# What you will learn
The real world is full of data, there are many types of data and collecting them is an art and comes through experience and doing it in an iterative process.
Also, this is one of the preliminary steps when you are about to do Machine Learning or Deep Learning models. The cascading steps or processes depend on quite a bit on this task but at the same time, this is a task that can be iterated many times.
And revisited depending on what we conclude or find in the subsequent steps and processes (you can look at the index to see what these steps are).
We will also introduce a flow diagram eventually to show you where in the flow this task/step sits.
The purpose of this guide is to show you the importance of these steps, but also how important data collection can be in the whole cycle.
- How to use this guide?
- The end-to-end process
- Research or business questions
- Start Small
- Small steps
- The Toolkit and sources
- Data Generation
- Sanity Check
- Automate The Boring Stuff!
# How to use this guide?
We'll give you some general tips to learn effectively and develop a decent foundation, that you can rely on to address and solve Data Science problems in the complexity of the real world (which is messy by definition).
In this guide, you'll find many different tips and the sub-topics answer these questions provided below:
- why? (see Research or business questions)
The last two "how?" and "what?" overlap a bit and expand into these:
And to keep track of the "big picture", refer to The end-to-end process section. Any other section(s) not mentioned are supporting sections to this guide.
As stated in Virgilio's Teaching Strategy Guide, read more about it here (opens new window).
# The end-to-end process
Here's a snapshot of the end-to-end flow (high-level perspective):
Details are omitted and focus is drawn to the current subject (Data Collection) we are talking about. But this is the "big picture" to also keep in mind when working through individual steps.
# Research or business questions
The first thing before or during the Data Collection process will be asking the right question(s). The right business questions (opens new window) may be hard to come up with, but it has the biggest impact (opens new window) on your performance of solving a particular problem. Remember, you want to solve a problem (opens new window), not to create new ones!
And the same goes for asking the right research question. If you intend to establish something based on a hunch or curiosity or other research then there has to be some good base to start with.
Your data collection process or step is entirely dependent on the research or the business problem you are trying to solve.
At this point it's also important to understand that the end-model you are on the quest to create, is a direct reflection of these two main factors:
- Data (or datasets used)
- Method (algorithm(s) used)
Although the second factor that is, "Method" will come a bit later, the first factor is the primary or core component of the end-model. Methods can be swapped and occasionally the end-results may not differ by much but good data or the right data is the key to getting a good model, otherwise you immediately can see the effects of GIGO (Garbage In, Garbage Out) (opens new window).
When coming up with a question or a problem statement, it's also necessary to know the definition of "done" and use examples if possible to establish what the end-results should/could be comparable to once we reach the endpoint of the process (each iteration or the last iteration).
# Start Small
It's not efficient to try to handle Gigabytes or Petabytes of data each time you want to create a dataset as part of the Data Collection process. Just use small subsets (opens new window) of the data (but take care that the data is representative and you know the problems or nuances it contains). Once you are certain about your process and the end-to-end flow has matured and stabilised you can consider using bigger chunks of data for the new iterations.
# Small steps
Just like the previous section Start Small, starting small and staying small or growing slowly (or incrementally) is the most optimum way to proceed with Machine Learning or Deep Learning problem solving - given there are so many variables and moving parts. It's best to minimise the moving targets and pin them or fix them to a minimum when working on any part of the end-to-end flow.
And so knowing clearly whether to increase the quantity of data to collect (or not collect at all) at the beginning (or end) of the end-to-end flow is a good way to go about. You will be able to determine this once you have gained enough intuition about the specific subject/domain and your end-to-end process to hand. Sometimes you may even decide to reduce the data to hand, if that makes sense based on the gathered observations and results.
# The Toolkit and sources
Here is a list of things you could be doing to gather fresh data or existing data (and repurpose it for your needs):
- Web Scraping tools and services
- Look for pre-existing datasets
- Kaggle Datasets
- Kaggle Kernels using custom datasets
- Other competition sites
- Other public domain sites
- Private domain sources (respecting the licenses)
- Existing data in excel sheets or other forms within teams/organisations
- Collecting log files of program or application executions from live systems
- Running surveys via automatic services: Google Forms, SurveyMonkey, etc..,
- Convert non-digitalised data using modern technology
- Data generation services and tools
Many of the above can be done in a "live" manner or through a
cron (scheduled) jobs and others through manual intervention. Which means you could have a "live" pipeline constantly collecting data and generating datasets for your Machine Learning / Deep Learning pipeline.
What is a dataset (opens new window)? And how to create them? Also another important question is what should the dataset look like? This will quite depend upon on your Research or business questions to hand.
Having a look at how others have created their datasets for the various domains or topics or research subjects, can also help in the process. See this Datasets resource (opens new window) which is a mix of both clean and unclean (raw) datasets. This can also answer the more specific question, "what file format should the dataset be saved in?" CSV (opens new window)? JSON (opens new window)? or any other know file formats.
They can throw light on types of data (features) to collect or not to collect, and in what form should they be in data types (opens new window) when collected such that it's useful for the next steps in the process.
# Data Generation
Data generation is also a good MVP (opens new window) step/process, and also could be incorporated as one of your steps during the first few iterations - if you discover or know from observation that the data to hand is not sufficient or correct or even not usable for the end-purpose (see Research or business questions).
Data Generation can be useful when a domain or field has limited data from previous experiments, or the exercises have resulted in a small dataset. But the dataset may be somewhat representative of the problem that is being solved - in these cases, Data Generation using existing data can help. This process is also called "synthetic data generation" if the data is generated from scratch, and "data augmentation" if it's generated based using the original data (parts of it or whole of it). See Data Augmentation for Deep Learning (opens new window).
This of course has it's pros and cons - you soon get a lot of data but then is the data biased and in what way? Is it balanced or not, if not what to do - all these kinds of questions come to the surface.
This topic may need a section on its own but many resources and facilities are propping up these days that do exactly this, see this resource on Data Generation (opens new window).
An important point to note, in the interest of privacy and ethics, the data is anonymised and/or balanced to reduce any injected bias via this process of data generation.
"A picture speaks a thousand words" or "A picture is worth a thousand words" or other variants - they all mean the same thing!
There is a misconception that Data Visualization (opens new window) takes place towards the end or during some other process in the end-to-end process. But the fact is, whenever you have data, no matter what state it is in, we can visualise it (partially, if not fully) . And gain an insight into what it is representing and what state it is in. It is an art and not just science to be able to do this, the science part of drawing pictures from it is more or less solved, but evaluating these pictures and diagrams is another skill. And depending on how well one knows their domain may or may not be able to make sense from them -- which again goes back to the Research or business questions.
But try to visualise as much as possible at every opportunity available, to be able to understand the changing nature of the data to hand.
# Sanity Check
You always want to be sure that your data are exactly how you want them to be, and because of this is a good rule of thumb is to apply a "sanity check" at the end of the Data collection process.
Although sometimes we won't know "what sanity checks to apply?" till we have run through the rest of the steps in the process and then come back the next iteration(s) and fine-tune this step. So that a process can catch the important low-hanging fruits and that do not hinder the rest of the processes. Sometimes you might just have to proceed with the data and the steps that follow i.e. Data Preparation, Data Cleaning, Data Visualization, etc... as later on these steps will help reveal much about the collected data and decisions to take at any point in the process.
As we talk about Data Visualization (opens new window), the Visualization step in the previous section itself is a good enough "sanity check" of your data - provided you have drawn those charts and have the skills to evaluate them. If not, hopefully, the rest of the steps will help you capture those insights (or other ones). And sometimes with experience and doing multiple iterations and observing various experiments, visualizations on older datasets (snapshots) start to make more sense.
Then there is also ethics you do not want to miss out on and the section to follow gives you insight into it. And finally, also interpretability/explainability which overlaps with ethics in a good way.
With rising concerns over privacy and bias, you want to be sure that the data collected does respect the ethics and standards in this field as much as possible.
To help with that as the awareness about things are improving, there are a lot of resources available, one such place to start would be here (opens new window). One of the resources mentioned there is that of a python package called Deon (opens new window). Interestingly it has a digital checklist you can consult and see if they apply to what you are about to do.
# Interpretability / Explainability
Another widely debated but essential topic is the ability for us to interpret or explain our models and what better place to start but during the Data Collection process. Another good resource (opens new window) to look into when although it's an ever-growing topic and subject to keep track of.
Other useful links to look at are TensorFlow: Responsible AI (opens new window) and https://pair.withgoogle.com/chapter/explainability-trust/ (you can find some more useful ones under People + AI GuideBook (opens new window)). Another important resource to check here is this course/tutorial on causality (opens new window).
Ensure the features in the dataset which will end up becoming your real-world model are explainable and not a blackbox (opens new window) to the real-world.
# Automate The Boring Stuff!
Data collection can be a tedious process but you could also automate (opens new window) the most you can. Also, automation is married with iteration, so this is the way you need to plan your data collection pipelines. It's not easy to recommend automation tips but here are some of the things you could consider doing (depending on your needs and the Research or business questions you are trying to answer), see The Toolkit and sources.
Beware that not in every case you may be able to automate the process, it will all depend on the domain your problem statement is created from and how digitalised it is. But sometimes some data may need to be extracted manually or through human-intervention or through traditional methods. Sometimes it's easier to not digitalise or automate something till the need for it arises, a few manual steps at a rare or occasionally is acceptable if Data Collection isn't performed regularly. At the end of the day try to strike a balance between need, efficiency and accuracy - during the initial steps you may be able to discount these a bit but as you do more iterations discipline and precision are important, so efficiency is our friend then.
These two resources are quite extensive and have been rated by experienced professionals in the industry:
- 👉 Effective Data Collection 👈 (opens new window)
- The Ultimate Guide to Effective Data Collection (opens new window)
- People + AI GuideBook: Data Collection + Evaluation (opens new window)
It would be good to have a glance at them to see if you have not considered something. Such resources can also help validate the things we already know and how we are using them in the context of the problem to hand (and how others have been using them as well).
Also look at this talk on "Do we know our data, as good as we know our tools?" (opens new window), it starts off with an introduction to "Data Collection" and then into the next steps in the end-to-end process, also look for additional resources towards the end of the talk/slides and the Appendix section of the slides.
But don't just stick to these build your acumen and also look for other resources everywhere else.
So now that you have an idea of what to do when you want to start solving a business problem using techniques like Machine Learning and Deep Learning. You can also start thinking about how to do this in easily and keep building on top of it, otherwise, you may be forever busy collecting data but not proceed further from there. And not know the impact or use of the data collected. And thereby not gain the insight needed to answer your query or solve the problem statement.
Do also keep in mind the due-diligence to follow when data is collected and converted into datasets to be passed on to the next process in the end-to-end pipeline.
May be another read of the guide, taking notes from the various sections and external resources, creating a mind-map of all of the above could help give you a better insight into this initial but important step into Data Science.