How to Set the Number of Trees in Random Forest

Forest — A Highly effective Device for Anybody Working With Knowledge

What’s Random Forest?

Have you ever ever wished you would make higher choices utilizing information — like predicting the danger of illnesses, crop yields, or recognizing patterns in buyer conduct? That’s the place machine studying is available in and some of the accessible and highly effective instruments on this discipline is one thing known as Random Forest.

So why is random forest so standard? For one, it’s extremely versatile. It really works effectively with many sorts of information whether or not numbers, classes, or each. It’s additionally broadly utilized in many fields — from predicting affected person outcomes in healthcare to detecting fraud in finance, from enhancing buying experiences on-line to optimising agricultural practices.

Regardless of the identify, random forest has nothing to do with timber in a forest — but it surely does use one thing known as Decision Trees to make good predictions. You may consider a call tree as a flowchart that guides a sequence of sure/no questions primarily based on the info you give it. A random forest creates an entire bunch of those timber (therefore the “forest”), every barely totally different, after which combines their outcomes to make one remaining resolution. It’s a bit like asking a bunch of specialists for his or her opinion after which going with the bulk vote.

However till not too long ago, one query was unanswered: What number of resolution timber do I really need? If every resolution tree can result in totally different outcomes, averaging many timber would result in higher and extra dependable outcomes. However what number of are sufficient? Fortunately, the optRF package deal solutions this query!

So let’s take a look at optimise Random Forest for predictions and variable choice!

Making Predictions with Random Forests

To optimise and to make use of random forest for making predictions, we will use the open-source statistics programme R. As soon as we open R, we now have to put in the 2 R packages “ranger” which permits to make use of random forests in R and “optRF” to optimise random forests. Each packages are open-source and obtainable through the official R repository CRAN. With the intention to set up and cargo these packages, the next traces of R code could be run:

> set up.packages(“ranger”)
> set up.packages(“optRF”)
> library(ranger)
> library(optRF)

Now that the packages are put in and loaded into the library, we will use the capabilities that these packages comprise. Moreover, we will additionally use the info set included within the optRF package deal which is free to make use of below the GPL license (simply because the optRF package deal itself). This information set known as SNPdata incorporates within the first column the yield of 250 wheat vegetation in addition to 5000 genomic markers (so known as single nucleotide polymorphisms or SNPs) that may comprise both the worth 0 or 2.

> SNPdata[1:5,1:5]
            Yield SNP_0001 SNP_0002 SNP_0003 SNP_0004
  ID_001 670.7588        0        0        0        0
  ID_002 542.5611        0        2        0        0
  ID_003 591.6631        2        2        0        2
  ID_004 476.3727        0        0        0        0
  ID_005 635.9814        2        2        0        2

This information set is an instance for genomic information and can be utilized for genomic prediction which is an important device for breeding high-yielding crops and, thus, to battle world starvation. The concept is to foretell the yield of crops utilizing genomic markers. And precisely for this objective, random forest can be utilized! That signifies that a random forest mannequin is used to explain the connection between the yield and the genomic markers. Afterwards, we will predict the yield of wheat vegetation the place we solely have genomic markers.

Subsequently, let’s think about that we now have 200 wheat vegetation the place we all know the yield and the genomic markers. That is the so-called coaching information set. Let’s additional assume that we now have 50 wheat vegetation the place we all know the genomic markers however not their yield. That is the so-called check information set. Thus, we separate the info body SNPdata in order that the primary 200 rows are saved as coaching and the final 50 rows with out their yield are saved as check information:

> Coaching = SNPdata[1:200,]
> Check = SNPdata[201:250,-1]

With these information units, we will now take a look at make predictions utilizing random forests!

First, we acquired to calculate the optimum variety of timber for random forest. Since we wish to make predictions, we use the perform opt_prediction from the optRF package deal. Into this perform we now have to insert the response from the coaching information set (on this case the yield), the predictors from the coaching information set (on this case the genomic markers), and the predictors from the check information set. Earlier than we run this perform, we will use the set.seed perform to make sure reproducibility although this isn’t vital (we’ll see later why reproducibility is a matter right here):

> set.seed(123)
> optRF_result = opt_prediction(y = Coaching[,1], 
+                               X = Coaching[,-1], 
+                               X_Test = Check)
  Advisable variety of timber: 19000

All the outcomes from the opt_prediction perform are actually saved within the object optRF_result, nevertheless, an important data was already printed within the console: For this information set, we should always use 19,000 timber.

With this data, we will now use random forest to make predictions. Subsequently, we use the ranger perform to derive a random forest mannequin that describes the connection between the genomic markers and the yield within the coaching information set. Additionally right here, we now have to insert the response within the y argument and the predictors within the x argument. Moreover, we will set the write.forest argument to be TRUE and we will insert the optimum variety of timber within the num.timber argument:

> RF_model = ranger(y = Coaching[,1], x = Coaching[,-1], 
+                   write.forest = TRUE, num.timber = 19000)

And that’s it! The article RF_model incorporates the random forest mannequin that describes the connection between the genomic markers and the yield. With this mannequin, we will now predict the yield for the 50 vegetation within the check information set the place we now have the genomic markers however we don’t know the yield:

> predictions = predict(RF_model, information=Check)$predictions
> predicted_Test = information.body(ID = row.names(Check), predicted_yield = predictions)

The information body predicted_Test now incorporates the IDs of the wheat vegetation along with their predicted yield:

> head(predicted_Test)
      ID predicted_yield
  ID_201        593.6063
  ID_202        596.8615
  ID_203        591.3695
  ID_204        589.3909
  ID_205        599.5155
  ID_206        608.1031

Variable Choice with Random Forests

A special method to analysing such a knowledge set can be to seek out out which variables are most vital to foretell the response. On this case, the query can be which genomic markers are most vital to foretell the yield. Additionally this may be finished with random forests!

If we deal with such a job, we don’t want a coaching and a check information set. We will merely use the complete information set SNPdata and see which of the variables are an important ones. However earlier than we try this, we should always once more decide the optimum variety of timber utilizing the optRF package deal. Since we’re insterested in calculating the variable significance, we use the perform opt_importance:

> set.seed(123)
> optRF_result = opt_importance(y=SNPdata[,1], 
+                               X=SNPdata[,-1])
  Advisable variety of timber: 40000

One can see that the optimum variety of timber is now greater than it was for predictions. That is truly usually the case. Nevertheless, with this variety of timber, we will now use the ranger perform to calculate the significance of the variables. Subsequently, we use the ranger perform as earlier than however we modify the variety of timber within the num.timber argument to 40,000 and we set the significance argument to “permutation” (different choices are “impurity” and “impurity_corrected”). 

> set.seed(123) 
> RF_model = ranger(y=SNPdata[,1], x=SNPdata[,-1], 
+                   write.forest = TRUE, num.timber = 40000,
+                   significance="permutation")
> D_VI = information.body(variable = names(SNPdata)[-1], 
+                   significance = RF_model$variable.significance)
> D_VI = D_VI[order(D_VI$importance, decreasing=TRUE),]

The information body D_VI now incorporates all of the variables, thus, all of the genomic markers, and subsequent to it, their significance. Additionally, we now have straight ordered this information body in order that an important markers are on the highest and the least vital markers are on the backside of this information body. Which signifies that we will take a look at an important variables utilizing the pinnacle perform:

> head(D_VI)
  variable significance
  SNP_0020   45.75302
  SNP_0004   38.65594
  SNP_0019   36.81254
  SNP_0050   34.56292
  SNP_0033   30.47347
  SNP_0043   28.54312

And that’s it! We’ve got used random forest to make predictions and to estimate an important variables in a knowledge set. Moreover, we now have optimised random forest utilizing the optRF package deal!

Why Do We Want Optimisation?

Now that we’ve seen how simple it’s to make use of random forest and the way rapidly it may be optimised, it’s time to take a better have a look at what’s occurring behind the scenes. Particularly, we’ll discover how random forest works and why the outcomes may change from one run to a different.

To do that, we’ll use random forest to calculate the significance of every genomic marker however as an alternative of optimising the variety of timber beforehand, we’ll stick to the default settings within the ranger perform. By default, ranger makes use of 500 resolution timber. Let’s attempt it out:

> set.seed(123) 
> RF_model = ranger(y=SNPdata[,1], x=SNPdata[,-1], 
+                   write.forest = TRUE, significance="permutation")
> D_VI = information.body(variable = names(SNPdata)[-1], 
+                   significance = RF_model$variable.significance)
> D_VI = D_VI[order(D_VI$importance, decreasing=TRUE),]
> head(D_VI)
  variable significance
  SNP_0020   80.22909
  SNP_0019   60.37387
  SNP_0043   50.52367
  SNP_0005   43.47999
  SNP_0034   38.52494
  SNP_0015   34.88654

As anticipated, the whole lot runs easily — and rapidly! In actual fact, this run was considerably sooner than once we beforehand used 40,000 timber. However what occurs if we run the very same code once more however this time with a unique seed?

> set.seed(321) 
> RF_model2 = ranger(y=SNPdata[,1], x=SNPdata[,-1], 
+                    write.forest = TRUE, significance="permutation")
> D_VI2 = information.body(variable = names(SNPdata)[-1], 
+                    significance = RF_model2$variable.significance)
> D_VI2 = D_VI2[order(D_VI2$importance, decreasing=TRUE),]
> head(D_VI2)
  variable significance
  SNP_0050   60.64051
  SNP_0043   58.59175
  SNP_0033   52.15701
  SNP_0020   51.10561
  SNP_0015   34.86162
  SNP_0019   34.21317

As soon as once more, the whole lot seems to work positive however take a better have a look at the outcomes. Within the first run, SNP_0020 had the very best significance rating at 80.23, however within the second run, SNP_0050 takes the highest spot and SNP_0020 drops to the fourth place with a a lot decrease significance rating of 51.11. That’s a big shift! So what modified?

The reply lies in one thing known as non-determinism. Random forest, because the identify suggests, includes a variety of randomness: it randomly selects information samples and subsets of variables at varied factors throughout coaching. This randomness helps forestall overfitting but it surely additionally signifies that outcomes can fluctuate barely every time you run the algorithm — even with the very same information set. That’s the place the set.seed() perform is available in. It acts like a bookmark in a shuffled deck of playing cards. By setting the identical seed, you make sure that the random decisions made by the algorithm observe the identical sequence each time you run the code. However whenever you change the seed, you’re successfully altering the random path the algorithm follows. That’s why, in our instance, an important genomic markers got here out in a different way in every run. This conduct — the place the identical course of can yield totally different outcomes as a result of inner randomness — is a traditional instance of non-determinism in machine studying.

As we simply noticed, random forest fashions can produce barely totally different outcomes each time you run them even when utilizing the identical information as a result of algorithm’s built-in randomness. So, how can we scale back this randomness and make our outcomes extra secure?

One of many easiest and handiest methods is to extend the variety of timber. Every tree in a random forest is educated on a random subset of the info and variables, so the extra timber we add, the higher the mannequin can “common out” the noise brought on by particular person timber. Consider it like asking 10 folks for his or her opinion versus asking 1,000 — you’re extra prone to get a dependable reply from the bigger group.

With extra timber, the mannequin’s predictions and variable significance rankings are likely to turn out to be extra secure and reproducible even with out setting a selected seed. In different phrases, including extra timber helps to tame the randomness. Nevertheless, there’s a catch. Extra timber additionally imply extra computation time. Coaching a random forest with 500 timber may take just a few seconds however coaching one with 40,000 timber may take a number of minutes or extra, relying on the dimensions of your information set and your laptop’s efficiency.

Nevertheless, the connection between the soundness and the computation time of random forest is non-linear. Whereas going from 500 to 1,000 timber can considerably enhance stability, going from 5,000 to 10,000 timber may solely present a tiny enchancment in stability whereas doubling the computation time. Sooner or later, you hit a plateau the place including extra timber provides diminishing returns — you pay extra in computation time however achieve little or no in stability. That’s why it’s important to seek out the precise stability: Sufficient timber to make sure secure outcomes however not so many who your evaluation turns into unnecessarily gradual.

And that is precisely what the optRF package deal does: it analyses the connection between the soundness and the variety of timber in random forests and makes use of this relationship to find out the optimum variety of timber that results in secure outcomes and past which including extra timber would unnecessarily enhance the computation time.

Above, we now have already used the opt_importance perform and saved the outcomes as optRF_result. This object incorporates the details about the optimum variety of timber but it surely additionally incorporates details about the connection between the soundness and the variety of timber. Utilizing the plot_stability perform, we will visualise this relationship. Subsequently, we now have to insert the identify of the optRF object, which measure we’re excited about (right here, we have an interest within the “significance”), the interval we wish to visualise on the X axis, and if the really helpful variety of timber ought to be added:

> plot_stability(optRF_result, measure="significance", 
+                from=0, to=50000, add_recommendation=FALSE)

This plot clearly exhibits the non-linear relationship between stability and the variety of timber. With 500 timber, random forest solely results in a stability of round 0.2 which explains why the outcomes modified drastically when repeating random forest after setting a unique seed. With the really helpful 40,000 timber, nevertheless, the soundness is close to 1 (which signifies an ideal stability). Including greater than 40,000 timber would get the soundness additional to 1 however this enhance can be solely very small whereas the computation time would additional enhance. That’s the reason 40,000 timber point out the optimum variety of timber for this information set.

The Takeaway: Optimise Random Forest to Get the Most of It

Random forest is a robust ally for anybody working with information — whether or not you’re a researcher, analyst, scholar, or information scientist. It’s simple to make use of, remarkably versatile, and extremely efficient throughout a variety of functions. However like every device, utilizing it effectively means understanding what’s occurring below the hood. On this submit, we’ve uncovered certainly one of its hidden quirks: The randomness that makes it robust may also make it unstable if not fastidiously managed. Fortuitously, with the optRF package deal, we will strike the proper stability between stability and efficiency, making certain we get dependable outcomes with out losing computational assets. Whether or not you’re working in genomics, drugs, economics, agriculture, or some other data-rich discipline, mastering this stability will provide help to make smarter, extra assured choices primarily based in your information.