How does domain knowledge affect the accuracy of machine learning models?

Domain knowledge refers to the expertise and understanding of a specific field or problem that is essential for effectively applying machine learning. It involves knowing the relevant context, factors, and measurements related to the problem you're trying to address. This knowledge guides decisions in problem definition, feature selection, model choice, and result interpretation, making it a crucial component in the successful application of machine learning techniques.

Machine learning models under the hood are pure mathematical models and they are oblivious to the business context. Domain knowledge plays a significant role in finding the right data sources, feature engineering, preventing data leakage in the model & and aligning accuracy, measurement of impact to the context of the problem, and many more.

Just as a chef's in depth knowledge of ingredients leads to unique and yummy dishes, a ML expert's understanding of a certain domain leads to special and effective models. High reliance on domain expertise in machine learning models can sometimes result in data-influenced decisions instead of data-driven ones. To strike a balance between preconceived notions and ML-generated outcomes, decision intelligence becomes crucial. Furthermore, domain expertise proves particularly valuable in tasks like data preparation (such as labeling) and feature selection. Consequently, teamwork in ML projects is essential to mitigate any potential biases throughout the process.

Acquiring domain knowledge is a gradual process that involves various methods. You can gain domain knowledge by reading relevant literature to understand background, concepts, terminology, and best practices. Consulting with experts in the field can provide valuable insights. Exploring and analyzing data can reveal patterns and relationships. Experimenting with machine learning models allows you to align your models with domain knowledge and goals, comparing performance and accuracy. It's a continuous learning process that involves a combination of these approaches.

Em PLN, aplicamos a técnica de fine-tuning (transfer-learning) para aplicar conhecimento de um domínio específico em um modelo pré-treinado. Por exemplo, treinamos modelos clínicos e biomédicos para língua portuguesa no nosso grupo de pesquisa, usando como checkpoint BERT multilingue.

In my experience, domain knowledge is needed to transform the available features into useful features that will enhance the accuracy of the machine learning model. For example, if we want to create a model to predict pump failure, it would be useful to know the Best Efficiency Point (BEP) of the pump at all time. Deviating from BEP will decrease the pump lifetime. The BEP data is not readily available, however it can be calculated using the pump head and flow rate. It takes domain knowledge to be able to calculate the BEP and using the BEP data in the model for the prediction.

Building an ML product is not a one way road. Its a road where two things - domain knowledge and technical knowledge meet at the middle. Lets say for example, you want to predict wind energy generation. Firstly it’s important to recognise what are the factors that actually affect the wind energy which is usually shared by domain expert or client. On the other hand, it is ML engineer’s responsibility to check the quality of the data and further preprocess it to make it suitable to train the model.