data decision making
Data Science project, rblog

Designing data driven decision making; Kaggle ColeRidge

There is an interesting challenge running on Kaggle at the moment. It has been designed in cooperation with the Coleridge Initiative ( . This initiative is established at the New York University, it’s goal is to facilitate data driven decision making by governments. In the challenge we get to optimize automated dataset detection In policy papers. The ultimate goal? We need to help civil servants make better and more efficient decisions.

What is the key to making better decisions? Use the most up to date scientific research and data. Data Science can help here to make this process more efficient. Specifically we can use Natural Language Processing (NLP) techniques to predict what papers use a dataset. This is the essence of the Kaggle challenge. If it succeeds, it can facilitate government decision making transparency.

In this initial blog I want to discuss some of my considerations when thinking about this challenge. There will also be some quick analysis on the dataset provided.

Challenge overview

So you are working on a problem, how many times have you looked for a good scientific source to find a solution? Put yourself in the position of someone working for a local government. You have to come up with a local policy impact analysis on the councils climate change efforts. You’ve been looking for evidence and scientific research all morning. Isn’t there a tool that immediately identifies all similar research where similar datasets are used?

I can imagine these decisions are made on a daily basis, all around the world. Designing better systems to navigate this huge information source is the key to data driven decision making. To improve such mass scale decision making, with direct societal impact. It triggers my imagination.

Next the dataset, we are given roughly 20.000 papers, and have to predict what datasets are used in these documents. There are 3 key technical aspects I want to discuss, that in my opinion are vital to reaching a solution.

  • Feature addition
  • Similarity index of scientific papers
  • Transfer learning

Feature addition

The dataset literally consists of 5 columns. A document ID, the publication title, the dataset used in the publication, the institute associated with it and a cleaned label for the dataset. Firstly one of the potential challenges with text data is to find reliable additional structured data from the raw text provided. We want to see if there are some other features available!

There are multiple reasons for attempting to find structured additional data from the provided text. If we for example can extract the date of the publication from the text, we can see if this date of publication is a relevant factor in the usage of different datasets. one of the most important reasons is that it makes our ultimate solution easier to understand for policy users. Hence it creates a simpler interpretation of our outcome if we can use these structured elements. In our efforts to improve decision making it is our duty as the data scientist to maximize understanding of our solution.

From a data science perspective, creating additional metadata also helps us improve our own understanding of the problem. In one of the upcoming blogs I will attempt to use the text to create these metadata elements. From the top of my head important elements are:

  • Publication title
  • Author / Organization
  • Total citations
  • Citated papers
  • Text length

There is probably more potential for data enrichment, but this is a starting consideration for me when going over this challenge.

Similarity index of scientific papers

When we look for different papers that cover similar datasets, we assume that there is some similarity. There are similar technical terms used and introduced, similar domain knowledge. It is logical that a dataset that covers economic indicators, has similar explanations of these indicators. A dataset has a theoretical limit to the angles it can be discussed in.

Aside from the challenge of predicting datasets used in the articles, it should be an interesting business case to cluster the scientific papers, and find similar articles. Are there for example very different articles that cover the same dataset? By zeroing in on these effects we can more optimally cover all the relevant knowledge on a given dataset. It also just interest me a lot this question, and i’m curious if there are very different articles that cover similar data.

One of the ways to approach this simply is to use a string matching alghorithm. Using this method is an interesting baseline approach for the challenge. When I start a challenge I am always looking for a baseline. What I find most helpful in establishing a simple baseline early is clarity on the challenge complexity. For me it answers the question, how difficult is this really? When we create a simple intuitive model it helps understand how humans work the problem in practice. Finally It will further help to understand the mountain we have to climb with our more advanced predictive model.

Transfer learning

When I read this challenge part of what peaked my interest is the potential to use transfer learning. In the field of Natural Language Processing there has been a big buzz around transfer learning. For a scientific breakdown see here , another great link is found here . One of the most promising recent models is BERT.

Personally I’ve never used transfer learning and am eager to give it a try in this challenge. For me It would be a great victory to beat my string matching baseline with a transfer learning method. This will take some studying on the links mentioned in the previous paragraph, but there is alot of great content available.

Improving data driven decision making

The purpose of this blog is to write down initial thoughts. I will start off in the challenge by tackling these aspects. It helps me create a structure for my project. Finally it clears the path for the focus of the challenge.

We are focusing on improving data driven decision making for governments. Creating a good prediction model is one part of this, and that is the focus of the Kaggle challenge. But working on this I also want to broaden the impact. Ultimately the prediction model needs to be explainable and usable to achieve impact. Hence keeping this in the back of our minds as we design it is vital.

To conclude I will leave you with a word cloud of the papers that mention the datasets “World Ocean Database” and “Census of Agriculture”. I expect they look sufficiently different to be of interest. The code used for creating them is below the pictures.

Ocean vs Agriculture dataset wordcloud
# Library

# load and process data

data <- read.csv("C:/Startup/ColeRidge/train.csv")

files_train <- list.files('C:/Startup/ColeRidge/train', full.names = T, recursive = FALSE)
files_id <- gsub('.json','',basename(files_train))

load_json <- function(file, files_id){
  output             <- jsonlite::fromJSON(file)
  output$publication <- files_id

  json_files <- read_feather('C:/Startup/ColeRidge/json_files.feather')
} else {
  json_files <- lapply(1:length(files_train), function(x){load_json(files_train[x], files_id[x])})
  json_files <- data.table::rbindlist(json_files)
  write_feather(json_files, path = 'C:/Startup/ColeRidge/json_files.feather')

data <- merge(data, json_files, by.x = 'Id', by.y = 'publication', all.x = T, all.y = F)

world_ocean <- data[which(data$dataset_title == 'World Ocean Database' & data$section_title =='Introduction'),]
agri_census <- data[which(data$dataset_title == 'Census of Agriculture' & data$section_title =='Introduction'),]

build_wordcloud <- function(text){
  data_corp <- Corpus(VectorSource(text))
  # Cleaning:
  data_corp <- tm_map(data_corp, content_transformer(tolower))
  data_corp <- tm_map(data_corp, removeWords, stopwords("english"))
  data_corp <- tm_map(data_corp, removePunctuation)
  # Document term matrix:
  data_corp <- TermDocumentMatrix(data_corp)
  data_corp <- as.matrix(data_corp)
  data_corp <- sort(rowSums(data_corp),decreasing=TRUE)
  data_corp <- data.frame(word = names(data_corp),freq=data_corp)
  wordcloud(words = data_corp$word, freq = data_corp$freq, min.freq = 1,
            max.words=200, random.order=FALSE, rot.per=0.35, 
            colors=brewer.pal(8, "Dark2"))


code review example R code Caret
PSA, rblog

Code Review Example R Code Caret

Today we will go through a practical code review example. I will analyze some code chunks I used during the Kaggle ACEA challenge. The code chunks come from a function I used to run my final model using Caret in R. Code review is a very important step in any project, I recommend doing it after any project. It helps condense your code and make it more efficient. In doing so it also allows you to share and preserve your work more efficiently. In this blog I will follow the top 3 of my checklist I shared earlier.

Obviously I am reviewing my own code in these examples, this is different from reviewing someone elses code. The simplest difference is that your not the original author, I already know what the code is supposed to achieve. In practical examples the first step is actually making sure you understand how the code works. After all how can you review something without understanding it?

If you want some background on the code, I used this function during the ACEA Smart Water Analytics Challenge on Kaggle. Its function is discussed in more detail here. In short, this function was designed to run the final model I ended up using for this challenge on all available datasets.

You can find the full unedited code in this github, named ‘modelling_function.R’. I will now go through some code chunks. This code hasn’t been reviewed or optimized since I did the project, and there is alot of room for improvement.

Code Review Example Function Headers

The code chunk below shows the header of my helper function wrapper.

model_handmade_folds <- function(data, horizon, dataset, lag, features, basefeatures){
  #basefeatures <- 'Depth'
  # Make lags:
  features <- grep(features,names(data),value = T)
  basefeatures <- grep(basefeatures,names(data),value = T)

There are 2 standout issues I noticed immediately and that should be addressed in order to improve this function. As described in my checklist on code review, helper functions are amazing in R code. But they are only amazing if done correctly. If done correctly they improve code readability and transfer between projects. Lets focus on readability.


the first issue is no explanation of the functions inputs. Basic explanation should include an example input and short descriptions of type and shape. The second is explanation of the functions use case, explain this shorthand at the function wrapper. Thirdly it is important to explain the output of the function in the wrapper. Finally reevaluate the name of the function, does it clearly relate to the shorthand explanation of its purpose.

All these alterations also improve its transferability to other projects. There is a secondary issue in this functions first 5 lines. The parameters ‘features’ and ‘basefeatures’ used as input for the function are immediately reassigned. This is bad coding practice as there is no reason why the reassigned variables couldn’t have been used as original input for the function.


The revised code is shown below. It now contains a precise explanation of all inputs, the functions use case and its output. The name of the function has been altered and the parameters more appropriately renamed. The reassignment of the 2 input parameters is removed and incorporated them into our function inputs/wrapper.

LLTO_Model <- function(data, name_of_dataset, lag, time_out_horizon, features, predictors){
  ## Function purpose: Use Leave Location and Time Out Moddeling (LLTO) on the 
  ## DataSets of the Kaggle ACEA Smart Water Analytics Challenge. Datasets are 
  ## time series with multiple predictors and features. Lags are used as features.
  ## A horizon gets specified to model the Time Out Cross validation.
  ## Function Output: Train Object
  ## Function input:
  # data                Dataset from Kaggle Acea Smart Water Analytics challenge, 
  #                     containing Date, numeric predictors & feature set
  # name_of_dataset     Name of input dataset, used to save trained model
  # lag                 Amount of lags used for features (integer input)
  # time_out_horizon    Horizon of folds used in Time Out Cross validation
  # features            Names of features used in dataset (numeric variables)
  # predictors          Names of predictors used in dataset (numeric variables)

Code Review Example Preprocessing Setup

The next part of the function runs through the different data preprocessing steps before the model training begins. From a user perspective I expect some control over this piece of code. As a user of a function you need these options to make it transferable. The different preprocessing steps need to be more precisely explained aswell. I also have some thoughts on code placement inside this segment.

  for(i in 1:length(features)){
    for(j in 1:lag){
      data$temp <- Lag(data[,features[i],+j])
      names(data)[which(names(data)=='temp')] <- paste(features[i],j, sep = '_')
  data <- data[,which(colMeans(!>.2)]
  # Inlude seasonality:
  data$year <- as.numeric(substr(data$Date,7,10))
  data$year <- data$year - min(data$year) + 1
  data$month <- as.numeric(substr(data$Date,4,5))
  data$quarter <- ifelse(data$month <= 3,1,
                         ifelse(data$month >=4 & data$month <= 6,2,
                                ifelse(data$month >=7 & data$month <= 9,3,
                                       ifelse(data$month >9,4,NA))))
  data_long <- tidyr::pivot_longer(data, predictors,names_to = 'location', values_to = 'depth_to_groundwater')
  data_long <- data_long[complete.cases(data_long),]
  data_long <- data_long[which(data_long$depth_to_groundwater != 0),]
  #data_model <- data_long[,-grep('location|Date|name',names(data_long))]
  temp <- data_long[,which(!names(data_long)%in%c('depth_to_groundwater','Date','location'))]
  nzv <- nearZeroVar(temp)                                                       # excluding variables with very low frequencies
  if(length(nzv)>0){temp <- temp[, -nzv]}
  i <- findCorrelation(cor(temp))                                                # excluding variables that are highly correlated with others
  if(length(i) > 0) temp <- temp[, -i]
  i <- findLinearCombos(temp)                                                    # excluding variables that are a linear combination of others
  if(!is.null(i$remove)) temp <- temp[, -i$remove]
  data_model <- data_long[,c('depth_to_groundwater','Date','location', names(temp))]
  data_model$Date <- as.Date(as.character(data_model$Date), format = '%d/%m/%Y')

Preprocessing steps explained

The code above consists of 5 preprocessing distinct steps:

  • Feature addition
  • data shaping for modeling
  • Removal of missing information
  • Statistical preprocessing steps
  • variable formatting

In the current code none of these steps are explained or identified as such. It is up to the reader of the code to figure this out. In the code below I explained these different distinct steps so they can be located easier. They can also be adjusted to a new project or dataset more easily. It is now easier to evaluate them individually and improve on the function.

Code placement

There is an argument here that the preprocessing steps need to be placed into their own function entirely. It is a distinct use case that can be seperated from training the model. Sometimes it even is a requirement in a production environment. This happens when we have to preprocess live production data as we did when training the model. The code is already greatly improved simply by reordening it into the 5 steps above.

Other considerations

One of the concerns I have reading this part of my code back is the lack of stepwise explanations. A clear example is the section on missing values, its as simple as it gets. Remove all rows that even contain a single missing value. As a new reader or user I want to know so many more things about this choice. This includes the impact of this choice, what if it removes all rows? The alternatives that were investigated and even why it removes the missing values.

Code review comes in to make projects stand up to production level quality. You could argue that this level of detail is prohibitive but I am here to argue it is a requirement. When this model is in operation any issue may arise, perhaps the missing data increases and the model fails specifically on this point. A developer may start a new project to implement imputation techniques. Maybe you already investigated this and it led to a dead end. One of the key characteristics of production level code is its ability to handle diverse outcomes and report on them accordingly.

  ## Variable formatting:
  # Date variable:
  data_model$Date <- as.Date(as.character(data_model$Date), format = '%d/%m/%Y')
  ## Feature addition:
  # Creating lags for i features in j lags
  for(i in 1:length(features)){
    for(j in 1:lag){
      data$temp <- quantmod::Lag(data[,features[i],+j])
      names(data)[which(names(data)=='temp')] <- paste(features[i],j, sep = '_')
  # Inlude seasonality:
  data$year    <- as.numeric(substr(data$Date,7,10))
  data$year    <- data$year - min(data$year) + 1
  data$month   <- as.numeric(substr(data$Date,4,5))
  data$quarter <- ifelse(data$month <= 3,1,
                         ifelse(data$month >=4 & data$month <= 6,2,
                                ifelse(data$month >=7 & data$month <= 9,3,
                                       ifelse(data$month >9,4,NA))))  
  # remove uncommon newly created features:
  data <- data[,which(colMeans(!>.2)]
  ## Data shaping for modelling
  # Long format by predictors for LLTO modelling:
  data_long <- tidyr::pivot_longer(data, predictors,names_to = 'location', values_to = 'depth_to_groundwater')
  ## Remove missing variables:
  # Complete cases
  data_long <- data_long[complete.cases(data_long),]
  ## Statistical preprocessing:
  # Remove outliers
  data_long <- data_long[which(data_long$depth_to_groundwater != 0),]
  # Preprocess features for modelling
  temp <- data_long[,which(!names(data_long)%in%c('depth_to_groundwater','Date','location'))]
  # excluding variables with very low frequencies
  nzv <- nearZeroVar(temp)                                                       
  if(length(nzv)>0){temp <- temp[, -nzv]}
  # excluding variables that are highly correlated with others
  i <- findCorrelation(cor(temp))                                               
  if(length(i) > 0) temp <- temp[, -i]
  # excluding variables that are a linear combination of others
  i <- findLinearCombos(temp)                                                   
  if(!is.null(i$remove)) temp <- temp[, -i$remove]
  # Selection of final feature set
  data_model <- data_long[,c('depth_to_groundwater','Date','location', names(temp))]

Code Review Example Cross validation indices

One of the implementations of this function is the handmade folds for cross validation. I implemented this in the code because there were no libraries that used LLTO modelling exactly how I wanted to use it here. Looking back at my checklist, this is an important stage to bring up library optimization. Here we have a usecase where our code deliberately includes a selfprogrammed aspect, and it is our responsibility as the reviewer to verify that this was correct.

In this particular case I know that there is a CAST package in R that claims to implement LLTO modelling, but I found that it didn’t create the folds for cross validation correctly. This is however not documented in the code below, which is a mistake. The code should reflect this piece of research (perhaps even so far as to show its wrong implementation). This also draws back to my previous point regarding the stepwise documentation. It is a smart move to keep initial implementations and checks of code so that later you can reuse your research to back up your final output. I have failed to do so and hence I am open to critism about why I self-programmed this set of functions.

The function of the code below is to create indices of cross validation folds for every predictor and time period defined in the dataset. Only the first time period is excluded as there is no prior data that can be used to train for it. For example a dataset with 4 predictors and 4 time periods gets 12 folds. Caret expects a names list item that includes the elements ‘IndexIn’ and IndexOut’.

# Handmade indexes:
  index_hand_design <- function(data,period, location, horizon, location_one = NULL){
    horizon2 <- max(period)-horizon
      indexin <- which(data$Date >= min(period) & data$Date <= horizon2)
      indexout <- which(data$Date > horizon2 & data$Date <= max(period))
    } else {
      indexin <- which(data$Date >= min(period) & data$Date <= horizon2 & data$location != location)
      indexout <- which(data$Date > horizon2 & data$Date <= max(period) & data$location == location)
    output <-c(list(indexin),list(indexout))
  periods <- round(length(seq.Date(from = min(data_model$Date),to = max(data_model$Date), by = 'day'))/horizon,0)
  dates   <- seq.Date(from = min(data_model$Date),to = max(data_model$Date), by = 'day')
  indices <- 1:periods*horizon
  periods_final <- dates[indices]
  periods_final <- periods_final[!]
  for(i in 3:length(periods_final)){
    output <- list(c(periods_final[i-2], periods_final[i]))
    if(i <= 3){
      output_final <- output
    } else {
      output_final <- c(output_final, output)
  locations <- unique(data_model$location)
  for(i in 1:length(locations)){
    for(j in 1:length(output_final)){
        output_temp <- index_hand_design(data_model,output_final[[j]], locations[i], horizon, location_one = 'yes') 
      } else {
        output_temp <- index_hand_design(data_model,output_final[[j]], locations[i], horizon)
      if(j == 1){
        final_inner <- output_temp
      } else {
        final_inner <- c(final_inner, output_temp)
    if(i == 1){
      final <- final_inner
    } else {
      final <- c(final, final_inner)
  index_final <- list(index = final[seq(1, length.out = length(locations)*length(output_final), by = 2)], 
                      indexOut = final[seq(2, length.out =length(locations)*length(output_final), by = 2)])

Code function

I find this part of the code hard to understand and read. My excuse is that it was written under time pressure after I found out it had to be done by hand. It is however a prime candidate for this reason to review here.

One of the problems is that original function, “index_hand_design”. I will refer back to section 1 about how to properly design helper functions. All problems discussed with the entire function apply to this inner helper function. There is also an argument to not define a function within a function.

To finish off these practical examples I would like to show how to change the nested for loop I used to run the index_hand_design function. Removing for loops is at the highest point of my code review checklist . Below is that code segment changed into an lapply and mapply format.

    final_inner <- lapply(1:length(output_final), function(x){index_hand_design(data_model,
                                                                                location_one = 'yes')})
    # Create dataframe with possible combinations:
    sequences <- data.frame(location = rep(locations,length(output_final)), 
                                           lengths = rep(1:length(output_final), length(locations)))
    final <- mapply(index_hand_design, data = data_model, sequences[,2], sequences[,1], horizon = horizon)

The code above is endlessly more concise then the initial nested for loop. In my final pieces of code I always focus on replacing any remaining for loops into lapply or when using nested for loops mapply.

Wrapping up code review example of R Code in Caret

There were some other problems in the cross validation section, which will end up in the final code on github. I highlighted the removal of the nested for loop as a practicle example. The function was almost done anyway. It is very revealing to work through your own code from a code review perspective. I definitly recommend investing the time in optimizing your work in this way. Hopefully these practicle examples benefit your own work.

code review example R code Caret
Code review checklist R code edition
Blog, PSA, rblog

Code Review Checklist R Code Edition Top 3

Great code review is one of the most underrated skills a Data Scientist can have. In this blog I will share my top 3 code review checklist, specifically for R Code. In our data science team we regularly do code review to make sure it is up to the standard it needs to be. The top 3 elements mentioned in this blog should in my opinion be included in any code review for R code.

One of the greatest pitfalls of code review is developer bias. The goal of code review is to create the best code, not the code you as a developer like the best. I can’t claim this blog is completely whiped clean of developer bias, we all have it. For me it is the main reason why I disliked code review so much. I wanted to write the code I like best. My reasoning was that if it works, why does it even matter how it works. Over the years I have learned as members of our team come and go what code review can add. By using best practices we design our code to be better understood and easier to maintain. Anyway, here is my code review checklist Top 3.

R Code Review Checklist: 1. Helper Functions

Helper functions are without a doubt my number 1 on any code review checklist. It is not an understatement that every for loop ever written in the R language should be replaced with a helper function. When I review someone elses code this is the first scan I make. At its core R is a functional programming language, this should say alot on why we should put focus on this aspect. Some other reasons are:

Code Seperation

The benefits of helper functions are many, for me, since I love a great function it is also makes code easier to read. But seriously, one of the main reasons for me is actually how it seperates your code into concise chunks. Great production ready code consists of a central script that loads different helper functions as needed. In the helper functions the actual data processing, modelling, etc is performed. As a reviewer of your code this allows me to evaluate it in the context of what this specific code chunk is ment to do.

Transferability of Helper Functions

Once we build our code in functional chuncks/functions, it allows us to improve transferability of our code. This is true both for projects and colleagues. A core benefit of code review in R is that it allows you and your team to derive a core set of functions for your company data. Ultimately these functions could be gathered and transformed into a company package.

I started this segment by focusing on for loops, in reality helper functions should be more common. Helper functions are helpful to prevent repitition not only through for loops but also through cross-project functionality. So if I see a step written out in a colleagues code which has a high likelihood of being in need of repitition, I will mark it as a potential helper function.

R Code Review Checklist: 2. Package Optimization

What I remember most from starting out with R is the bewonderement of packages. Now I never had much other programming experience so stick with me on this. What was this magic that you could load functions to do stuff to your data as needed? And there are currently how many packages on Cran? The answer is 17260 packages. Thats a hole lot of code read to work with for you in your projects. One of the great skills an experienced developer can have is to be library agnostic, and this is my focus when reviewing code.

Why We Dont Hammer Down Screws

What exactly is library agnostic? it means that you are willing and capable to switch packages when needed. I’m trying to put emphasis on using the right tool for the right job, this should be a central element in code review. An area where this long was a problem for me is BaseR, I would nearly always use BaseR code as this is what I learned and was comfortable with. I had no immediate reason to use packages such as the tidyverse or data.table, my code was working fine. But it ended up holding me back, because I was incapable of switching packages as needed my code was inefficient in specific areas.

It is in my top 3 jobs of the reviewer to identify these areas and give feedback on optimizing the usage of packages. Look at what packages might fit the specific use case, what packages are implemented in the code? When providing feedback in this area I like to explain different options and their potential benefit (speed, understandability, etc). This in my opinion also helps the reviewer think more about the code and is an area that benefits both.

R Code Review Checklist: 3. Code Placement

When I write an initial script for a project, it always starts out quite structured. Its like a ritual, load data, select, transform, make visualizations, etc. But then one thing always happens, I notice something in my output, a weird pattern. Perhaps its a colleague that wants a slight alteration of the insights. What happens to my code now? I end up placing a quick filter on row 241. A week goes by, maybe even a month goes by, I revisit the project. The scope has changed slightly and I start adding code again.

If you are lucky this is a project that ends at some point, and your code gets retired with it. But what if it becomes a success? We are going to need production ready code that is shareable within the organization. Code review is all about this, we focus on reordering and optimizing code placement. As a code reviewer split the code in front of you into use cases and their approriate chunks. If someone loads data outside the ‘load_data’ chunk, mark it. If someone filters their data within a data visualization, mark it.

The value of great code placement is difficult to overestimate. It hits all the relevant boxes of code review. It improves readability of code and transferability between projects, and very often increases its functionality too.

Wrapping Up Code Review Checklist

These 3 points together make up the core of my code review, and belong on any code review checklist. These 3 points also have one thing in common. They all focus on outer appearance of code, I am a strong believer that this is the biggest benefit of code review. By optimizing readability and transferability we create synergies within data science teams and proejcts that are much needed.

You will find online articles that see validation of the output of code as part of code review. I personally think that any validation is part of the code writing process itself. If you dont have checks and balances in your code that validates your inputs and outputs thats a big problem. The role of the code reviewer is to check for these elements, but they are not responsible for writing or doing it themselves.

This blog was very focused on the ‘how’ of code review. If you are interested in an introduction on the what and why of code review, why it benefits your team or company or a practical example of code review I will release content on those subjects soon. Hopefully put together some of it can benefit you in starting out with code review.


My personal backstory; A road to Data Science for a living

As the title says, today I will share my own person backstory and  how it shaped me for data science to come into my life. As every person on the planet will tell you, their story is a little bit different, and so is mine. Now lets go down memory lane.

From a young age I’ve always been competitive, into strategy games and numbers. If I’m honest it wasn’t always completely healthy, I was always looking to confirm my own uniqueness and intellect when I was very young. For example in primary school from 6th grade you could play chess in a school selection, they would select 5 people and the winner would be  table number 1, all the way down to 5 to make a team. We would then as a team of 5 play against all the other schools in the area and number 1 would play against the number 1 of the other schools. I was number 1 3 years in a row, the first time that happened in school history and in my first year I was 9 and playing against these 12 year old’s. I thought it was awesome and it just reaffirmed everything I thought of myself.

Aside from the probably mental health concerns, what I enjoyed so much about chess from an early age is the mix between smart logic and time constraints. At the time I would also play age of empires, a real-time strategy (RTS) game that has the exact same tradeoffs between time and smart logic. This was my next addiction, during high school I competitively played RTS games online, I played a lot of e-sports but this was when the entire esports scene was in its infancy. RTS games taught me 2 things, it taught me how to search for optimal strategies given limited information, and it taught me the concept of studying and improving your strategies outside of the competitive environment, training.

By the time I started my economics studies in 2008 I hit a different period in my life, it was right around when online poker became main stream. This game just perfectly hit my interests, it was a natural evolution after playing RTS games, only the stakes were far more interesting. What was new about Poker? It was more math intensive, it brought me into the world of databases and statistics. It connected to my studies that brought me game theory and behavioral economics.

It was through online poker that I first got into contact with a programming language, there was this professional player who developed his own tool to calculate game theory optimal strategies in a given situation. He did this in python and shared his notebooks with everyone who followed his course. This was a very interesting new challenge for me, by going through his code and altering it to my own needs I self-taught basic Python. Later during my master we did some Stata, which really isn’t programming, but we also had the option to use R. I taught myself R too, just by copying code, looking online and using it for my master thesis.

By the time I finished my studies in Health Economics I was ready for a real job, interestingly enough I was not directly looking for a job in data science, it was economic analysis that was my principal interest. I actually had a job offer in hand from a consultancy in Rotterdam specialist in health economic analysis. At the time of that job offer I just started working part time doing some programming work for a data scientist who worked for this insurance company, MediRisk. He told me that they are looking to expand and got me an interview there. It was my curiosity for something new that brought me to the decision to go with the data science position at MediRisk.

See it can be easy to put something on a page and make it seem like a logical buildup and story, and in a way it was, but the real red line was always my curiosity for something new. Perhaps not completely new, but always one step forward and one step sideways. I love to broaden my view of the world by learning new skills, ideally by doing them. In my job currently I get to constantly explore new tools and technologies, understand them for their potential and excite others to join me in using them to create value. Learning by doing something new is something I aspire to always continue doing.

In a next blog I will share some tips and tricks into what I feel is most important if you want to start as a data scientist.