Part two: the impact of poor data quality

In the first article in this series, published on Wednesday 14^th August, we discussed the importance that data quality plays in the effectiveness of data analyses by Machine Learning (ML) and AI. Characteristics that define data quality were listed as:

• Completeness
• Consistency
• Lack of bias
• Accuracy

To understand the impact of these characteristics, one needs to appreciate the sensitivity that algorithms have to them, as well as the importance of choosing the right algorithm for the analysis being planned. These are complex topics requiring detailed explanations that will be discussed in future articles. 

Below are some of the key problems that poor data quality creates:

• Incomplete data can cause ML models to miss patterns or relationships. For example, if crucial bioactivity data for certain compounds is missing, the model may not fully account for the structure-activity relationships, leading to inaccurate predictions.
• Inconsistencies, such as variations in how data is recorded (eg, different units of measurement or naming conventions), can confuse models and lead to erroneous predictions. For instance, if the same compound is labelled differently in different datasets, the model might treat it as different entities, skewing the results.
• Bias in data can lead to models that are not generalisable or that perform poorly on certain subsets of data. For instance, if your training data is biased toward a particular chemical scaffold or a specific set of biological targets, the model might be less effective in predicting the activity of compounds outside these categories.
• Noise in data, which may arise from experimental errors, variability in biological assays, or inconsistent conditions, can obscure true signals and reduce the model’s ability to learn relevant patterns. This can result in a higher rate of false positives or negatives.
• Duplicate records can distort the training process by giving undue weight to certain data points, leading to overfitting and a lack of generalisation in the model.
• In cases where the data is highly imbalanced, such as having many more inactive compounds than active ones, models can become biased towards predicting the majority class. This might lead to poor performance in identifying active compounds.
• Redundant features or data points can inflate the dimensionality of the data without adding new information, leading to overfitting and reduced model performance.

• Scientific reproducibility is compromised when data quality is poor, as other researchers or systems might not replicate the findings.

In the next article in this series, which will be published Friday 13^th September, we will discuss pragmatic guidelines to help support better data quality and to discover if your data quality is compromised.