How to Prepare Data For Machine Learning

Machine learning algorithms learn from data. It is critical that you feed them the right data for the problem you want to solve. Even if you have good data, you need to make sure that it is in a useful scale, format and even that meaningful features are included.

In this post you will learn how to prepare data for a machine learning algorithm. This is a big topic and you will cover the essentials.

lots of data

Lots of Data
Photo attributed to cibomahto, some rights reserved

Data Preparation Process

The more disciplined you are in your handling of data, the more consistent and better results you are like likely to achieve. The process for getting data ready for a machine learning algorithm can be summarized in three steps:

  • Step 1: Select Data
  • Step 2: Preprocess Data
  • Step 3: Transform Data

You can follow this process in a linear manner, but it is very likely to be iterative with many loops.

Step 1: Select Data

This step is concerned with selecting the subset of all available data that you will be working with. There is always a strong desire for including all data that is available, that the maxim “more is better” will hold. This may or may not be true.

You need to consider what data you actually need to address the question or problem you are working on. Make some assumptions about the data you require and be careful to record those assumptions so that you can test them later if needed.

Below are some questions to help you think through this process:

  • What is the extent of the data you have available? For example through time, database tables, connected systems. Ensure you have a clear picture of everything that you can use.
  • What data is not available that you wish you had available? For example data that is not recorded or cannot be recorded. You may be able to derive or simulate this data.
  • What data don’t you need to address the problem? Excluding data is almost always easier than including data. Note down which data you excluded and why.

It is only in small problems, like competition or toy datasets where the data has already been selected for you.

Step 2: Preprocess Data

After you have selected the data, you need to consider how you are going to use the data. This preprocessing step is about getting the selected data into a form that you can work.

Three common data preprocessing steps are formatting, cleaning and sampling:

  • Formatting: The data you have selected may not be in a format that is suitable for you to work with. The data may be in a relational database and you would like it in a flat file, or the data may be in a proprietary file format and you would like it in a relational database or a text file.
  • Cleaning: Cleaning data is the removal or fixing of missing data. There may be data instances that are incomplete and do not carry the data you believe you need to address the problem. These instances may need to be removed. Additionally, there may be sensitive information in some of the attributes and these attributes may need to be anonymized or removed from the data entirely.
  • Sampling: There may be far more selected data available than you need to work with. More data can result in much longer running times for algorithms and larger computational and memory requirements. You can take a smaller representative sample of the selected data that may be much faster for exploring and prototyping solutions before considering the whole dataset.

It is very likely that the machine learning tools you use on the data will influence the preprocessing you will be required to perform. You will likely revisit this step.

So much data

So much data
Photo attributed to Marc_Smith, some rights reserved

Step 3: Transform Data

The final step is to transform the process data. The specific algorithm you are working with and the knowledge of the problem domain will influence this step and you will very likely have to revisit different transformations of your preprocessed data as you work on your problem.

Three common data transformations are scaling, attribute decompositions and attribute aggregations. This step is also referred to as feature engineering.

  • Scaling: The preprocessed data may contain attributes with a mixtures of scales for various quantities such as dollars, kilograms and sales volume. Many machine learning methods like data attributes to have the same scale such as between 0 and 1 for the smallest and largest value for a given feature. Consider any feature scaling you may need to perform.
  • Decomposition: There may be features that represent a complex concept that may be more useful to a machine learning method when split into the constituent parts. An example is a date that may have day and time components that in turn could be split out further. Perhaps only the hour of day is relevant to the problem being solved. consider what feature decompositions you can perform.
  • Aggregation: There may be features that can be aggregated into a single feature that would be more meaningful to the problem you are trying to solve. For example, there may be a data instances for each time a customer logged into a system that could be aggregated into a count for the number of logins allowing the additional instances to be discarded. Consider what type of feature aggregations could perform.

You can spend a lot of time engineering features from your data and it can be very beneficial to the performance of an algorithm. Start small and build on the skills you learn.

Summary

In this post you learned the essence of data preparation for machine learning. You discovered a three step framework for data preparation and tactics in each step:

  • Step 1: Data Selection Consider what data is available, what data is missing and what data can be removed.
  • Step 2: Data Preprocessing Organize your selected data by formatting, cleaning and sampling from it.
  • Step 3: Data Transformation Transform preprocessed data ready for machine learning by engineering features using scaling, attribute decomposition and attribute aggregation.

Data preparation is a large subject that can involve a lot of iterations, exploration and analysis. Getting good at data preparation will make you a master at machine learning. For now, just consider the questions raised in this post when preparing data and always be looking for clearer ways of representing the problem you are trying to solve.

Resources

If you are looking to dive deeper into this subject, you can learn more in the resources below.

Do you have some data preparation process tips and tricks. Please leave a comment and share your experiences.

78 Responses to How to Prepare Data For Machine Learning

  1. Fraser March 31, 2014 at 4:30 am #

    I enjoyed your concise overview, Jason.

    Perhaps you can delve a little into the dangers/opportunities in your Step 2: Cleaning stage.

    It has been my experience that those data you may want to remove contain the more interesting data to the client (perhaps only after the requested client questions are addressed).

    Fraser

    • jasonb March 31, 2014 at 5:36 am #

      Hi Fraser, good question.
      Indeed, it can difficult to know if data is bad and you may not always have a domain expert at hand to comment. Sometimes it is obvious though, like 0 values that are impossible in the domain like a blood pressure. I’ve also seen -999 used to signal “not provided”. In these cases we can mark attributes as missing and think about possible rules for imputing if we so desire.
      Where do you draw the line though? Should severe outliers be marked as missing? Sometimes. I like to try a lot of stuff, for example, I would try removing instances with large outliers in one dimension and see what that did to my models, I’d also try removing instances with missing values and try models on variations of the data with imputed value. Almost always, modeling ground truth is not the goal, there are performance metrics like classification accuracy or AUC that we are being optimized.
      You’re right though, sometimes the broken data can represent something very interesting – anomalies that signal something useful in and of themselves in the domain.

  2. Fraser March 31, 2014 at 6:05 am #

    Yes, indeed. Is it an outlier, or a poorly encoded result, or a result with atypical calibration, or does it represent a distinct and real combination of natural conditions …

    I work a lot with chemical concentration data in water and sediment and I run into censored data routinely. Mostly from the 1000 mg/L. Censored data of this particular type is handled differently by different people and as you suggest values need to be imputed (with an appropriate sampling distribution) if the rest of a multi-parameter time-sample result is to remain in the analysis.

    For me this is what makes data analysis fun.

    I just arrived at your site, and I see so many articles of interest. Thank you for making this available.

    Fraser

  3. Fraser March 31, 2014 at 6:07 am #

    The use of the angle brackets got lost in my post above.

    “Mostly of the type “less than” .01 pg/L but occasionally the other side, say “greater than” 1000 mg/L.”

    • jasonb March 31, 2014 at 7:42 am #

      Insightful comments Fraser, thanks. Reach out any time if you want kick around some ideas on a tough problem.

  4. Fraser March 31, 2014 at 11:50 am #

    Thanks, Jason. I will do that. Fraser

  5. Surajit August 25, 2015 at 10:23 pm #

    I like “Getting good at data preparation will make you a master at machine learning”. This is indeed a good post.

    Thanks Dr Jason.

    • Rohita Gupta February 10, 2017 at 5:43 pm #

      Can you please share the link of this article

      • Jason Brownlee February 11, 2017 at 4:55 am #

        I believe Surajit was quoting from this article.

  6. Kiran Garimella November 6, 2015 at 9:20 am #

    Great set of articles!

    One issue that I run into is that the data sometimes lacks semantic integrity. This is not an issue of missing values, but just having improper values. When values are of different data types within a column, it is easy to detect and fix.

    However, when the data type is the same but the meaning changes, then it’s much more difficult. For example, I’ve seen sales data where a column named ‘marketing plan code’ would have string data type denoting marketing plan codes, except in a few cases where the users put in vendor codes because they didn’t have any other field to record that information.

    Any insights and anecdotes about this issue?

  7. KLeyn May 4, 2016 at 10:14 pm #

    Jason, does it affect an algoritm if, during the preparation process I transform the list of rows (like tables, where the key columns repeats) to pivot tablee, where the key colums shows once and a lot of columns (say hundreds) have parcial sums or counts for the different conditions (let’s say sales of january in one column, sales of february in a second column and so on).
    Does it would make muiltcorrelation as some columns could be aggregated to one?

  8. mokhtar May 12, 2016 at 4:44 am #

    Thank you for your valuable information for this important area Machine learning, started with data structure and going further to build it complete.

  9. ali October 21, 2016 at 9:50 am #

    how can i make one attribute as decision attribute in the data set in order to classification model depend on the selected attribute

    • Jason Brownlee October 22, 2016 at 6:54 am #

      Hi Ali,

      Different algorithms will chose which variables to use and how to use them. You can force a model to use one variable by deleting all of the other variables.

  10. Avin October 31, 2016 at 11:21 am #

    Hi Jason,

    Appreciate the effort you put into the great article.

    I am currently working on a project on a government data set to find if an entity(person or an individual) were involved in a a positive or a negative way. I took a flat file containing some test data and prepared the code to perform sentiment analysis using Naive Bayes algorithm using NLTK python modules.

    – In most cases we have a defined trained data set tagged as ‘positive’ or ‘negative’ (e.g movie reviews, twitter data set). In my case there is no existing trained government data set.
    – The training data is available but I need to categorize the training data set as ‘positive’ or ‘negative’.
    – My question here is, how do we go about classifying my government data as ‘positive’ or ‘negative’.

    I’m looking forward on your advice on how to categorize my government training data as positive or negative. This is very important for me to get my sentiment analysis with best possible accuracy.

    • Jason Brownlee November 1, 2016 at 7:58 am #

      Hi Avin, I would advise you locate a subject matter expert to prepare a high-quality training dataset for you (manual classifications).

  11. Mayur November 4, 2016 at 6:01 am #

    What is the best way to process large amount of data for machine learning?

    • Jason Brownlee November 4, 2016 at 11:14 am #

      Hi Mayur,

      That depends on the problem and how the data is currently represented and stored. No silver bullets, sorry.

  12. Ivan November 9, 2016 at 12:27 am #

    My current and first ML project has natural language as it’s input and I spent a huge chunk of time on preparing it.

    I stopped once the data reached a “reasonable” level so that I could continue with the project, i.e. I’m dropping the hard to parse cases and might return to them later once the whole pipeline is ready for testing.

    Keeping the 80/20 rule in mind.

  13. ted January 3, 2017 at 3:14 am #

    Thank you for your valuables posts, my question is how to apply machine learning to Cancer Registry data set?

    I have two datasets:

    1. Dataset1 :

    About 18K observation and 22 variables: the five years data set includes
    Demographic, Diagnoses, and treatments,

    2. Dataset2:
    aggregate vitals based on race grouping on: regions, stages, vitals

    Thank you for your help

    Ted

  14. José Alberto Ramos Silva March 11, 2017 at 11:46 pm #

    Hi Jason, thank you for the great effort and knowledge put into all these posts!
    My question will probably be silly, but since I’m a complete n00b I’ll do it just the same.
    Data prep, feature analysis and engineering will get you a set of data in a format completely different from original data. These data transformation steps may be very hard to do automatically. My problem is related to classification, I am using NN which may not be the best choice, but hey, humor me 😉
    So, cutting short. Originally, I get raw data, I prep and transform it. The transformed data will train and test “my” NN. Now, the “real world” will challenge my model with raw data, presumably with the same format as my original training set, minus the classification ( of course…). Now, I suppose I’ll have to go through the same data transformation of the data before the trained model can be fed with it. Right? Doesn’t this mean extra care must be taken to make the data transformation process (at least ideally) automatic itself?
    Sorry for the long question, hope to hear your thoughts on these points. And thank you once again!

    • Jason Brownlee March 12, 2017 at 8:27 am #

      Very good question José!

      Yes. Any data transformation performed on data used to fit your model must be performed on data when making predictions.

      This means we need a very clear recipe for this transform, ideally automatic and also in the case of regression problems it must be reversible so that we can convert predictions back into their predictions scale for use.

      • José Alberto Ramos Silva March 16, 2017 at 9:47 am #

        Thank you very much Jason. And keep up the excellent job you are doing!

  15. dhanpal singh April 29, 2017 at 7:51 am #

    what is the best book to learn how to prepare the datasets for machine learning models

  16. Eric Kraemer August 12, 2017 at 7:19 am #

    I would like to offer that within your topic of “Select Data” you offer a bit more explicit guidance on the topic of assessing and characterizing data quality. It’s cliché, but garbage-in-garbage-out is a fundamental concept. I so often come across advanced analytic initiatives that have started out with Assumptions for quality of “selected” data and moved on – only to find out months later that everything has to reset to basic principles of data acquisition and management.

    What transforms have been applies to source data by systems that precede the database you are selecting from?

    If sensor data is involved, what formatting, precision, transformations, signal processing, etc. have been applied?

    If data is being acquired from multiple, disparate systems what formatting, scale, and precision differences are being masked by the database system you are selecting from?

    Just a few examples.

    • Jason Brownlee August 13, 2017 at 9:43 am #

      Really good points Eric.

      It’s hard to give general advice on data prep because of all of the detail in specific data matters.

      It’s not like algorithms where you can say “try everything and see what works on your data”.

  17. Rahul Shukla September 28, 2017 at 2:37 pm #

    how we save our preprocessed data into database and how we train model using this data.

  18. Aniket Saxena October 20, 2017 at 3:17 pm #

    Hi Jason,

    when I go through UCI Machine Learning Repository following doubts have occured:

    1. in bike sharing dataset, I saw two .csv files(one is day.csv and another is hour.csv). So,i can’t understand how to make this dataset suitable for me to apply machine learning algorithm on it to make predictive model by splitting the whole dataset into train and test sets?

    2. in this repository, I saw dataset characteristics as multivariate and univariate, what does this mean?

    3. in this repository, whenever I explore any of the dataset, there is no statement present there to mention which is the feature to be predict by applying machine learning algorithms?

    4. what if both numeric types(float as well as integer) values in any of the feature exist in a dataset? Should we scale the feature values(integer) to float in order to get good predictive model?

    Please help…….

    • Jason Brownlee October 21, 2017 at 5:25 am #

      Each dataset is different. You will need to take care and discover how to prepare each one.

      Univariate means one variable, feature or column (all the same thing), multivariate means many.

      You might have to check the data or read the associated paper.

      Depending on the algorithm used, you might need to convert all features to numeric.

  19. Aniket Saxena October 21, 2017 at 1:41 pm #

    So, this means that we have to convert the integer values of all feature exist in a dataset into float values in order to increase the accuracy of our model? Correct me if I am wrong.

    What do you think if there are two .csv files in a dataset, how should we prepare this type of dataset? Please recommend me a way to do this.

    Thanks…

    • Jason Brownlee October 22, 2017 at 5:15 am #

      Perhaps, it depends on the algorithms being used. I would recommend trying it.

      If there are two files, I would recommend combining the data into one file.

  20. Aniket Saxena October 22, 2017 at 2:37 pm #

    Hello Jason,

    When I saw bike sharing dataset in UCI Machine Learning Dataset, UCI mention it’s dataset characteristics as univariate despite having total of 16 features(columns). Why is this so? Shouldn’t it be multivariate, instead.

    Secondly, as you have recommended to join two .csv files into one, but when I use this dataset, I noticed that both of the files have same features(except hr(hour) available only in hour.csv file not in day.csv file) with different values in each of the same features available in both the files. In this kind of situation, if I join both the files, values get redundant and even features as well. So, what do you recommend, how to prepare my dataset in this type of situation?
    Thanks for your quick response to previous question….

    • Jason Brownlee October 23, 2017 at 5:39 am #

      Perhaps they define univariate in terms of output variable only.

      Sorry, perhaps I don’t have enough information to give you good advice on how to prepare your data.

  21. Aniket Saxena October 25, 2017 at 3:55 am #

    Thanks for your help on this topic but please whenever you will come to know about how to prepare this type of dataset, tell me or recommend me as well at that moment of time.

    Thank you so much for guiding me how to prepare any dataset by creating this amazing post.

  22. Gene November 4, 2017 at 7:59 pm #

    What happens if I use a data that does not have a normal distribution?
    Are some ML algorithms only suitable for data that are assumed to be normal?
    How can I identify whether an algorithm works with normal/non-normal or just normal data?

    • Jason Brownlee November 5, 2017 at 5:16 am #

      In practice, you can often get good results by breaking these rules.

      I would recommend testing a suite of algorithms on your data and double down on what seems to be working.

  23. Krish December 12, 2017 at 1:48 am #

    Thanks for your help. Can you please suggest me what is the best way to deal with a dataset that contains a lot of text columns. Also the values of these columns too have a huge set of different values.

  24. Everton January 10, 2018 at 12:20 am #

    Hi Dr. Jason,

    Thank you for your work, I really appreciate your efforts in helping us.
    I am a BIG fan.

    First off all, I’m planning to use a LSMT-RNN in multivariate time series problem.

    I’m beginning my studies in machine learning and probably my question is very silly, but to me is a big issue.

    I have a time series database with 221 features not supervisioned yet, wich I would to transform to a input with to 6 up to 10 features. After this, I would like to supervise the output up to 10 time-steps with 1 feature.

    I had preprocessing my database by: cleaning, detrending, normalizing, correlating and clustering by affinty. I got 27 clusters from my 221 features.

    My question is:

    Now I think I can choose my input data, but how? Should I pick from the same cluster that my output have affinity with output, or should I pick from other clusters that don’t have affinty with my output?

    Thx for your time, sry about the big text.

    • Jason Brownlee January 10, 2018 at 5:28 am #

      Perhaps try a few methods and see which is easier to model.

  25. Everton January 11, 2018 at 1:54 am #

    Sorry, but I didn’t get the answer.

  26. Jesús Martínez February 8, 2018 at 4:51 am #

    Good article, Jason.

    Another data processing technique that is commonly used today, particularly in computer vision, is data augmentation where basically we introduce small changes such as rotations, coloring, and translations to images in order to emulate different conditions.

  27. Surya Gupta February 11, 2018 at 5:05 pm #

    hello,
    Actually, I am new toML, I want to know that when we apply data preprocessing on a dataset, whether we have to change the existing dataset or we have to create a new dataset for the modified data? Means after preprocessing is done will we be having two datasets, one the actual dataset and the other preprocessed dataset or there will be only one dataset with preprocessed data?

    • Jason Brownlee February 12, 2018 at 8:27 am #

      Create and save a new dataset or views on your raw data.

  28. chini February 15, 2018 at 5:43 pm #

    hi sir actually i want to prepare a data set for speaker recognition project for that i would like to prepare audio recorded data will you please mention the best procedure for that.

    • Jason Brownlee February 16, 2018 at 8:33 am #

      Sorry, I don’t have material on preparing audio data. I hope to cover it in the future.

  29. vikash February 19, 2018 at 4:41 pm #

    Hi ,
    I am vikash I want to know about the assumptions means about the pre-validation and post validation of data.for example for linear regression we have pre-validation or diagnosis like
    1.normal distribution of data
    2.No multicollinearity
    3.Linear relationship
    and
    4.Missing values
    for Post validation or diagnosis after creating the linear regression mode there are
    1.Normality of errors
    2.Homoschedasticity
    3ouliers and levrges
    5auto corelation.
    these are the assumptions for Linear Regression .What about the rest of algorithms assumptions ? can you guide me the assumptions for other algorithms .

    Thank you.

    • Jason Brownlee February 21, 2018 at 6:23 am #

      Often you can get good results or even better results if you ignore these type of assumptions. The reason is that is in predictive modeling, model skill is more important than theoretical correctness.

  30. gayathri April 6, 2018 at 2:08 pm #

    I have a CSV file timestamp, hostname, metric (CPU,MEM,PAGESCAN), metric vaule (0.7). I need to find the increase in metric value due to cpu or mem or pagescan.

    If CPU is increased then which host is maximum utilizing CPU like that finding the root cause.

    The data set contains both categorical values and numerical values. Do I want to convert the categorical data like hostname and metric value to numerical.

    Do i need to do data transformation?

    What machine learning techniques will well predict the root cause ? which algol.

    I am trying to use spark ML.

    Any suggestions.

    Thanks

    • Jason Brownlee April 6, 2018 at 3:51 pm #

      Yes, I would recommend converting categorical data to integer or even one hot encoding prior to modeling.

      I would recommend testing a suite of methods on your data to see what works best. Then double down on that.

      Sorry, I don’t have examples for Spark.

  31. Ernest April 16, 2018 at 2:33 am #

    Hello,
    I’m new in Machine Learning so I have a question. Input data have to be the same size?
    I mean, I have 10 matrixs with data, but matrixs have size for example [60, 120], [60, 460], [60, 340] and so on. I want to use Tensorflow Engine.
    I would be grateful if you could answer my question.
    Regards!

    • Jason Brownlee April 16, 2018 at 6:11 am #

      Yes, generally data must have the same shape.

      • Adama April 17, 2018 at 1:47 pm #

        Bonjour Jason,
        moi je sollicite votre soutient documentaire par rapport à mon projet “Techniques pour la préparation des données pour des projets de science de données”. j’ai lu les différents commentaires, mais mon projet en demande plus d’avantages sur les différentes et méthodes. on me demande de:
        Faire un état de l’art des techniques et outils pour la préparation des données et regrouper les approches en fonction des méthodes et techniques utilisées.
        Faire une synthèse des avantages et inconvénients des méthodes les plus pertinentes de l’état de l’art et proposer un processus pour la préparation des données.
        je suis vraiment dans le besoin d’orientation et de documentation. Merci

        • Jason Brownlee April 17, 2018 at 2:57 pm #

          Hi Adama, I think if you’re having trouble with your homework project that my best advice is to talk your professor and teaching staff. You are paying them to teach you.

          Data preparation is really specific to aa given type of data and predictive modeling problem. Perhaps you can focus your attention more to make the project easier.

  32. Cheyne Ravenscroft April 18, 2018 at 11:53 pm #

    Hi Jason,

    i really like the site and there is a lot of really useful things here, i’m presented at the moment with a problem.

    I’m attempting to classify a number of scanned PDFs based on the machine read text within them, i’ve got to the point where i have a relatively large test set.

    The documents themselves have extremely predictable sentences which tie in very closely with the classification however all i’ve managed to really find on this is using the BoW model.

    Would using a neural net to achieve this be a viable option? Also i’m having some problems with the pre-processing of the data. I’m not 100% on the best way to remove ‘\n’ characters and other punctuation from the large text strings.

    any help or pointers would be greatly appreciated.

    Many thanks,

    Cheyne

    • Jason Brownlee April 19, 2018 at 6:34 am #

      It is hard for me to tell. I would generally recommend testing a suite of methods to see what works best for your specific data.

      Let me know how you go.

  33. zino April 26, 2018 at 7:43 am #

    hi Jason , i like a lot your way to explain machine learning.
    i am working on combining machine learning techniques , and my question : there is ML problems where there are enough datasets to validate my work.

  34. Nil May 3, 2018 at 3:01 am #

    Hi DR Jason,

    It is a very good guide.

    I have a question, I am writing a neural network from scratch (back propagation algorithm) using sigmoid function so I have scaled my data in a range between -1 and 1 ]-1,1[ but sigmoid function give results between 0 and 1. So I would like to know if I must scale my data in a range between 0 and 1 [0,1] for sigmoid function?. Or would DR Jason please make me clear if there is a recommended scale of data when using a sigmoid function? or what is the recommended scale for sigmoid function?

    Best Regards.

  35. Jeremy May 19, 2018 at 10:49 am #

    Hi Dear Jason,
    Thanks four this overview. I would like to know in which format I should prepare my data for Non-dominated Sorting Genetic Algorithm 2 MATLAB. Thanks!

    • Jason Brownlee May 20, 2018 at 6:34 am #

      I believe that is an optimization algorithm, not a supervised learning algorithm. I don’t know what you mean exactly?

  36. Ekaterina June 1, 2018 at 2:41 am #

    Hi Jason, I am working with a dataset that has a lot of similar data items (e.g. mobile phone data). So, I would like to do the diversity-based sampling. What is the best way and tools to do it?

    • Jason Brownlee June 1, 2018 at 8:25 am #

      Perhaps clustering and filter based on distance to cluster centroids?

      Perhaps check the literature?

  37. Ansh July 23, 2018 at 4:34 am #

    Hi Jason,

    I’m trying to create a classification LSTM model. I have three categorical variables apart from my predictor variable. I have label encoded all the three variables. Do i need to scale the variables or I could use them as is .

Leave a Reply