TabularDateTimeHistograms
Generates histograms to provide graphical insight into the distribution of time intervals in a model’s datetime data.
Purpose
The TabularDateTimeHistograms metric is designed to provide graphical insight into the distribution of time intervals in a machine learning model’s datetime data. By plotting histograms of differences between consecutive date entries in all datetime variables, it enables an examination of the underlying pattern of time series data and identification of anomalies.
Test Mechanism
This test operates by first identifying all datetime columns and extracting them from the dataset. For each datetime column, it next computes the differences (in days) between consecutive dates, excluding zero values, and visualizes these differences in a histogram. The Plotly library’s histogram function is used to generate histograms, which are labeled appropriately and provide a graphical representation of the frequency of different day intervals in the dataset.
Signs of High Risk
- If no datetime columns are detected in the dataset, this would lead to a ValueError. Hence, the absence of datetime columns signifies a high risk.
- A severely skewed or irregular distribution depicted in the histogram may indicate possible complications with the data, such as faulty timestamps or abnormalities.
Strengths
- The metric offers a visual overview of time interval frequencies within the dataset, supporting the recognition of inherent patterns.
- Histogram plots can aid in the detection of potential outliers and data anomalies, contributing to an assessment of data quality.
- The metric is versatile, compatible with a range of task types, including classification and regression, and can work with multiple datetime variables if present.
Limitations
- A major weakness of this metric is its dependence on the visual examination of data, as it does not provide a measurable evaluation of the model.
- The metric might overlook complex or multi-dimensional trends in the data.
- The test is only applicable to datasets containing datetime columns and will fail if such columns are unavailable.
- The interpretation of the histograms relies heavily on the domain expertise and experience of the reviewer.