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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968
/*!
Naïve Bayes examples
# Naïve Bayes
The naïve bayes assumption is that all features in the labelled data are independent of each other
given the class they correspond to. This means the probability of some input given a class
can be computed as the product of each individual feature in that input conditioned on that class.
By Baye's Theorum we can relate the probability of the class given the input to the probability
of the input given the class. As a classifier only needs to determine which class some input
is most likely to be we can compare just the product of the probability of the input given a class
and the probability of that class.
## Bayes' Theorum
posterior = ( prior x likelihood ) / evidence
P(C<sub>k</sub> | **x**) = ( P(C<sub>k</sub>) * P(**x** | C<sub>k</sub>) ) / P(**x**)
P(C<sub>k</sub> | **x**) ∝ P(C<sub>k</sub>) * P(**x** | C<sub>k</sub>)
where C<sub>k</sub> is the kth class and **x** is the input to classify.
## Classifier
taking logs on Bayes' rule yields
log(P(C<sub>k</sub> | **x**)) ∝ log(P(C<sub>k</sub>)) + log(P(**x** | C<sub>k</sub>))
given the naïve bayes assumption
log(P(C<sub>k</sub> | **x**)) ∝ log(P(C<sub>k</sub>)) + the sum over all i features of (log(P(x<sub>i</sub> | C<sub>k</sub>)))
Then to determine the class we take the class corresponding to the largest
log(P(C<sub>k</sub> | **x**)).
Computing the individual probabilities of a feature conditioned on a class depends on what the
type of data is.
For categorical data this is simply occurances in class / total occurances. In practise laplacian
smoothing (adding one occurance for each class to the computed probabilities) may be used to
avoid computing a probability of 0 when some category doesn't have any samples for a class.
For continuous data we can model the feature as distributed according to a Gaussian distribution.
## Simple Naïve Bayes Example with F-1 score analysis
Naïve Bayes can be done by hand (with a calculator) which is what the below example will show.
We have a list of data about the environment and want to predict if we should go outside based
on the conditions. Some of these are categorical values and others are real valued.
```
use easy_ml::distributions::Gaussian;
use easy_ml::linear_algebra;
#[derive(Clone, Copy, PartialEq, Debug)]
enum Weather {
Stormy, Rainy, Cloudy, Clear, Sunny
}
type WindSpeed = f64;
#[derive(Clone, Copy, PartialEq, Debug)]
enum Pandemic {
Pandemic, NoPandemic
}
#[derive(Clone, Copy, PartialEq, Debug)]
enum Decision {
GoOut, StayIn
}
#[derive(Clone, Copy, PartialEq, Debug)]
struct Observation {
weather: Weather,
wind: WindSpeed,
pandemic: Pandemic,
decision: Decision,
}
impl Observation {
fn new(
weather: Weather, wind: WindSpeed, pandemic: Pandemic, decision: Decision
) -> Observation {
Observation {
weather,
wind,
pandemic,
decision
}
}
}
let observations = vec![
Observation::new(Weather::Clear, 0.5, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Clear, 0.9, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Clear, 0.8, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Stormy, 0.7, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Rainy, 0.1, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Rainy, 0.5, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Rainy, 0.6, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Rainy, 0.7, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Cloudy, 0.3, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Cloudy, 0.5, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Cloudy, 0.2, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Cloudy, 0.8, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Sunny, 0.3, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Sunny, 0.9, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Sunny, 0.5, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Sunny, 0.5, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Sunny, 0.5, Pandemic::Pandemic, Decision::StayIn),
Observation::new(Weather::Clear, 0.1, Pandemic::Pandemic, Decision::StayIn),
Observation::new(Weather::Clear, 0.9, Pandemic::Pandemic, Decision::StayIn)
];
fn predict(
observations: &Vec<Observation>, weather: Weather, wind: WindSpeed, pandemic: Pandemic
) -> Decision {
let total = observations.len() as f64;
// first compute the number of each class in the data
let total_stay_in = observations.iter()
.filter(|observation| observation.decision == Decision::StayIn)
.count() as f64;
let total_go_out = observations.iter()
.filter(|observation| observation.decision == Decision::GoOut)
.count() as f64;
let weather_log_probability_stay_in = {
// compute how many rows in the data are this weather and stay in
let total = observations.iter()
.filter(|observation| observation.decision == Decision::StayIn)
.filter(|observation| observation.weather == weather)
.count() as f64;
// there are 5 variants for the weather and we use laplacian smoothing
// to avoid introducing zero probabilities, +1 / +5 treats the data
// as if there is one stay in for each weather type.
((total + 1.0) / (total_stay_in + 5.0)).ln()
};
let weather_log_probability_go_out = {
// compute how many rows in the data are this weather and go out
let total = observations.iter()
.filter(|observation| observation.decision == Decision::GoOut)
.filter(|observation| observation.weather == weather)
.count() as f64;
// there are 5 variants for the weather and we use laplacian smoothing
// to avoid introducing zero probabilities, +1 / +5 treats the data
// as if there is one go out for each weather type.
((total + 1.0) / (total_go_out + 5.0)).ln()
};
// we're modelling the wind as a Gaussian so we get a probability by
// computing the probability density for the wind we have, if it is
// the same as the mean wind for stay in then it will be closest to 1
// and the further away the closer to 0 it will be
let wind_speed_model_stay_in: Gaussian<WindSpeed> = Gaussian::approximating(
observations.iter()
.filter(|observation| observation.decision == Decision::StayIn)
.map(|observation| observation.wind)
);
let wind_log_probability_stay_in = wind_speed_model_stay_in.probability(&wind);
let wind_speed_model_go_out: Gaussian<WindSpeed> = Gaussian::approximating(
observations.iter()
.filter(|observation| observation.decision == Decision::GoOut)
.map(|observation| observation.wind)
);
let wind_log_probability_go_out = wind_speed_model_go_out.probability(&wind);
let pandemic_log_probability_stay_in = {
// compute how many rows in the data are this pandemic and stay in
let total = observations.iter()
.filter(|observation| observation.decision == Decision::StayIn)
.filter(|observation| observation.pandemic == pandemic)
.count() as f64;
// there are 2 variants for the pandemic type and we use laplacian smoothing
// to avoid introducing zero probabilities, +1 / +2 treats the data
// as if there is one stay in for each pandemic type.
((total + 1.0) / (total_stay_in + 2.0)).ln()
};
let pandemic_log_probability_go_out = {
// compute how many rows in the data are this pandemic and go out
let total = observations.iter()
.filter(|observation| observation.decision == Decision::GoOut)
.filter(|observation| observation.pandemic == pandemic)
.count() as f64;
// there are 2 variants for the pandemic type and we use laplacian smoothing
// to avoid introducing zero probabilities, +1 / +2 treats the data
// as if there is one go out for each pandemic type.
((total + 1.0) / (total_go_out + 2.0)).ln()
};
let prior_log_probability_stay_in = total_stay_in / total;
let prior_log_probability_go_out = total_go_out / total;
let posterior_log_probability_stay_in = (
prior_log_probability_stay_in
+ weather_log_probability_stay_in
+ wind_log_probability_stay_in
+ pandemic_log_probability_stay_in
);
let posterior_log_probability_go_out = (
prior_log_probability_go_out
+ weather_log_probability_go_out
+ wind_log_probability_go_out
+ pandemic_log_probability_go_out
);
if posterior_log_probability_go_out > posterior_log_probability_stay_in {
Decision::GoOut
} else {
Decision::StayIn
}
}
let test_data = vec![
Observation::new(Weather::Sunny, 0.8, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Sunny, 0.2, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Stormy, 0.2, Pandemic::NoPandemic, Decision::StayIn),
Observation::new(Weather::Cloudy, 0.3, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Rainy, 0.8, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Rainy, 0.5, Pandemic::NoPandemic, Decision::GoOut),
Observation::new(Weather::Stormy, 0.6, Pandemic::Pandemic, Decision::StayIn),
Observation::new(Weather::Rainy, 0.1, Pandemic::Pandemic, Decision::StayIn),
];
let predictions = test_data.iter()
.map(|data| predict(&observations, data.weather, data.wind, data.pandemic))
.collect::<Vec<Decision>>();
println!("Test data and predictions\n{:?}", test_data.iter()
.cloned()
.zip(predictions.clone())
.collect::<Vec<(Observation, Decision)>>());
println!("Accuracy: {:?}", test_data.iter()
.zip(predictions.clone())
.map(|(data, decision)| if data.decision == decision { 1.0 } else { 0.0 })
.sum::<f64>() / (test_data.len() as f64));
// To compute Recall and Precision it is neccessary to decide which class should be
// considered the positive and which should be the negative. For medical diagnosis
// this is always diagnosing something as present. For this example we take GoOut to
// be the positive case.
// True Positives are where the model predicts the positive class when it is
// the correct decision, eg in this scenario, going outside when it should
// decide to go outside
let true_positives = test_data.iter()
.cloned()
.zip(predictions.clone())
.filter(|(data, _)| data.decision == Decision::GoOut)
.filter(|(_, decision)| decision == &Decision::GoOut)
.count() as f64;
// False Positives are when the model predicts the positive class when it is
// not the correct decision, eg in this scenario, going outside when it should
// decide to stay in
let false_positives = test_data.iter()
.cloned()
.zip(predictions.clone())
.filter(|(data, _)| data.decision == Decision::StayIn)
.filter(|(_, decision)| decision == &Decision::GoOut)
.count() as f64;
// True Negatives are when the model predicts the negative class when it is
// the correct decision, eg in this scenario, staying in when it should
// decide to stay in
// there's a pandemic and no good reason to go outside
let true_negatives = test_data.iter()
.cloned()
.zip(predictions.clone())
.filter(|(data, _)| data.decision == Decision::StayIn)
.filter(|(_, decision)| decision == &Decision::StayIn)
.count() as f64;
// False Negatives are when the model predicts the negative class when it is
// not the correct decision, eg in this scenario, staying in when it should
// decide to go outside
// there's a pandemic and no good reason to go outside
let false_negatives = test_data.iter()
.cloned()
.zip(predictions.clone())
.filter(|(data, _)| data.decision == Decision::GoOut)
.filter(|(_, decision)| decision == &Decision::StayIn)
.count() as f64;
// Precision measures how good the model is at identifying the positive
// class (you can trivially get 100% precision by never predicting the
// positive class, as this means you can't get a false positive).
let precision = true_positives / (true_positives + false_positives);
// Recall is the true positive rate which is how good the model is at
// identifying the positive class out of all the positive cases (you can
// trivially get 100% recall by always predicting the positive class).
let recall = true_positives / (true_positives + false_negatives);
// The F-1 score is the harmonic mean of precision and recall and
// averages them for an accuracy measure.
// In this case the two classes are roughly equally likely, so the F-1
// score and Accuracy are similar. However, if the model had learned
// to always predict GoOut, then it would still have an accuracy of
// roughly 50% because of the equal likelihood, whereas its F-1 score
// would be much lower than 50%.
let f1_score = linear_algebra::f1_score(precision, recall);
println!("F1-Score: {:?}", f1_score);
assert!(f1_score > 0.8);
```
## 3 Class Naïve Bayes Example
[See submodule for 3 Class Naïve Bayes Example](three_class)
*/
#[rustfmt::skip]
pub mod three_class {
/*!
# 3 Class Naïve Bayes Example
For this example some population data is generated for a fictional alien race as I didn't
have any real datasets to hand. This alien race has 3 sexes (mysteriously no individuals
are ever intersex or trans) and is sexually dimorphic, meaning we can try to determine their
sex from various measuremnts.
As with humans, a gaussian distribution for physical charateristics is sensible due to
evolutionary and biological factors.
The example includes categorical features such as marking color and real valued
features such as height in order to show how both can be modelled with Naïve Bayes.
Note that most of the code below is for generating and clustering the data to perform
Naïve Bayes on, not doing it.
### Clustering
After generating the unlabelled data clustering is performed on a very small subset of that data
to find three cluster centres that are then used to assign the whole dataset sex class labels.
This creates a labelled dataset to perform Naïve Bayes on to see if we can predict sex given
the various features.
By clustering only a very small bit of the data, by chance we can expect there to be large gaps
in the subset because our data has many dimensions.
```
use easy_ml::matrices::Matrix;
use easy_ml::matrices::slices::{Slice2D, Slice};
use easy_ml::linear_algebra;
use easy_ml::distributions::Gaussian;
use rand::{Rng, SeedableRng};
use rand::distributions::{DistIter, Standard};
use rand_chacha::ChaCha8Rng;
#[derive(Clone, Copy, PartialEq, Debug)]
struct Alien {
height: f64,
primary_marking_color: AlienMarkingColor,
tail_length: f64,
metabolic_rate: f64,
spiked_tail: bool,
sex: AlienSex,
}
#[derive(Clone, Copy, Eq, PartialEq, Debug)]
enum AlienMarkingColor {
Red = 1, Yellow, Orange, Blue, Purple, White, Black
}
#[derive(Clone, Copy, Eq, PartialEq, Debug)]
enum AlienSex {
A, B, C
}
/**
* ===============================
* = Data Generation =
* ===============================
*/
// rather than explicitly define a generative function that already knows the relationship
// between alien charateristics instead the samples are generated without an assigned sex
// and then clustered using k-means
// use a fixed seed random generator from the rand crate
let mut random_generator = ChaCha8Rng::seed_from_u64(16);
let mut random_numbers: DistIter<Standard, &mut ChaCha8Rng, f64> =
(&mut random_generator).sample_iter(Standard);
/**
* Generates height data for creating the alien dataset.
*/
fn generate_heights(samples: usize, random_numbers: &mut impl Iterator<Item = f64>) -> Vec<f64> {
// the average height shall be 1.5 meters with a standard deviation of 0.25
let heights_distribution = Gaussian::new(1.5, 0.25 * 0.25);
let mut heights = heights_distribution.draw(random_numbers, samples).unwrap();
// ensure all aliens are at least 0.25 meters tall
heights.drain(..).map(|x| if x > 0.25 { x } else { 0.25 }).collect()
}
/**
* Generates tail length data for creating the alien dataset.
*/
fn generate_tail_length(samples: usize, random_numbers: &mut impl Iterator<Item = f64>) -> Vec<f64> {
// the average length shall be 1.25 meters with more variation in tail length
let tails_distribution = Gaussian::new(1.25, 0.5 * 0.5);
let mut tails = tails_distribution.draw(random_numbers, samples).unwrap();
// ensure all tails are at least 0.5 meters long
tails.drain(..).map(|x| if x > 0.5 { x } else { 0.5 }).collect()
}
/**
* Generates color data for creating the alien dataset.
*
* Note that floats are still returned despite this being a category because we need all the
* data types to be the same for clustering
*/
fn generate_colors(samples: usize, random_numbers: &mut impl Iterator<Item = f64>) -> Vec<f64> {
let mut colors = Vec::with_capacity(samples);
for i in 0..samples {
let x = random_numbers.next().unwrap();
if x < 0.2 {
colors.push(AlienMarkingColor::Red as u8 as f64);
} else if x < 0.3 {
colors.push(AlienMarkingColor::Yellow as u8 as f64);
} else if x < 0.45 {
colors.push(AlienMarkingColor::Orange as u8 as f64);
} else if x < 0.59 {
colors.push(AlienMarkingColor::Blue as u8 as f64);
} else if x < 0.63 {
colors.push(AlienMarkingColor::Purple as u8 as f64);
} else if x < 0.9 {
colors.push(AlienMarkingColor::White as u8 as f64);
} else {
colors.push(AlienMarkingColor::Black as u8 as f64);
}
}
colors
}
/**
* Recovers the color type which is the closest match to the input floating point color
*/
fn recover_generated_color(color: f64) -> AlienMarkingColor {
let numerical_colors = [
AlienMarkingColor::Red as u8 as f64,
AlienMarkingColor::Yellow as u8 as f64,
AlienMarkingColor::Orange as u8 as f64,
AlienMarkingColor::Blue as u8 as f64,
AlienMarkingColor::Purple as u8 as f64,
AlienMarkingColor::White as u8 as f64,
AlienMarkingColor::Black as u8 as f64,
];
let colors = [
AlienMarkingColor::Red,
AlienMarkingColor::Yellow,
AlienMarkingColor::Orange,
AlienMarkingColor::Blue,
AlienMarkingColor::Purple,
AlienMarkingColor::White,
AlienMarkingColor::Black,
];
// look for the closest fit, as manipulated floating points may not be exact
let color_index = numerical_colors.iter()
// take the absolute difference so an exact match will become 0
.map(|c| (c - color).abs())
.enumerate()
// find the element with the smallest difference in the list
.min_by(|(_, a), (_, b)| a.partial_cmp(b).expect("NaN should not be in list"))
// discard the difference
.map(|(index, difference)| index)
// retrieve the value from the option
.unwrap();
colors[color_index]
}
/**
* Generates metabolic rate data for creating the alien dataset.
*/
fn generate_metabolic_rate(
samples: usize, random_numbers: &mut impl Iterator<Item = f64>
) -> Vec<f64> {
// the average rate shall be 100 heart beats per minute with a standard deviation of 20
let metabolic_rate_distribution = Gaussian::new(100.0, 20.0 * 20.0);
let mut metabolic_rates = metabolic_rate_distribution.draw(random_numbers, samples).unwrap();
// ensure all rates are at least 50 and less than 200
metabolic_rates.drain(..).map(|x| {
if x <= 50.0 {
50.0
} else if x >= 200.0 {
200.0
} else {
x
}
}).collect()
}
/**
* Generates spiked tailness data for creating the alien dataset.
*/
fn generate_spiked_tail(
samples: usize, random_numbers: &mut impl Iterator<Item = f64>
) -> Vec<f64> {
let mut spikes = Vec::with_capacity(samples);
for i in 0..samples {
let x = random_numbers.next().unwrap();
if x < 0.4 {
// 0 shall represent a non spiked tail
spikes.push(0.0)
} else {
// 1 shall represent a spiked tail
spikes.push(1.0)
}
}
spikes
}
/**
* Recovers the spiked tail type which is the closest match to the input floating point
*/
fn recover_generated_spiked_tail(spiked: f64) -> bool {
return spiked >= 0.5
}
// We shall generate 1000 samples for the dataset
const SAMPLES: usize = 1000;
// Collect all the float typed features into a matrix
let unlabelled_dataset = {
let mut dataset = Matrix::column(generate_heights(SAMPLES, &mut random_numbers));
dataset.insert_column_with(1,
generate_colors(SAMPLES, &mut random_numbers).drain(..));
dataset.insert_column_with(2,
generate_tail_length(SAMPLES, &mut random_numbers).drain(..));
dataset.insert_column_with(3,
generate_metabolic_rate(SAMPLES, &mut random_numbers).drain(..));
dataset.insert_column_with(4,
generate_spiked_tail(SAMPLES, &mut random_numbers).drain(..));
dataset
};
/**
* ===============================
* = Clustering =
* ===============================
*/
// Create a subset of the first 30 samples from the full dataset to use for clustering
let unlabelled_subset = unlabelled_dataset.retain(
Slice2D::new()
.columns(Slice::All())
.rows(Slice::Range(0..30))
);
// We normalise all the features to 0 mean and 1 standard deviation because
// we will use euclidean distance as the distance matric, and our features
// have very different variances. This avoids the distance metric being
// dominated by any particular feature.
let mut means_and_variances = Vec::with_capacity(unlabelled_subset.columns());
// The normalised subset is computed taking the mean and variance from the subset,
// these means and variances will be needed later to apply to the rest of the data.
let mut normalised_subset = {
let mut normalised_subset = unlabelled_subset;
for feature in 0..normalised_subset.columns() {
let mean = linear_algebra::mean(normalised_subset.column_iter(feature));
let variance = linear_algebra::variance(normalised_subset.column_iter(feature));
// save the data for normalisation and denormalisation for each feature
// for use later
means_and_variances.push(vec![ mean, variance ]);
for row in 0..normalised_subset.rows() {
let x = normalised_subset.get(row, feature);
normalised_subset.set(row, feature, (x - mean) / variance);
}
}
normalised_subset
};
// create a 5 x 2 matrix where each row is the mean and variance of each of the 5 features
let means_and_variances = Matrix::from(means_and_variances);
// pick the first 3 samples as the starting points for the cluster centres
// and place them into a 3 x 5 matrix where we have 3 rows of cluster centres
// and 5 features which are all normalised
let mut clusters = Matrix::from(vec![
normalised_subset.row_iter(0).collect(),
normalised_subset.row_iter(1).collect(),
normalised_subset.row_iter(2).collect()]);
// add a 6th column to the subset to track the closest cluster for each sample
const CLUSTER_ID_COLUMN: usize = 5;
normalised_subset.insert_column(CLUSTER_ID_COLUMN, -1.0);
// set a threshold at which we consider the cluster centres to have converged
const CHANGE_THRESHOLD: f64 = 0.001;
// track how much the means have changed each update
let mut absolute_changes = -1.0;
// loop until we go under the CHANGE_THRESHOLD, reassigning points to the nearest
// cluster then cluster centres to their mean of points
while absolute_changes == -1.0 || absolute_changes > CHANGE_THRESHOLD {
// assign each point to the nearest cluster centre by euclidean distance
for point in 0..normalised_subset.rows() {
let mut closest_cluster = -1.0;
let mut least_squared_distance = std::f64::MAX;
for cluster in 0..clusters.rows() {
// we don't actually need to square root the distances for finding
// which is least because least squared distance is the same as
// least distance
let squared_distance = {
let mut sum = 0.0;
for feature in 0..clusters.columns() {
let cluster_coordinate = clusters.get(cluster, feature);
let point_coordiante = normalised_subset.get(point, feature);
sum += (cluster_coordinate - point_coordiante).powi(2);
}
sum
};
if squared_distance < least_squared_distance {
closest_cluster = cluster as f64;
least_squared_distance = squared_distance;
}
}
// save the cluster that is closest to each point in the 6th column
normalised_subset.set(point, CLUSTER_ID_COLUMN, closest_cluster);
}
// update cluster centres to the mean of their points
absolute_changes = 0.0;
for cluster in 0..clusters.rows() {
// construct a list of the points this cluster owns
let owned = normalised_subset.column_iter(CLUSTER_ID_COLUMN)
// zip together the id values with their index
.enumerate()
// exclude the points that aren't assigned to this cluster
.filter(|(index, id)| (*id as usize) == cluster)
// drop the cluster ids from each item and copy over the data
// for each point for each feature
.map(|(index, id)| {
// for each point copy all its data in each feature excluding the
// final cluster id column into a new vec
normalised_subset.row_iter(index)
// taking the first 5 excludes the 6th column due to 0 indexing
.take(CLUSTER_ID_COLUMN)
.collect::<Vec<f64>>()
})
// collect into a vector of vectors containing each feature's data
.collect::<Vec<Vec<f64>>>();
// pass the vector of vectors into a matrix so we have
// a matrix where each row is the data of a point this cluster owns
let owned = Matrix::from(owned);
// construct a vector of the mean for each feature that this cluster
// now has
let new_means = {
let mut means = Vec::with_capacity(owned.rows());
for feature in 0..owned.columns() {
let mean = owned.column_iter(feature).sum::<f64>() / (owned.rows() as f64);
means.push(mean);
}
means
};
// update each new mean for the cluster
for feature in 0..clusters.columns() {
let previous_mean = clusters.get(cluster, feature);
// track the absolute difference between the new mean and the old one
// so we know when to stop updating the clusters
absolute_changes += (previous_mean - new_means[feature]).abs();
clusters.set(cluster, feature, new_means[feature]);
}
}
}
println!(
"Denormalised clusters at convergence:\n{:?}\n{:.3}",
vec![ "H", "C", "T", "M", "S" ],
clusters.map_with_index(|x, _, feature| {
let mean = means_and_variances.get(feature, 0);
let variance = means_and_variances.get(feature, 1);
(x * variance) + mean
}));
// Now we will assign every alien in the full dataset a sex using these cluster centres
let mut aliens: Vec<Alien> = Vec::with_capacity(unlabelled_dataset.rows());
fn assign_alien_sex(index: u8) -> AlienSex {
if index == 0 {
AlienSex::A
} else if index == 1 {
AlienSex::B
} else {
AlienSex::C
}
}
for i in 0..unlabelled_dataset.rows() {
let alien_data = Matrix::column(unlabelled_dataset.row_iter(i).collect());
// normalise the alien data first so comparisons are on unit variance
// and zero mean
let normalised_alien_data = alien_data.map_with_index(|x, feature, _| {
let mean = means_and_variances.get(feature, 0);
let variance = means_and_variances.get(feature, 1);
// normalise each feature in the alien data
(x - mean) / variance
});
let mut distances = Vec::with_capacity(clusters.rows());
for j in 0..clusters.rows() {
let cluster_data = Matrix::row(clusters.row_iter(j).collect());
// use euclidean distance to compare the alien with the cluster, the cluster
// is already normalised
let sum_of_squares = (cluster_data * &normalised_alien_data).scalar();
distances.push(sum_of_squares);
}
// find the cluster with the lowest distance to each point and get its index
let chosen_cluster = distances.iter()
.enumerate()
.min_by(|(_, a), (_, b)| a.partial_cmp(b).expect("NaN should not be in list"))
.map(|(i, _)| i)
.unwrap();
// convert each float to the correct data type
aliens.push(Alien {
height: alien_data.get(0, 0),
primary_marking_color: recover_generated_color(alien_data.get(1, 0)),
tail_length: alien_data.get(2, 0),
metabolic_rate: alien_data.get(3, 0),
spiked_tail: recover_generated_spiked_tail(alien_data.get(4, 0)),
sex: assign_alien_sex(chosen_cluster as u8),
})
}
println!("First 10 aliens");
for i in 0..10 {
println!("{:?}", aliens[i]);
}
// Put the aliens in a matrix for convenience
let aliens = Matrix::row(aliens);
println!("Sex A aliens total: {}", aliens.row_reference_iter(0)
.fold(0, |accumulator, alien| accumulator + if alien.sex == AlienSex::A { 1 } else { 0 }));
println!("Sex B aliens total: {}", aliens.row_reference_iter(0)
.fold(0, |accumulator, alien| accumulator + if alien.sex == AlienSex::B { 1 } else { 0 }));
println!("Sex C aliens total: {}", aliens.row_reference_iter(0)
.fold(0, |accumulator, alien| accumulator + if alien.sex == AlienSex::C { 1 } else { 0 }));
/**
* ===============================
* = Naïve Bayes =
* ===============================
*/
// Each class is roughly one third so we should not have a strong prior to a particular class
// In order to evaluate the performance of the Naïve Bayes classifier we will hold out
// the last 100 aliens from the dataset and use them as training data
let training_data = Matrix::column(aliens.row_iter(0).take(900).collect());
let test_data = Matrix::column(aliens.row_iter(0).skip(900).collect());
/**
* Predicts the most probable alien sex for each test input alien (disregarding
* the sex field in those inputs)
*
* For the real valued features the probabilities are computed by modelling
* the features (conditioned on each class) as gaussian distributions.
* For categorical features laplacian smoothing of the counts is used to
* estimate probabilities of the features (conditioned on each class).
*/
fn predict_aliens(training_data: &Matrix<Alien>, test_data: &Matrix<Alien>) -> Matrix<AlienSex> {
let mut relative_log_probabilities = Vec::with_capacity(3);
for class in &[ AlienSex::A, AlienSex::B, AlienSex::C ] {
let training_data_class_only = training_data.column_iter(0)
.filter(|a| &a.sex == class)
.collect::<Vec<Alien>>();
// compute how likely each class is in the training set
let prior = (training_data_class_only.len() as f64) / (training_data.rows() as f64);
// We model the real valued features as Gaussians, note that these
// are Gaussian distributions over only the training data of each class
let heights: Gaussian<f64> = Gaussian::approximating(
training_data_class_only.iter().map(|a| a.height)
);
let tail_lengths: Gaussian<f64> = Gaussian::approximating(
training_data_class_only.iter().map(|a| a.tail_length)
);
let metabolic_rates: Gaussian<f64> = Gaussian::approximating(
training_data_class_only.iter().map(|a| a.metabolic_rate)
);
// gradually build up the sum of log probabilities to get the
// log of the prior * likelihood which will be proportional to the posterior
let relative_log_probabilities_of_class = test_data.column_reference_iter(0)
.map(|alien| {
// probabilitiy of the alien sex and the alien
let mut log_relative_probability = prior.ln();
// Compute the probability using the Gaussian model for each real valued feature.
// Due to floating point precision limits and the variance for some of these
// Gaussian models being extremely small (0.01 for heights) we
// check if a probability computed is zero or extremely close to zero
// and if so increase it a bit to avoid computing -inf when we take the log.
let mut height_given_class = heights.probability(&alien.height);
if height_given_class.abs() <= 0.000000000001 {
height_given_class = 0.000000000001;
}
log_relative_probability += height_given_class.ln();
let mut tail_given_class = tail_lengths.probability(&alien.tail_length);
if tail_given_class.abs() <= 0.000000000001 {
tail_given_class = 0.000000000001;
}
log_relative_probability += tail_given_class.ln();
let mut metabolic_rates_given_class = metabolic_rates.probability(
&alien.metabolic_rate);
if metabolic_rates_given_class.abs() <= 0.000000000001 {
metabolic_rates_given_class = 0.000000000001;
}
log_relative_probability += metabolic_rates_given_class.ln();
// compute the probability of the categorical features using lapacian smoothing
let color_of_class = training_data_class_only.iter()
.map(|a| a.primary_marking_color)
// count how many aliens of this class have this color
.fold(0, |acc, color|
acc + if color == alien.primary_marking_color { 1 } else { 0 });
// with laplacian smoothing we assume there is one datapoint for each color
// which avoids zero probabilities but does not distort the probabilities much
// there are 7 color types so we add 7 to the total
let color_given_class = ((color_of_class + 1) as f64)
/ ((training_data_class_only.len() + 7) as f64);
log_relative_probability += color_given_class.ln();
let spiked_tail_of_class = training_data_class_only.iter()
.map(|a| a.spiked_tail)
// count how many aliens of this class have a spiked tail or not
.fold(0, |acc, spiked| acc + if spiked == alien.spiked_tail { 1 } else { 0 });
// again we assume one alien of the class with a spiked tail and one without
// to avoid zero probabilities
let spiked_tail_given_class = ((spiked_tail_of_class + 1) as f64)
/ ((training_data_class_only.len() + 2) as f64);
log_relative_probability += spiked_tail_given_class.ln();
if log_relative_probability == std::f64::NEG_INFINITY {
println!("Individual probs P:{} H:{} T:{} M:{} C:{} S:{}",
prior, height_given_class, tail_given_class, metabolic_rates_given_class,
color_given_class, spiked_tail_given_class);
}
log_relative_probability
}).collect();
relative_log_probabilities.push(relative_log_probabilities_of_class);
}
// collect the relative probabilitiy estimates for each class and each alien
// into a 3 x 100 matrix respectively
let probabilities = Matrix::from(relative_log_probabilities);
let predictions = (0..probabilities.columns()).map(|i| {
let predicted_class_index = probabilities.column_iter(i)
.enumerate()
// find the class with the highest relative probability estimate
.max_by(|(_, a), (_, b)| a.partial_cmp(b).expect("NaN should not be in list"))
// discard the probability
.map(|(index, p)| index)
// retrieve the value from the option
.unwrap();
if predicted_class_index == 0 {
AlienSex::A
} else if predicted_class_index == 1 {
AlienSex::B
} else {
AlienSex::C
}
}).collect();
Matrix::column(predictions)
}
let predictions = predict_aliens(&training_data, &test_data);
println!("First 10 test aliens and predictions");
for i in 0..10 {
println!("Predicted: {:?} Input: {:?}", predictions.get(i, 0), test_data.get(i, 0));
}
let accuracy = test_data.column_iter(0)
.zip(predictions.row_iter(0))
.map(|(alien, prediction)| if alien.sex == prediction { 1 } else { 0 })
.sum::<u16>() as f64 / (test_data.rows() as f64);
println!("Accuracy {}%", accuracy * 100.0);
/**
* ===============================
* = Analysis =
* ===============================
*/
// We can get a better sense of how well our classifier has done by
// printing the confusion matrix
// Construct a confusion matrix of actual x predicted classes, using A as 0, B as 1 and C as 2
// for indexing. If the accuracy was 100% we would see only non zero numbers on the diagonal
// as every prediction would be the actual class.
let confusion_matrix = {
let mut confusion_matrix = Matrix::empty(0, (3, 3));
// loop through all the actual and predicted classes to fill the confusion matrix
// with the total occurances of each possible combination
for (actual, predicted) in test_data.column_iter(0).zip(predictions.column_iter(0)) {
match actual.sex {
AlienSex::A => {
match predicted {
AlienSex::A => confusion_matrix.set(0, 0, confusion_matrix.get(0, 0) + 1),
AlienSex::B => confusion_matrix.set(0, 1, confusion_matrix.get(0, 1) + 1),
AlienSex::C => confusion_matrix.set(0, 2, confusion_matrix.get(0, 2) + 1),
}
},
AlienSex::B => {
match predicted {
AlienSex::A => confusion_matrix.set(1, 0, confusion_matrix.get(1, 0) + 1),
AlienSex::B => confusion_matrix.set(1, 1, confusion_matrix.get(1, 1) + 1),
AlienSex::C => confusion_matrix.set(1, 2, confusion_matrix.get(1, 2) + 1),
}
},
AlienSex::C => {
match predicted {
AlienSex::A => confusion_matrix.set(2, 0, confusion_matrix.get(2, 0) + 1),
AlienSex::B => confusion_matrix.set(2, 1, confusion_matrix.get(2, 1) + 1),
AlienSex::C => confusion_matrix.set(2, 2, confusion_matrix.get(2, 2) + 1),
}
}
}
}
confusion_matrix
};
println!("Confusion matrix: Rows are actual class, Columns are predicted class\n{}",
confusion_matrix);
println!(" A B C");
```
The above code prints 60% accuracy which isn't amazing as if we learned the cluster centers
that labelled the data in the first place we should be able to get 100% accuracy but
60% is still far better than guessing for a three class problem. As Naïve Bayes has no
hyperparameters it is good for establishing a good baseline to compare other examples such
as [Logistic Regression](super::super::logistic_regression) to.
*/
}