Curse of dimensionality: (Or the unintended consequences when a human mind faces real complexity)

begin quote from:

https://en.wikipedia.org/wiki/Curse_of_dimensionality

Curse of dimensionality

From Wikipedia, the free encyclopedia

Jump to navigationJump to search

The curse of dimensionality refers to various phenomena that arise when analyzing and organizing data in high-dimensional spaces that do not occur in low-dimensional settings such as the three-dimensional physical space of everyday experience. The expression was coined by Richard E. Bellman when considering problems in dynamic programming.^[1][2]

Dimensionally cursed phenomena occur in domains such as numerical analysis, sampling, combinatorics, machine learning, data mining and databases. The common theme of these problems is that when the dimensionality increases, the volume of the space increases so fast that the available data become sparse. This sparsity is problematic for any method that requires statistical significance. In order to obtain a statistically sound and reliable result, the amount of data needed to support the result often grows exponentially with the dimensionality. Also, organizing and searching data often relies on detecting areas where objects form groups with similar properties; in high dimensional data, however, all objects appear to be sparse and dissimilar in many ways, which prevents common data organization strategies from being efficient.

Contents

• 1Domains
  • 1.1Combinatorics
  • 1.2Sampling
  • 1.3Optimization
  • 1.4Machine learning
  • 1.5Distance functions
  • 1.6Nearest neighbor search
  • 1.7Anomaly detection
  • 1.8Blessing of dimensionality
• 2See also
• 3References

Domains[edit]

Combinatorics[edit]

In some problems, each variable can take one of several discrete values, or the range of possible values is divided to give a finite number of possibilities. Taking the variables together, a huge number of combinations of values must be considered. This effect is also known as the combinatorial explosion. Even in the simplest case of ${\displaystyle d}$ binary variables, the number of possible combinations already is ${\displaystyle 2^{d}}$, exponential in the dimensionality. Naively, each additional dimension doubles the effort needed to try all combinations.

Sampling[edit]

There is an exponential increase in volume associated with adding extra dimensions to a mathematical space. For example, 10²=100 evenly spaced sample points suffice to sample a unit interval (a "1-dimensional cube") with no more than 10⁻²=0.01 distance between points; an equivalent sampling of a 10-dimensional unit hypercube with a lattice that has a spacing of 10⁻²=0.01 between adjacent points would require 10²⁰=[(10²)¹⁰] sample points. In general, with a spacing distance of 10⁻ⁿ the 10-dimensional hypercube appears to be a factor of 10^n(10-1)=[(10ⁿ)¹⁰/(10ⁿ)] "larger" than the 1-dimensional hypercube, which is the unit interval. In the above example n=2: when using a sampling distance of 0.01 the 10-dimensional hypercube appears to be 10¹⁸ "larger" than the unit interval. This effect is a combination of the combinatorics problems above and the distance function problems explained below.

Optimization[edit]

When solving dynamic optimization problems by numerical backward induction, the objective function must be computed for each combination of values. This is a significant obstacle when the dimension of the "state variable" is large.^[3]

Machine learning[edit]

In machine learning problems that involve learning a "state-of-nature" from a finite number of data samples in a high-dimensional feature space with each feature having a range of possible values, typically an enormous amount of training data is required to ensure that there are several samples with each combination of values. A typical rule of thumb is that there should be at least 5 training examples for each dimension in the representation.^[4] With a fixed number of training samples, the predictive power of a classifier or regressor first increases as number of dimensions or features used is increased but then decreases,^[5] which is known as Hughes phenomenon^[6] or peaking phenomena.^[4]

Distance functions[edit]

When a measure such as a Euclidean distance is defined using many coordinates, there is little difference in the distances between different pairs of samples.

One way to illustrate the "vastness" of high-dimensional Euclidean space is to compare the proportion of an inscribed hypersphere with radius ${\displaystyle r}$ and dimension ${\displaystyle d}$, to that of a hypercube with edges of length ${\displaystyle 2r.}$ The volume of such a sphere is ${\displaystyle {\frac {2r^{d}\pi ^{d/2}}{d\;\Gamma (d/2)}}}$, where ${\displaystyle \Gamma }$ is the gamma function, while the volume of the cube is ${\displaystyle (2r)^{d}}$. As the dimension ${\displaystyle d}$ of the space increases, the hypersphere becomes an insignificant volume relative to that of the hypercube. This can clearly be seen by comparing the proportions as the dimension ${\displaystyle d}$ goes to infinity:

${\displaystyle {\frac {V_{\mathrm {hypersphere} }}{V_{\mathrm {hypercube} }}}={\frac {\pi ^{d/2}}{d2^{d-1}\Gamma (d/2)}}\rightarrow 0}$ as ${\displaystyle d\rightarrow \infty }$.

Furthermore, the distance between the center and the corners is ${\displaystyle r{\sqrt {d}}}$, which increases without bound for fixed r. In this sense, nearly all of the high-dimensional space is "far away" from the centre. To put it another way, the high-dimensional unit hypercube can be said to consist almost entirely of the "corners" of the hypercube, with almost no "middle".

This also helps to understand the chi-squared distribution. Indeed, the (non-central) chi-squared distribution associated to a random point in the interval [-1, 1] is the same as the distribution of the length-squared of a random point in the d-cube. By the law of large numbers, this distribution concentrates itself in a narrow band around d times the standard deviation squared (σ²) of the original derivation. This illuminates the chi-squared distribution and also illustrates that most of the volume of the d-cube concentrates near the surface of a sphere of radius ${\displaystyle \sigma {\sqrt {d}}}$.

A further development of this phenomenon is as follows. Any fixed distribution on ℝ induces a product distribution on points in ℝ^d. For any fixed n, it turns out that the minimum and the maximum distance between a random reference point Q and a list of n random data points P₁,...,P_n become indiscernible compared to the minimum distance:^[7]

${\displaystyle \lim _{d\to \infty }E\left({\frac {\operatorname {dist} _{\max }(d)-\operatorname {dist} _{\min }(d)}{\operatorname {dist} _{\min }(d)}}\right)\to 0}$.

This is often cited as distance functions losing their usefulness (for the nearest-neighbor criterion in feature-comparison algorithms, for example) in high dimensions. However, recent research has shown this to only hold in the artificial scenario when the one-dimensional distributions ℝ are independent and identically distributed.^[8] When attributes are correlated, data can become easier and provide higher distance contrast and the signal-to-noise ratio was found to play an important role, thus feature selection should be used.^[8]

Nearest neighbor search[edit]

The effect complicates nearest neighbor search in high dimensional space. It is not possible to quickly reject candidates by using the difference in one coordinate as a lower bound for a distance based on all the dimensions.^[9][10]

However, it has recently been observed that the mere number of dimensions does not necessarily result in difficulties,^[11] since relevant additional dimensions can also increase the contrast. In addition, for the resulting ranking it remains useful to discern close and far neighbors. Irrelevant ("noise") dimensions, however, reduce the contrast in the manner described above. In time series analysis, where the data are inherently high-dimensional, distance functions also work reliably as long as the signal-to-noise ratio is high enough.^[12]

k-nearest neighbor classification[edit]

Another effect of high dimensionality on distance functions concerns k-nearest neighbor (k-NN) graphs constructed from a data set using a distance function. As the dimension increases, the indegree distribution of the k-NN digraph becomes skewed with a peak on the right because of the emergence of a disproportionate number of hubs, that is, data-points that appear in many more k-NN lists of other data-points than the average. This phenomenon can have a considerable impact on various techniques for classification (including the k-NN classifier), semi-supervised learning, and clustering,^[13] and it also affects information retrieval.^[14]

Anomaly detection[edit]

In a 2012 survey, Zimek et al. identified the following problems when searching for anomalies in high-dimensional data:^[8]

1. Concentration of scores and distances: derived values such as distances become numerically similar
2. Irrelevant attributes: in high dimensional data, a significant number of attributes may be irrelevant
3. Definition of reference sets: for local methods, reference sets are often nearest-neighbor based
4. Incomparable scores for different dimensionalities: different subspaces produce incomparable scores
5. Interpretability of scores: the scores often no longer convey a semantic meaning
6. Exponential search space: the search space can no longer be systematically scanned
7. Data snooping bias: given the large search space, for every desired significance a hypothesis can be found
8. Hubness: certain objects occur more frequently in neighbor lists than others.

Many of the analyzed specialized methods tackle one or another of these problems, but there remain many open research questions.

Blessing of dimensionality[edit]

Surprisingly and despite the expected "curse of dimensionality" difficulties, common-sense heuristics based on the most straightforward methods "can yield results which are almost surely optimal" for high-dimensional problems.^[15] The term "blessing of dimensionality" was introduced in the late 1990s.^[15] Donoho in his "Millennium manifesto" clearly explained why the "blessing of dimensionality" will form a basis of future data mining.^[16] The effects of the blessing of dimensionality were discovered in many applications and found their foundation in the concentration of measure phenomena.^[17] One example of the blessing of dimensionality phenomenon is linear separability of a random point from a large finite random set with high probability even if this set is exponentially large: the number of elements in this random set can grow exponentially with dimension. Moreover, this linear functional can be selected in the form of the simplest linear Fisher discriminant. This separability theorem was proven for a wide class of probability distributions: general uniformly log-concave distributions, product distributions in a cube and many other families (reviewed recently in ^[17]).

"The blessing of dimensionality and the curse of dimensionality are two sides of the same coin."^[18] For example, the typical property of essentially high-dimensional probability distributions in a high-dimensional space is: the squared distance of random points to a selected point is, with high probability, close to the average (or median) squared distance. This property significantly simplifies the expected geometry of data and indexing of high-dimensional data (blessing),^[19] but, at the same time, it makes the similarity search in high dimensions difficult and even useless (curse).^[20]

Zimek et al.^[8] noted that while the typical formalizations of the curse of dimensionality affect i.i.d. data, having data that is separated in each attribute becomes easier even in high dimensions, and argued that the signal-to-noise ratio matters: data becomes easier with each attribute that adds signal, and harder with attributes that only add noise (irrelevant error) to the data. In particular for unsupervised data analysis this effect is known as swamping.

See also[edit]

• Bellman equation
• Clustering high-dimensional data
• Concentration of measure
• Dimension reduction
• Model Order Reduction
• Dynamic programming
• Fourier-related transforms
• Linear least squares
• Multilinear PCA
• Multilinear subspace learning
• Principal component analysis
• Singular value decomposition

References