Getting started in applied machine learning can be difficult, especially when working with real-world data.

Often, machine learning tutorials will recommend or require that you prepare your data in specific ways before fitting a machine learning model.

One good example is to use a one-hot encoding on categorical data.

- Why is a one-hot encoding required?
- Why can’t you fit a model on your data directly?

In this post, you will discover the answer to these important questions and better understand data preparation in general in applied machine learning.

Let’s get started.

## What is Categorical Data?

Categorical data are variables that contain label values rather than numeric values.

The number of possible values is often limited to a fixed set.

Categorical variables are often called nominal.

Some examples include:

- A “
*pet*” variable with the values: “*dog*” and “*cat*“. - A “
*color*” variable with the values: “*red*“, “*green*” and “*blue*“. - A “
*place*” variable with the values: “first”, “*second*”*and*“*third*“.

Each value represents a different category.

Some categories may have a natural relationship to each other, such as a natural ordering.

The “*place*” variable above does have a natural ordering of values. This type of categorical variable is called an ordinal variable.

## What is the Problem with Categorical Data?

Some algorithms can work with categorical data directly.

For example, a decision tree can be learned directly from categorical data with no data transform required (this depends on the specific implementation).

Many machine learning algorithms cannot operate on label data directly. They require all input variables and output variables to be numeric.

In general, this is mostly a constraint of the efficient implementation of machine learning algorithms rather than hard limitations on the algorithms themselves.

This means that categorical data must be converted to a numerical form. If the categorical variable is an output variable, you may also want to convert predictions by the model back into a categorical form in order to present them or use them in some application.

## How to Convert Categorical Data to Numerical Data?

This involves two steps:

- Integer Encoding
- One-Hot Encoding

### 1. Integer Encoding

As a first step, each unique category value is assigned an integer value.

For example, “*red*” is 1, “*green*” is 2, and “*blue*” is 3.

This is called a label encoding or an integer encoding and is easily reversible.

For some variables, this may be enough.

The integer values have a natural ordered relationship between each other and machine learning algorithms may be able to understand and harness this relationship.

For example, ordinal variables like the “place” example above would be a good example where a label encoding would be sufficient.

### 2. One-Hot Encoding

For categorical variables where no such ordinal relationship exists, the integer encoding is not enough.

In fact, using this encoding and allowing the model to assume a natural ordering between categories may result in poor performance or unexpected results (predictions halfway between categories).

In this case, a one-hot encoding can be applied to the integer representation. This is where the integer encoded variable is removed and a new binary variable is added for each unique integer value.

In the “*color*” variable example, there are 3 categories and therefore 3 binary variables are needed. A “1” value is placed in the binary variable for the color and “0” values for the other colors.

For example:

1 2 3 4 |
red, green, blue 1, 0, 0 0, 1, 0 0, 0, 1 |

The binary variables are often called “dummy variables” in other fields, such as statistics.

## Further Reading

- Categorical variable on Wikipedia
- Nominal category on Wikipedia
- Dummy variable on Wikipedia

## Summary

In this post, you discovered why categorical data often must be encoded when working with machine learning algorithms.

Specifically:

- That categorical data is defined as variables with a finite set of label values.
- That most machine learning algorithms require numerical input and output variables.
- That an integer and one hot encoding is used to convert categorical data to integer data.

Do you have any questions?

Post your questions to comments below and I will do my best to answer.

You didn’t mention that if we have a categorical variable with 3 categories, we only need to define 2 one-hot variables to save us from linear dependency.

HHi jason.I truly following you alot and really appreciate your effort and ease of tutorials.just a question,How one hot encoding would work for multilabel class and in coming tutorials could you help in featureselection of text data for muticlass and multilabel classification using keras.i tried multiclass for 90 datapoints. And used keras for mlp,cnn and rnn where each datapoint is long paragraph with labels but accuracy i got is 37.5 prcent. Let me know if you have any suggestions

The one hot vector would have a length that would equal the number of labels, but multiple 1 values could be specified.

Thanks for the suggestion.

This post suggests ways to lift deep learning model skill:

http://machinelearningmastery.com/improve-deep-learning-performance/

What are the cons of one hot encoding ??? Supposed that you have some categorical features with each one with 500 or more differents values !! So when you do one hot encoding you will have many colums in the dataset does it still good for a machine learning algorithm ???

Great question!

The vectors can get very large, e.g. the length of all words in your vocab in an NLP problem.

Large vectors make the method slow (increased computational complexity).

In these cases, a dense representation could be used, e.g. word embeddings in NLP.

Hi Jason, thanks again for your amazing pedagogy.

Back to the Espoirt question, I face this problem with 84 user_ID. I do a OHE of them and, like you said when I fit the data with a SVM classifier, it’s look like I fall in a infinite loop. So taking in to account the fact that I am not in the NLP case, how can I fixe this ?

Thanks.

What do you mean you fall into an infinite loop?

Very helpful post, Jason!

Espoir raised my question here but I did not undestand how to apply your answer to my case. I have 11+ thousand different products id. The database has about 130 thousand entries. This easily leads to MemoryError when using OHE. What approach/solution should I look for?

Ouch.

Maybe you can use efficient sparse vector representations to cut down on memory?

Maybe try exploring dense vector methods that are used in NLP. Maybe you can something like a word embedding and let the model (e.g. a neural net) learn the relationship between different input labels, if any.

Hello Jason how do we retrieve the features back after OHE if we need to present it visually?

You can reverse the encoding with an argmax() (e.g. numpy.argmax())

Thank you for these wonderful posts!

Does data have to be one-hot encoded for classification trees and random forests as well or they can handle data without it? Or just try which gives better results?

No, trees can deal with categories as-is.

Hi Jason, this post is very helpful, thank you!!

Question- In general what happens to model performance, when we apply One Hot Encoding to a ordinal feature? Would you suggest only to use integer encoding in case of ordinal features?

It really depends on the problem and the meaning of the feature being encoded.

If in doubt, test.

I see, thanks!

hey Jason,

As usual this is another useful post on feature representation of categorical variables. Since logistic regression fits a separation line on the data points of the form w1X1 + w2X2 +.. where X are features such as categorical variables- Places,color etc, and w are weights, intuitively X can take only numerical values for the line to fit. Is this a right intuition?

Yes, regression algorithms like logistic regression require numeric input variables.

Thanks a lot for your clarifying. I love your blogs and daily email digests. They help me to understand key concepts & practical tips easily.

Thanks Raj.

nice!

Thanks.

very well explained..thanks

Thanks, I’m glad it helped.

I love your blog!

One question: if we use tree based methods like decision tree, etc. Do we still need one-hot encoding?

Thanks you very much!

No Jie. Most decision trees can work with categorical inputs directly.

Thank you very much!

No probs.

Hi Jason, loving the blog … a lot!

I’m using your binary classification tutorial as a template (thanks!) for a retail sales data predictor. I’m basically trying to predict future hourly sales using product features and hourly weather forecasts, trained on historical sales and using above/below annual average sales as my binary labels.

I have encoded my categorical data and I get good accuracy when training my data (87%+), but this falls down (to 26%) when I try to predict using an unseen, and much smaller data set.

As far as I can see my problem is caused by encoding the categorical data – the same categories in my unseen set have different codes than in my model. Could this be the cause of my poor prediction performance: the encoded prediction categories are not aligned to those used to train and test the model? If so how do you overcome these challenges in practice?

Hope it makes sense.

Nice work Andrew!

Your model might be overfitting, try a smaller model, try regularization, try a large dataset, try less training.

Here are more ideas:

http://machinelearningmastery.com/improve-deep-learning-performance/

I hope that helps as a start.

Hey Jason, didn’t think I had ‘that’ problem, but I probably do 🙂

Many thanks.

Appreciable and very helpful post, thank you!!!

Question: What is the best way to one hot encode an array of categorical variables?

I have also startup with a AI post you can also find some knowledge over there: Thebigmoapproach.com/

There are many ways and “best” is defined by the tools and problem.

Here are a few ways:

http://machinelearningmastery.com/how-to-one-hot-encode-sequence-data-in-python/

hi Jason:

One question, take the “color” variable as an example,if the color is ‘red’ , then after one-hot encoding ,it becomes 1,0,0 . So,can we think that it generates three features from one feature?

It has been added two columns，is that right？

Correct Tom!

Hi Jason, if my input data is [1 red 3 4 5], if use one hot encoder, red become [1,0,0], ]does it mean that the whole features of the input data is extended?

input data now is [1 1 0 0 3 4 5]

Sorry, I don’t follow. Perhaps you can restate your question?

Hi Jason, Thank you for the reply.

For example, if I have 4 feature of my input, [121 4 red 10; 100 3 green 7; 110 8 blue 6]

For the first row, the value related to each feature–feature 1:121, feature 2:4, , feature: red, feature 4: 10.

I want to use one hot encoder now, red = [1,0,0], green = [0,1,0], blue = [0,0,1].

So my input become [121 4 1,0,0 10; 100 3 0,1,0] 7; 110 8 0,0,1 6] , after one hot encoding, we now have 6 features, so I use the new data for training, it that right?

Thanks.

Hello DR. Brownlee,

I am training a model to detect attacks and i need someone like you to help me detect the mistakes in my code because my training is not producing any better results. Kindly alert me if you will be interested to help me.

Thank you

Sorry, I do not have the capacity to review your code.

I am using Tensorflow developing the mode,and would want to know how your book can help me do that. since it makes reference to Keras. Thank you

My deep learning book shows how to bring deep learning to your projects using the Keras library. It does not cover tensorflow.

Keras is a library that runs on top of tensorflow and is much easier to use.

This is not entirely correct as far as I understand. As Varun mentioned, you need to have one less column (n-1 columns). What has been described is dummy-encoding (which is not one-hot-encoding). There is a major problem with dummy encoding which is perfect collinearity with the intercept value. As the sum of all the dummy values of one category (n columns) is ALWAYS equal to 1. So it’s basically an intercept

Other way around I believe. Dummy encoding is n-1 columns, one hot has n columns.

Hi Jason,I have 6 categorical values which are present in the data that I have. The data that I have has many missing categorical values that are left as empty strings. What to do if I have missing categorical values? Do I need to OHE them also? or how to deal with the categorial feature with missing values?

I’m using sci-kit learn and trying out many algorithms for my dataset.

I list some ways to handle missing data here:

https://machinelearningmastery.com/handle-missing-data-python/

Hi Jason,

First thank you for your post !

There is something i did not understand about your explanation : let’s take the color example (so red is 1, green is 2, blue is 3).

I did not understand the “ordinal relationship between catégories” : does the One-Hot-Encode allow better accuracy for some learning algorithms than these categories? (So far here’s what I thought: the algorithm reads 1,2 or 3 instead of red, green or blue, and makes the necessary correlations for predictions, and that has no impact on the predictions accuracy.)

Hmm. Sorry for not being clearer.

Ordinal means ordered. Some categories are naturally ordered and in these cases some algorithms may give better results by using just an integer encoding.

For problems where the categories are not ordered, the integer encoding may result in worse performance than one hot encoding because the algorithm may assume a false ordering based on the assigned numbers.

Does that help?

Bit confused. In the case we ordered integer labels correctly, do we need one hot encoding? Actually it is bit stupid that labeling impacts to acc. I thought one hot labeling is for simplicity but you say that in the case of integer label which is not well ordered.

Is there any reason why we use one hot encoding in the case we order integer labels correctly?

I accept what u explained why we need to use integer encoding instead of character labeling.

Thank you

I was saying that if your variable values are not ordinal and you treat them as ordinal when fitting the model (e.g. not use one hot encoding), you may loose skill.

Does that help?

I have data from 20 000 stores. Each store has it’s integer ID. This ID is meaningless, just ID. Should I add 20 000 binary variables to datatset? And 20 000 neurons in input layer of LTSM? It sounds frightening…

No, drop the id unless you have a hunch that it is predictive (e.g. numbering maps to geographical region and regions have similar outcomes).

Ok, I have latitude and longitude of each store. Should I use them instead of ID? Similar question. I have 17 states of weather (cloudy, rainy, etc.). Should I replace them with 17 binary variables? Or should I try to give integer code to them to show similarity of heavy rain to rain and light rain, sunny to partial clouds and heavy clouds?

There are no rules, I would encourage you to try many different framings and see what works best for your specific data.

I have some biases that I could suggest, but it would be better (your results would be better) if you use experiments to discover what works for your problem.

Yes, it’s right, thanks

Great post Jason! I’m glad I came across it. It really helped me to understand the need for one hot encoding. I’m new to machine learning and I am currently running xgboost in R for a classification problem.

I have 2 questions:

(1) If my target variable (the variable I want to predict) is categorical, should I also convert it into numeric form using hot encoding or will a simple label encoding suffice?

(2) Are there specific R packages for one hot encoding of features?

It really depends on the method. It can help.

Sorry, I don’t recall if you must encode variables for xgboost in R, it has been a long time.

Hello Jason

I have dataset having numeric and nominal type. It also has missing values. For nominal datatype, first I applied Labelencoder() to convert them into numeric values, but along with my two categories(normal, abnormal), it also assigns a code to NaN. In such scenario how can I impute values by its Mean?

You can impute with the mode in this case.

Hi Jason,

Since the number of columns created for a categorical column after applying OneHotEncoding is equal to the number of unique values in that categorical column; often it happens that the number of features in the tested model is not equal to the number of features on the dataset to be predicted after applying OHE similarly on the categorical fields. In such cases model throws an error while predicting since it expects equal number of features both in the training and to be predicted dataset. Can you please advise how to handle such situation ?

The same transform object used for training is then used for test or any other data. It can be saved to disk if need be.

Hi Jason,

I couldn’t get what do you mean by “same transform object”. ? The training dataset structure (number of initial features) is same both for Training and Testing/to-be-predict dataset. But the uniqueness of values under each feature/column may differ which is quite natural. Therefore OneHotEncoding or pandas get_dummies create different number of encoded features in Test/to-be-predict dataset than the training dataset. How to deal with this issue – that is what my question.

Need your advise please.

Thanks.

Sorry. To be clearer, you can train the transform objects on the training data and use them on the test set and other new data.

The transform objects may be the label encoder and the one hot encoder.

The training data should be such that it covers all possible labels for a given feature.

Does that help?

Hi Jason,

should I do one-hot encode for two level categorical variables? like variable only contains (yes. no) converts to two variable (0,1) and (1,0)

Thanks.

Generally, this is not needed.

For One-Hot Encoding (OHE) of a categorical variable State with 4 values: NJ, NY, PA, DE

We can remove one of them, say DE, to reduce complexity.

So if NJ=0, and NY=0, and PA=0, then it is DE

Is removing one recommended?

This becomes more obvious in the case of a binary categorical variable.

Thanks.

If you can simplify the data, then I would recommend doing that.

Always test the change on model skill though.

One Hot Encoding via pd.get_dummies() works when training a data set however this same approach does NOT work when predicting on a single data row using a saved trained model.

For example, if you have a ‘Sex’ in your train set then pd.get_dummies() will create two columns, one for ‘Male’ and one for ‘Female’. Once you save a model (say via pickle for example) and you want to predict based on a single row you can only have either ‘Male’ or ‘Female’ in the row and therefore pd.get_dummies() will only create one column. When this occurs the number of columns no longer matching the number of columns you trained your model on and errors out.

Do you know a solution to this issue? My actual need uses Zip Code rather than Sex which is more complex.

I recommend using LabelEncoder and OneHotEncoders from sklearn on a reasonable sample of your data (all cases covered) and then pickle the encoders for later use.

Thank you!!!

Hi Jason,

I am piggybacking on some of the other questions re: n-1 encoding and n encoding. I have a dataset where I predict price based on day of week using sklearn LinearRegression (also playing with Ridge). I used DictVectorizer in sklearn to prep my data and I end up with 7 columns for day of week, rather than 6. In some of the questions above, you indicate simpler is better…though you do say to “test the change on model skill.” Could you elaborate on that – for example, what are the practical implications of using one or the other for a dataset like mine (features = days of week; target = price)? My model seems to spit out a reasonable y-intercept, though I’m not sure exactly what the y-intercept is because my model has no [0, 0, 0, 0, 0, 0, 0] for day (i.e., no “reference” day).

Is there a mathematical reason to use n-1 vs n encoding? I hope this makes sense. I’ve Googled like 50 times and can’t find an article that really gets into this. Thank you.

If your goal is the best model skill, then use whatever works to improve that skill.

No need for idealized justifications.

Very helpful, thanks.

May I ask that if I have 4 possible letters in a string that I would like to encode (let’s say A B C D), what is better for neural networks? One-ho or integer encoding. The groups have no order, so I would say one-hot but I do not know whether neural network could deal with integer encoding in this case (it would mean a quarter of the features as one-hot encoding).

Thank you!

One hot if there is no ordinal relationship between the labels.

I am total noob to this so maybe a silly question but this can only be applied when categories are less in number and the problem is about classification right?

One hot encoding can be used on input features for any type of problem and on the output feature for classification problems.

Hello jason i have question: If there are categorical variable like 1st class, 2nd class, 3rd class fror housing price prediction , if i am converting with OneHotcoding so how algorithm will judge the ranking part of housing ? Yes it does convert it into binary but does it also taking car of ranking of that categorical variable? & One more question is to get binary output “pd.get_dummy” is useful or One HotEncoder is useful ?

A one hot encoding is for a classification problem, not regression.

The house price is a regression problem where we predict a quantity.

Hi Jason..I have a data with string values having more than 500 unique values. How can I encode it so that can pass it to Ml algorithm. Is this good candidate for categorical encoding?

That is a lot. I would recommend NLP representations such as bag of words or word embeddings.

I have posts on both, start here:

https://machinelearningmastery.com/start-here/#nlp

Hi Jason,

This is a great post, thanks for providing such valuable info. My question is:

If we have many colors in the color column, say 25 colors, what if we encode the colors in 3 columns with RGB values instead of 25 binary columns? Do you see any abnormality with this approach?

No problem, it would be a binary vector with 25 elements.

Hi Jason.

I am an intern in data science, no exprience in datas. Thanks for your posts and e-mails that boosted my confidence to start my intern on ML and Deep learning.

currentley, i got a dateset of 10 GB and after i make a preliminary investagation on the data i found out the following.

feature ‘x1’ has 78 unique categories

feature ‘x2’ has 24 unique categories

feature ‘x3’ has 24 unique categories

feature ‘x4’ has 35 unique categories

feature ‘x5’ has 40 unique categories

feature ‘x6’ has 106 unique categories

feature ‘x7’ has 285629 unique categories

feature ‘x8’ has 523912 unique categories

feature ‘x8’ has 27 unique categories

feature .x9’ has 224 unique categories

feature ‘x10’ has 108 unique categories

feature ‘x11’ has 98 unique categories

feature ‘x12’ has 10 unique categories

feature ‘x13’ has 1508604 unique categories

feature ‘x14’ has 15 unique categories

feature ‘x15’ has 1323136 unique categories

feature ‘x16’ has 3446828 unique categories feature ‘x17’ has 10 unique categories

feature ‘x18’ has 200 unique categories

feature ‘x19’ has 2575092 unique categories

feature ‘x20’ has 197957 unique categories

how you you deal with this data set….. it has categorical and int attributes. it a classification problem. just predict the out come either lets say 0 nor 1. how you handle the category or how would you encode this attributes.

should i simply use label encoder , one hot encoder or dummies. is it possible to encode such a big categories after all?

i am confused where to start.

Looking forwards for your suggestions and help

That is a lot of categories.

Perhaps you can remove some features?

Perhaps you can consolidate the categories for each feature?

You can get started with feature selection here:

http://machinelearningmastery.com/an-introduction-to-feature-selection/

The article boils down to one sentence:

“using this encoding and allowing the model to assume a natural ordering between categories may result in poor performance or unexpected results (predictions halfway between categories)”

And that’s enough said. Thanks.

Not quite.

That applies to the integer encoding, not the one hot encoding.

In fact, that is the problem that the one hot encoding will over come.

I disagree with one hot encoding approach. I mean, it depends on the algorithm. My opinion is based on playing around with categorical data and various algorithms on many Kaggle competitions with real world data.

For example, LightGBM can offer a good accuracy when using native categorical features. Not like simply one-hot coding, LightGBM can find the optimal split of categorical features. Such an optimal split can provide the much better accuracy than one-hot coding solution. (official documentation: http://lightgbm.readthedocs.io/en/latest/Advanced-Topics.html)

PS. Compared to other GBMs (native gbm, h2o gbm or even xgboost), lightgbm is far ahead in terms of speed and accuracy.

Thanks for the note Pranav.