I became a Data Science practitioner in probably the most peculiar, counter-intuitive way of becoming one. Before being confronted by the embroidered Python or R dilemma, before being exposed to Data Analytics Platforms, Math and Statistical theory; I concerned myself primarily with what I can best term “Data Transmutation”. I intuitively knew that before I could do Data Analytics, I had to somehow transmute the Data. I would primarily be mining Legal Data; its unstructured nature and textual rigidity would compel me to develop a way of modifying it into a form that is Analytics-receptive. This was before I even knew what ETL was.
I soon began quantifying Legal permutations and expressing them as mathematically weighted numerals for the purposes of efficient Data Mining. This method is designed specifically for Machine Learning Algorithms that require numerical attributes and weights for their calculations. I however realized that beyond mathematically quantifying Legal Data, it was in fact altering the Data completely and (to my surprise) not just Legal Data. That is how Data Transmutation as I understand it, was born.
Data Transmutation creates a synthesis of multiple events that are first encapsulated into a single expression, then transposed into a math function and finally transmuted into a coherent data point. It condenses Data into mathematically calculable numerals and symbolic expressions that weight the factual permutations of events and occurrences. The results are highly potent Data points; it’s like condensing Lite Beer into a liquid with an alcohol content of 100%. This distillation of Large Data sets into highly concentrated but rational Data points is very helpful. Data Transmutation is a bit like gene extraction. All data tells a story (some stories more exciting than others) and every story has a core phenotypical structure and genome. Data Transmutation is a way of delineating the Data into “DNA” strands that represent the foundational archetype of the story the Data is telling. Transmutations mine the primary consequence of events and occurrences by isolating the systemic functions of data; thereby extracting only salient truths. Think of the process of Diffusion in Biology, which means something that goes from a very high concentration to a low one once it expands and occupies larger spaces. Through Diffusion, a gas loses its potency and efficacy as it begins to spread, very similar to what happens during the collection and architecture of Data. When you Extract, Transform and Load Data, you’re essentially taking a series of events and fragmenting them into scalable features for the purposes of Algorithmic enquiry. This fragmentation increases density, widens factual parameters, increases variation and ultimately “diffuses” the efficacy of the story. However when Data is transmuted into an condensed form; factual parameters aren’t unnecessarily expanded, the features become more salient, the density remains the same and variation is kept at a healthy level.
here is of course Sampling, Feature Optimisation, Feature Generation (combinational vector creation) and many other tools which are all used to perform the function of distilling data into a state of optimum lucidity. However there is a difference between Transmutation and Segmentation, which the above tools essentially are. They minimize and optimally abridge the Data to be analysed, they do not fundamentally mutate it.
Discretization methods are ubiquitous on all good Analytic platforms, that is probably the closest you can come to changing the aesthetic identity of data points without using Data Transmutation. You could certainly use Discretization to convert numerical attributes (where some entries are “0”) into binary attributes detailing “Yes” or “No”. This however cannot be done without compromising the structural and probative integrity of those data points.
Nevertheless, the definitive feature of Data Transmutation is the ability to mathematically calculate the value of a transmuted data point, without using Analytics to do so. Consider a classification model for Gold for example. One of the data points under the attribute “Pressure and Temperature Data” has been Transmuted from 27.0 GPa of Sheer Modulus (original data point), into a mathematically weighted expression of (P+) 4.833 (Transmutation Value). Because the data point has been delineated into a math function, it is possible to calculate the mathematically representative value of Sheer Modulus, in short-hand form, without using any code or software: with pen and pad. Think of Machine learning Algorithms that have the ability to produce Formulas for their results or at a very abstract level, even Map Reduce; Data Transmutation works in a similar way.
I am not saying that this method of altering data is a divine panacea, just like a Machine Learning Algorithm there are conditions and parameters that it must satisfy to perform optimally. All I am saying is that there is a way of changing data to facilitate a more advanced method of Machine Learning. Unfortunately Transmutation will inevitably lengthen the already protracted ETL process, however the rewards are bountiful. As Data Science practitioners we should let go of the fear of “corrupting” data. Change it as you see fit and you may be pleasantly surprised by the results.