-
Notifications
You must be signed in to change notification settings - Fork 346
/
MomentsGroup.scala
408 lines (338 loc) · 13.6 KB
/
MomentsGroup.scala
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
/*
Copyright 2012 Twitter, Inc.
Licensed under the Apache License, Version 2.0 (the "License");
you may not use this file except in compliance with the License.
You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software
distributed under the License is distributed on an "AS IS" BASIS,
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
See the License for the specific language governing permissions and
limitations under the License.
*/
package com.twitter.algebird
import algebra.{CommutativeGroup, CommutativeMonoid}
/**
* A class to calculate the first five central moments over a sequence of Doubles. Given the first five
* central moments, we can then calculate metrics like skewness and kurtosis.
*
* m{i} denotes the ith central moment.
*
* This code manually inlines code to make it look like a case class. This is done because we changed the
* count from a Long to a Double to enable the scale method, which allows exponential decays of moments, but
* we didn't want to break backwards binary compatibility.
*/
sealed class Moments(val m0D: Double, val m1: Double, val m2: Double, val m3: Double, val m4: Double)
extends Product
with Serializable {
def this(m0: Long, m1: Double, m2: Double, m3: Double, m4: Double) =
this(m0.toDouble, m1, m2, m3, m4)
def m0: Long = m0D.toLong
def count: Long = m0
def totalWeight: Double = m0D
def mean: Double = m1
// Population variance, not sample variance.
def variance: Double =
m2 / m0D
// Population standard deviation, not sample standard deviation.
def stddev: Double = math.sqrt(variance)
def skewness: Double =
m3 / (m2 * stddev)
def kurtosis: Double =
m0D * m4 / (m2 * m2) - 3
/**
* Combines this instance with another [[Moments]] instance.
* @param b
* the other instance
* @return
* a [[Moments]] instances representing the combined moments of this instance and `b`
*/
def +(b: Moments): Moments = Moments.momentsMonoid.plus(this, b)
/**
* Returns a new [[Moments]] instance generated by merging in the new observation `b`.
* @param b
* a new observation
* @return
* a [[Moments]] instance representing the combined moments of this instance and `b`.
*/
def +(b: Double): Moments = {
val n = m0D + 1
val delta = b - mean
val delta_n = delta / n
val delta_n2 = delta_n * delta_n
val term1 = delta * delta_n * m0D
val meanCombined = Moments.getCombinedMeanDouble(m0D, mean, 1.0, b)
val m2combined = m2 + term1
val m3combined = m3 + term1 * delta_n * (n - 2) - 3 * delta_n * m2
val m4combined = m4 + term1 * delta_n2 * (n * n - 3 * n + 3) +
6 * delta_n2 * m2 - 4 * delta_n * m3
new Moments(n, meanCombined, m2combined, m3combined, m4combined)
}
/**
* Returns a [[Fold]] instance that uses `+` to accumulate deltas into this [[Moments]] instance.
*/
def fold: Fold[Double, Moments] =
Fold.foldMutable[Moments.MomentsState, Double, Moments](
{ case (state, x) =>
state += x
},
_ => Moments.MomentsState.fromMoments(this),
(state: Moments.MomentsState) => state.toMoments
)
override def productArity: Int = 5
override def productElement(idx: Int): Any =
idx match {
case 0 => count
case 1 => m1
case 2 => m2
case 3 => m3
case 4 => m4
}
override def canEqual(that: Any): Boolean =
that.isInstanceOf[Moments]
def copy(c0: Long = count, v1: Double = m1, v2: Double = m2, v3: Double = m3, v4: Double = m4): Moments = {
val v0 = if (c0 == count) m0D else c0.toDouble
new Moments(m0D = v0, m1 = v1, m2 = v2, m3 = v3, m4 = v4)
}
override def toString: String =
s"Moments($m0D, $m1, $m2, $m3, $m4)"
override def hashCode: Int = scala.util.hashing.MurmurHash3.productHash(this)
override def equals(that: Any): Boolean =
that match {
case thatM: Moments =>
(m0D == thatM.m0D) &&
(m1 == thatM.m1) &&
(m2 == thatM.m2) &&
(m3 == thatM.m3) &&
(m4 == thatM.m4)
case _ => false
}
/**
* Scale all the moments by a constant. This allows you to use Moments with exponential decay
*/
def scale(z: Double): Moments =
if (z < 0.0) // the "extraneous" if here is to avoid allocating the error message unless necessary
throw new IllegalArgumentException(s"cannot scale by negative value: $z")
else if (z == 0)
Moments.momentsMonoid.zero
else
new Moments(m0D = z * m0D, m1 = m1, m2 = z * m2, m3 = z * m3, m4 = z * m4)
}
object Moments {
final class MomentsState(
var count: Double,
var mean: Double,
var m2: Double,
var m3: Double,
var m4: Double
) {
def +=(b: Moments): this.type = {
/*
* Unfortunately we copy the code from the monoid's plus implementation,
* but we do it to avoid allocating a new Moments on every item in the
* loop. the Monoid laws test that sum matches looping on plus
*/
val countCombined = count + b.m0D
if (countCombined == 0.0) {
mean = 0.0
m2 = 0.0
m3 = 0.0
m4 = 0.0
} else {
val delta = b.mean - mean
val delta_n = delta / countCombined
val delta_n2 = delta_n * delta_n
val delta_n3 = delta_n2 * delta_n
val count_sq = count * count
val rn_sq = b.m0D * b.m0D
val meanCombined = Moments.getCombinedMeanDouble(count, mean, b.m0D, b.mean)
val m2Combined = m2 + b.m2 + delta * delta_n * count * b.m0D
val m3Combined = m3 + b.m3 +
delta * delta_n2 * count * b.m0D * (count - b.m0D) +
3 * delta_n * (count * b.m2 - b.m0D * m2)
val m4Combined = m4 + b.m4 +
delta * delta_n3 * count * b.m0D *
(count_sq - count * b.m0D + rn_sq) +
6 * delta_n2 * (count_sq * b.m2 + rn_sq * m2) +
4 * delta_n * (count * b.m3 - b.m0D * m3)
mean = meanCombined
m2 = m2Combined
m3 = m3Combined
m4 = m4Combined
}
count = countCombined
this
}
def +=(b: Double): this.type = {
val prevCount = count
count += 1
val delta = b - mean
val delta_n = delta / count
val delta_n2 = delta_n * delta_n
val term1 = delta * delta_n * prevCount
mean = Moments.getCombinedMeanDouble(prevCount, mean, 1.0, b)
m4 += term1 * delta_n2 * (count * count - 3 * count + 3) +
6 * delta_n2 * m2 - 4 * delta_n * m3
m3 += term1 * delta_n * (count - 2) - 3 * delta_n * m2
m2 += term1
this
}
def toMoments: Moments = new Moments(count, mean, m2, m3, m4)
def resetFromMoments(m: Moments): this.type = {
count = m.m0D
mean = m.m1
m2 = m.m2
m3 = m.m3
m4 = m.m4
this
}
}
object MomentsState {
def fromMoments(m: Moments): MomentsState =
new MomentsState(m.m0D, m.m1, m.m2, m.m3, m.m4)
def newEmpty(): MomentsState =
new MomentsState(0.0, 0.0, 0.0, 0.0, 0.0)
}
@deprecated("use monoid[Moments], this isn't lawful for negate", "0.13.8")
def group: Group[Moments] with CommutativeGroup[Moments] =
MomentsGroup
implicit val momentsMonoid: Monoid[Moments] with CommutativeMonoid[Moments] =
new MomentsMonoid
val aggregator: MomentsAggregator.type = MomentsAggregator
val fold: Fold[Double, Moments] = momentsMonoid.zero.fold
def numericAggregator[N](implicit num: Numeric[N]): MonoidAggregator[N, Moments, Moments] =
Aggregator.prepareMonoid { n: N => Moments(num.toDouble(n)) }
/**
* Create a Moments object given a single value. This is useful for initializing moment calculations at the
* start of a stream.
*/
def apply[V: Numeric](value: V)(implicit num: Numeric[V]): Moments =
new Moments(1.0, num.toDouble(value), 0, 0, 0)
def apply[V](m0: Long, m1: V, m2: V, m3: V, m4: V)(implicit num: Numeric[V]): Moments =
new Moments(m0, num.toDouble(m1), num.toDouble(m2), num.toDouble(m3), num.toDouble(m4))
/**
* This it the legacy apply when count was a Long
*/
def apply(m0: Long, m1: Double, m2: Double, m3: Double, m4: Double): Moments =
new Moments(m0, m1, m2, m3, m4)
/**
* This it the legacy unapply when count was a Long
*/
def unapply(m: Moments): Option[(Long, Double, Double, Double, Double)] =
Some((m.m0, m.m1, m.m2, m.m3, m.m4))
/**
* When combining averages, if the counts sizes are too close we should use a different algorithm. This
* constant defines how close the ratio of the smaller to the total count can be:
*/
private[this] val STABILITY_CONSTANT = 0.1
/**
* Given two streams of doubles (weightN, an) and (weightK, ak) of form (weighted count, mean), calculates
* the mean of the combined stream.
*
* Uses a more stable online algorithm which should be suitable for large numbers of records similar to:
* http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm
*
* This differs from the implementation in MomentsGroup.scala only in that here, the counts are weighted,
* and are thus doubles instead of longs
*/
def getCombinedMeanDouble(weightN: Double, an: Double, weightK: Double, ak: Double): Double =
if (weightN < weightK) getCombinedMeanDouble(weightK, ak, weightN, an)
else
(weightN + weightK) match {
case 0.0 => 0.0
case newCount if newCount == weightN => an
case newCount =>
val scaling = weightK / newCount
// a_n + (a_k - a_n)*(k/(n+k)) is only stable if n is not approximately k
if (scaling < STABILITY_CONSTANT) an + (ak - an) * scaling
else (weightN * an + weightK * ak) / newCount
}
}
class MomentsMonoid extends Monoid[Moments] with CommutativeMonoid[Moments] {
/**
* When combining averages, if the counts sizes are too close we should use a different algorithm. This
* constant defines how close the ratio of the smaller to the total count can be:
*/
private val STABILITY_CONSTANT = 0.1
/**
* Given two streams of doubles (n, an) and (k, ak) of form (count, mean), calculates the mean of the
* combined stream.
*
* Uses a more stable online algorithm which should be suitable for large numbers of records similar to:
* http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Parallel_algorithm
*
* we no longer use this, but we can't remove it due to binary compatibility
*/
@deprecated("Use Moments.getCombinedMeanDouble instead", since = "0.13.8")
def getCombinedMean(n: Long, an: Double, k: Long, ak: Double): Double =
if (n < k) getCombinedMean(k, ak, n, an)
else
(n + k) match {
case 0L => 0.0
case newCount if newCount == n => an
case newCount =>
val scaling = k.toDouble / newCount
// a_n + (a_k - a_n)*(k/(n+k)) is only stable if n is not approximately k
if (scaling < STABILITY_CONSTANT) an + (ak - an) * scaling
else (n * an + k * ak) / newCount
}
override val zero: Moments = new Moments(0.0, 0.0, 0.0, 0.0, 0.0)
// Combines the moment calculations from two streams.
// See http://en.wikipedia.org/wiki/Algorithms_for_calculating_variance#Higher-order_statistics
// for more information on the formulas used to update the moments.
override def plus(a: Moments, b: Moments): Moments = {
val countCombined = a.m0D + b.m0D
if (countCombined == 0.0) zero
else {
val delta = b.mean - a.mean
val delta_n = delta / countCombined
val delta_n2 = delta_n * delta_n
val delta_n3 = delta_n2 * delta_n
val ln_sq = a.m0D * a.m0D
val rn_sq = b.m0D * b.m0D
val meanCombined = Moments.getCombinedMeanDouble(a.m0D, a.mean, b.m0D, b.mean)
val m2 = a.m2 + b.m2 + delta * delta_n * a.m0D * b.m0D
val m3 = a.m3 + b.m3 +
delta * delta_n2 * a.m0D * b.m0D * (a.m0D - b.m0D) +
3 * delta_n * (a.m0D * b.m2 - b.m0D * a.m2)
val m4 = a.m4 + b.m4 +
delta * delta_n3 * a.m0D * b.m0D * (ln_sq - a.m0D * b.m0D + rn_sq) +
6 * delta_n2 * (ln_sq * b.m2 + rn_sq * a.m2) +
4 * delta_n * (a.m0D * b.m3 - b.m0D * a.m3)
new Moments(countCombined, meanCombined, m2, m3, m4)
}
}
override def sumOption(items: TraversableOnce[Moments]): Option[Moments] =
if (items.isEmpty) None
else {
val iter = items.toIterator
val init = iter.next()
// If there is only a single item, skip the MomentsState instantiation and
// return it.
if (!iter.hasNext) {
Some(init)
} else {
val state = Moments.MomentsState.fromMoments(init)
while (iter.hasNext) {
state += iter.next()
}
Some(state.toMoments)
}
}
}
/**
* This should not be used as a group (avoid negate and minus). It was wrongly believed that this was a group
* for several years in this code, however it was only being tested with positive counts (which is to say the
* generators were too weak). It isn't the case that minus and negate are totally wrong but (a - a) + b in
* general isn't associative: it won't equal a - (a - b) which it should.
*/
@deprecated("use Moments.momentsMonoid, this isn't lawful for negative counts", "0.13.8")
object MomentsGroup extends MomentsMonoid with Group[Moments] with CommutativeGroup[Moments] {
override def negate(a: Moments): Moments =
new Moments(-a.m0D, a.m1, -a.m2, -a.m3, -a.m4)
}
object MomentsAggregator extends MonoidAggregator[Double, Moments, Moments] {
override val monoid: MomentsGroup.type = MomentsGroup
override def prepare(input: Double): Moments = Moments(input)
override def present(m: Moments): Moments = m
}