How do you use Python or R to implement machine learning models?

In my experience, understanding the domain, from where the data is coming, is extremely important. Speaking to the domain expert - understanding each feature and its importance in the context of the problem is very important. It is important to understand if we can expect a relationship between the features - correlated or non linear relationship, etc.

Using Python or R for ML implementation must originally start from the business understanding (knowing the exact problems, objectives and goals), and requirement gathering (knowing the kind of useful resources, this could mean the data that may help in tackling the problem or even expert knowledge). Subsequently, problem formulation (knowing the best way/technique to solve the business problem and knowing the best way possible on the use of the available resources/methodologies) and lastly knowing how best to serve the models to make the most out of it (i.e., deployment/production platforms or integration). Until all these are sorted, Python or R would just remain a tool.

Everything has got pytonified in the field. Through well implemented wrappers on other high performance languages, python offers the perfect medium to develop data science, machine learning and deep learning solutions. However, interestingly there may be few R libraries on very niche subjects which has not been implemented yet in python. There is everyday new libraries being added to python environment and make sure to catch them up by reading latest blogs on medium and try out!

The choice of a statistical model and the associated methods to estimate its parameters are to be done very carefully. It is to be noted that a statistical model is only an approximation of the underlying data generating process. So, a parameter estimation method like maximum likelihood method which is considered asymptotically optimal, may not be necessarily good when a grossly wrong model is assumed. Here, it's good to recall once renowned statistician G.E.P. Box said "All models are wrong, fortunately some are useful". Thus, some form of model validation must be carefully carried out based on test data. However, the choice of 'test' and 'train' can also be tricky when the observations are autocorrelated. So, ML is an iterative process.

A very crucial point is to know the extent and capabilities of the statistical model or ML algorithm you will use. For instance, by their intrinsic nature, no tree based algorithm would be able to interpolate out of bounds or truly learn a functional optimization. This single example on tree algorithms is one of the mistakes many data scientists make due to popularity of random forest and xgboost; that certain tabular problems require functional interpolation.

The splitting of the data into 'train' and 'test' is useful only when performed iteratively even when the observations are collected independently from a data generating process (often assumed fixed). The choice of splitting the data can be tricky when the observations are autocorrelated (e.g., time series data or spatially correlated data). One needs to make sure that the inherent autocorrelation (or more generally auto-dependence) structure is not destroyed by a random split of the data set.

The choice of validation strategy is very subjective -like many aspects of statistics. There is no single correct answer. Such as proportions and number of folds, testing sets etc. One interesting choice one requires to make is on the aggregation of fold metrics. Each folded experiment is a unique ML modeling itself. So, shall we take the mean of the fold statistics, median or the worst metric observed ? Depending on your use case's requirements this may have a serious impact. So, just don't blindly use libraries' default validation procedures.

The model evaluation or validation method critically depends on the metric chosen to measure the discrepancy between the models. E.g., we can use the popular squared error loss or absolute error loss but if we use more intrinsic type of loss like entropy or KL divergence, it may lead to different choices of optimal models. Thus, one should clearly state the metric of the model evaluation and acknowledge its sensitivity to such a choice.

Cherish the autoML tools and low-code/no-code platforms -even you are an expert in coding. They open up immense perspectives and increase your productivity on whole machine learning stack. However, if you are a newbie they can limit you or even sometimes lead you to wrong conclusion, if you are not knowing the nuts and bolds inside. So, make sure you learn in detail what the tool itself is doing during automation of all steps : preprocessing, dimensionality reduction, model training, validation ..

Machine learning operations, or MLOps, are a set of robust practices to deploy and maintain machine learning models in production. The vetiver framework is built as a simple tool stack for both R and Python workflows.

I’ll argue to go with Python instead of R. Is a more tech friendly language, more easy to deploy and couple with the backend code and all the tools necessary are there! You can still use R for research, however if your implementation in production will be done in Python, you can save a lot of time doing the research also in Python.

Virtually everything in the ML implementation process has already been automated. The train and test population selection, model and algorithm selection, hyper-parameter tuning are all automated using various libraries. We don’t have to manually make any selections here. With deep learning, feature engineering also becomes semi-automated and generative AI infrastructure fine-tuned on your company data and artifacts will take care of 99% of pre-processing. What will become invaluable in the ML implementation process is the deep business acumen to create the solution and the deep engineering knowledge to create efficient and low cost model training, inference and deployment architecture and actually executing on them.

Performing Feature Analysis (Needs to be done after gathering business requirements, objective and data): Performing feature analysis based on business goals and requirements is a critical step in building effective machine learning models that align with the specific needs of your business. Feature analysis helps you understand the relevance and impact of different features (variables) in your dataset, allowing you to make informed decisions about which features to include, transform, or exclude.

Data Preprocessing activity is important, and key to success of of any Data Science project. Data Preprocessing activities like handling null data, outliers, imbalance data, conversion of categorical data to numerical data, selecting appropriate features. For feature selection different method could be used based on type of problem. Here domain expertise is a key along with different statistical methods for selection.