It works on my machine Have you ever founds some great pieces of software only to realize that the code provided by the developer does not work anymore? The problem of software reproducibility in scientific research is becoming more and more important as number of new packages increases exponentially. But the problem is even more complex - there are not only strictly reproducible or irreproducible projects but also some that fall in between. That is why a zero-one rating, which is common approach to measure reproducibility, may be misleading. One metric to rule them all To enable comparison between different software, in this article group of students from the Warsaw University of Technology analyzed packages published in The R Journal and developed a 6-point scale to measure reproducibility of the article based on six main categories of problems that can be faced while trying to reproduce scientific paper results:

Our story begins The story of Hajada is perhaps not an Arabian epic, but still, but just like one of Shecherezade’s stories, it’ll keep you, dear Reader, on your toes. Sometimes, the data is incomplete and we need to cope with that. There are many methods of so-called “imputation’’, i.e. filling gaps in the data. Hanna Zdulska, Jakub Kosterna, Dawid Przybyliński took an effort to compare those methods. Here’s what they found out. Challengers approach What kinds of imputation did they compare? The popular ones, you may know them or not, namely: “Bogo replace”, replacing with mode or median, “MICE” imputation method, “missForest” imputation and “VIM’s k-NN”.

What is this blog entry about? Black-boxes are commonly used in computer vision. But do we have to use it? This article looks at this issue and we try to understand it with our small (but developed after one semester of machine learning experience) brains and summarize it here. What is this article about? Computer vision is cool. But it would be just as cool to understand how it works, and it’s not so obvious. Explainable methods of image recognition - which is de facto classification - cannot use logistic regression and decision trees, because every model loses transparency as its performance increases - not to mention understanding neural networks.

Meet Hajada! Have you heard of the Indian mathematician Hajada? We started to think about it, having read the title of the article “The Hajada Imputation Test” - it sounded somehow familiar… But you probably haven’t had any contact with him, because not so long ago there was no such man. He was born by the authors of the test and the article, and his name comes from the first letters of their names. So what is his test? Hajada decided to study the effectiveness and time efficiency of various methods of dealing with missing data. He juxtaposed three simple (or even naive) methods such as deleting rows or inserting random values and three more sublime methods, including mice and missForest algorithms.

Black vs white Machine learning seems to be all about creating a model with best performance - balancing well its variance and accuracy. Unfortunately, the pursuit of that balance makes us forget about the the fact, that - in the end - model will serve human beings. If that’s the case, a third factor should be considered - interpretability. When a model is unexplainable (AKA black-box model), it may be treated as untrustworthy and become useless. It is a problem, since many models known for its high performance (like XGBoost) happen to be parts of the black-box team. A false(?) trade-off So it would seem, that explainability is, and has to be, sacrificed for better performance of the model.