Data processing challenges and the relationship between data and models

June 27, 2018 | By Goran Marjanovic

Data processing is the cornerstone of how computers streamline business operations. Before computers, practices like bookkeeping and taking a census were entirely manual. Now we can handle hundreds of thousands of data sets per day, which allow businesses to scale and meet the evolving market demands.

AI and machine learning take data processing to the next level. Data scientists use sophisticated algorithms to analyze and refine even larger datasets at a much faster rate.

But as state-of-the-art as it is, data processing isn’t a magic wand. Computers are only capable of processing what they’re taught, and it takes meticulous planning and step-by-step execution to implement everything.

This guide provides a high-level overview of what steps are necessary to reap all the benefits of advanced computer technology. Each step is equally important, and all of them must be performed for everything to function correctly.

1. Pre-processing data

During the pre-processing phase, data scientists collect raw data and strip it of any unnecessary elements. This extraneous information is noisy and erroneous data that can slow down or even halt the process. For example, when gathering retail receipts to calculate revenues, information like the store address or customer’s name isn’t necessary and should be parsed out.

Data should be reduced to indicating the fields that are necessary to express a relationship between input and output. It’s important during this phase to gain a full picture of what data is necessary and available. This means accounting for missing data (blank fields can cause errors) and filtering incorrect values.

Sometimes it’s necessary to aggregate data. By aggregating data into broader records, meaning dividing it into multiple groups and then drawing the number for each group, you can reduce the data size and the computing time without tangible prediction losses. For example, instead of looking for the most purchased item of a given day through multiple years of a store existence, aggregate them to monthly or quarterly scores.

Normalisation (re-scaling) and dimensionality reduction are also very important preprocessing procedures before fitting any sort of model. Normalisation rescales the features because different features might have different scales. This is followed by dimensionality reduction with, say, PCA, which allow one to choose the right features for the modelling phase.

When helping clients, we’re inquisitive to ensure that nothing is missed because businesses can’t always fully articulate their needs at this stage. Once the data is cleaned, it’s then manipulated into a form that enables analysis.

2. Fitting the model

Once the data is prepared, it’s time to determine the model that must be used followed by the algorithm for parameter estimation. Modelling and Algorithms are the heart of data processing, and it’s important to use the algorithms that are scalable with your business. Sometimes this means building your own; at other times off-the-shelf algorithms can be customized to perform the necessary functions.

Running test data through an algorithm provides an output that can be measured for accuracy and viability through performance measures, histograms and other scientific methods.

Algorithm choice depends on the size of the data fed into it, the methodology considered for its parameter tuning, as well as the number of parameters to tune.

3. Post-processing data

Post-processing is as important as fitting the model. It involves employing procedures that allow the data scientist to determine the validity of their model output. In other words, they need to find ways to measure the variability and correctness of the output, and subsequently create safeguard measures against illogical and/or inconsistent predictions.

4. Presenting Results

Once all the work is done, presenting the output in an easy-to-digest manner is what separates you from everyone else. It’s an art that requires a data scientist who’s able to communicate and present technical content in a digestible way.


Data processing is a regular part of any business, and far too often, it’s done manually. Using machine-learning algorithms reduces processing time and errors, thereby leaving staff available to focus on more important matters. It takes an up-front commitment to get it done the right way, but once the initial work is done, your company’s bottlenecks are relieved.

Subscribe to HIVERY updates