In the hype of the moment, people sometimes forget that. It is tempting to believe that A.I. can solve every problem, but blindly following that hype can lead to unrealistic expectations.
I have said it here before: quality data is the lifeline of artificial intelligence. Around 90% of building an A.I. model is preparation: collecting, cleaning, preparing, analysing and enriching data.
Right... good preparation done, now train the A.I. model and hey presto, fantastic results? No. Even then, it can still happen that no useful patterns can be found in a high-quality dataset. Why exactly?
1️⃣ Complexity: some problems are inherently more complex than others and can be difficult to model with A.I., especially when there are too many unpredictable variables.
2️⃣ Representativeness and variability: a dataset may not be sufficiently representative of the variability of the real-world problem. This can lead to a model that performs well on the dataset but generalises poorly in practice.
3️⃣ Hidden variables: important factors or variables may be missing from the dataset that are essential for making accurate predictions.
Still rather abstract? A practical example to clarify:
Suppose you want to predict bookings for a hotel on the Belgian coast based on historical booking data provided by the hotel. The data looks consistent, complete and extensive, with no gaps. That has to produce good predictions, right?
Not necessarily. Our dataset is missing crucial data. Which data? Current weather data, COVID-19 pandemic data from recent years, school holidays, data about local events, public holidays and possibly promotions, competitive data (e.g. the opening of a new hotel nearby), etc.
Each of these may have influenced our historical booking data, and prediction accuracy will probably suffer from the absence of this critical information.
Failing to identify patterns is not necessarily a loss; it can also produce valuable insights. The iterative process of building, testing and adjusting AI models is a fundamental aspect of our field, where learning from "failures" is part of the work. Preliminary data research ensures that we build on solid foundations. Taking the time to understand the data increases the chance of success.
In short: there is nothing wrong with being ambitious, but realise that if you want to integrate artificial intelligence into business processes, not every AI model will be successful in every situation.
