-values in an image. The data is scaled already to the range 0-255.
*Distance based algorithm need scaling
*There is no need of scaling in tree based algorithms
But it is good to scale your data and train model ,if possible compare the model accuracy and other evaluations before scaling and after scaling and use the best possibility
These is as per my knowledge
I have about 44 Million training examples across about 6200 categories.
After training, the model comes out to be ~ 450MB
And while testing, with 5 parallel mappers (each given enough RAM), the classification proceeds at a rate of ~ 4 items a second which is WAY too slow.
How can speed things up?
One way i can think of is to reduce the word corpus, but i fear losing accuracy. I had maxDFPercent set to 80.
Another way i thought of was to run the items through a clustering algorithm and empirically maximize the number of clusters while keeping the items within each category restricted to a single cluster. This would allow me to build separate models for each cluster and thereby (possibly) decrease training and testing time.
Any other thoughts?
Edit :
After some of the answers given below, i started contemplating doing some form of down-sampling by running a clustering algorithm, identifying groups of items that are "highly" close to one another and then taking a union of a few samples from those "highly" close groups and other samples that are not that tightly close to one another.
I also started thinking about using some form of data normalization techniques that involve incorporating edit distances while using n-grams (http://lucene.apache.org/core/4_1_0/suggest/org/apache/lucene/search/spell/NGramDistance.html)
I'm also considering using the hadoop streaming api to leverage some of the ML libraries available in Python from listed here http://pydata.org/downloads/ , and here http://scikit-learn.org/stable/modules/svm.html#svm (These I think use liblinear mentioned in one of the answers below)
Prune stopwords and otherwise useless words (too low support etc.) as early as possible.
Depending on how you use clustering, it may actually make in particular the test phase even more expensive.
Try other tools than Mahout. I found Mahout to be really slow in comparison. It seems that it somewhere comes at a really high overhead.
Using less training exampes would be an option. You will see that after a specific amount of training examples you classification accuracy on unseen examples won't increase. I would recommend to try to train with 100, 500, 1000, 5000, ... examples per category and using 20% for cross validating the accuracy. When it doesn't increase anymore, you have found the amount of data you need which may be a lot less then you use now.
Another approach would be to use another library. For document-classification i find liblinear very very very fast. It's may be more low-level then mahout.
"but i fear losing accuracy" Have you actually tried using less features or less documents? You may not lose as much accuracy as you fear. There may be a few things at play here:
Such a high number of documents are not likely to be from the same time period. Over time, the content of a stream will inevitably drift and words indicative of one class may become indicative of another. In a way, adding data from this year to a classifier trained on last year's data is just confusing it. You may get much better performance if you train on less data.
The majority of features are not helpful, as #Anony-Mousse said already. You might want to perform some form of feature selection before you train your classifier. This will also speed up training. I've had good results in the past with mutual information.
I've previously trained classifiers for a data set of similar scale and found the system worked best with only 200k features, and using any more than 10% of the data for training did not improve accuracy at all.
PS Could you tell us a bit more about your problem and data set?
Edit after question was updated:
Clustering is a good way of selecting representative documents, but it will take a long time. You will also have to re-run it periodically as new data come in.
I don't think edit distance is the way to go. Typical algorithms are quadratic in the length of the input strings, and you might have to run for each pair of words in the corpus. That's a long time!
I would again suggest that you give random sampling a shot. You say you are concerned about accuracy, but are using Naive Bayes. If you wanted the best model money can buy, you would go for a non-linear SVM, and you probably wouldn't live to see it finish training. People resort to classifiers with known issues (there's a reason Naive Bayes is called Naive) because they are much faster than the alternative but performance will often be just a tiny bit worse. Let me give you an example from my experience:
RBF SVM- 85% F1 score - training time ~ month
Linear SVM- 83% F1 score - training time ~ day
Naive Bayes- 82% F1 score - training time ~ day
You find the same thing in the literature: paper . Out of curiosity, what kind of accuracy are you getting?