Frequently Asked Questions¶

Longitudinal data records the same subjects across multiple time points, usually called waves, whilst tracking multiple variables for each subject. In biomedical research, this could mean measuring blood pressure, cholesterol, and smoking status for the same patient across repeated visits. Its main value is that it captures change over time rather than a single snapshot.

Time-series data and longitudinal data both involve observations made over time, but they differ in several important ways:

• Focus: Time-series data usually follows one variable, or a small set of variables, across many time points. Longitudinal data usually follows multiple variables for the same subjects across repeated waves or visits.
• Nature: Time-series workflows often focus on ordered temporal sequences, frequently numerical, whereas longitudinal data commonly includes both numerical and categorical variables collected repeatedly for each subject.
• Time structure: Time-series data often involves many closely spaced observations, whilst longitudinal data usually involves fewer observations spread across longer periods such as months or years.
• Irregularity: Time-series data is often regularly spaced, whereas longitudinal data more commonly includes irregular gaps between observations or missing visits.
• Machine-learning goal: Time-series methods are often used to predict future values or detect temporal patterns in sequences. Longitudinal classification usually aims to predict a class observed at a single reference wave from features measured across several waves.

In summary, the main differences lie in how time is represented, how subjects and variables are organised, and which prediction task is being addressed. This is why methods designed for long numerical sequences do not transfer directly to the repeated-measures classification setting targeted by Sklong.

In Sklong, longitudinal classification means supervised learning where features are observed across several waves and the goal is to predict a categorical outcome. More precisely, the main setting is one where the features are observed across multiple time points, whilst the class variable is observed at a single reference wave.

The library is therefore designed around longitudinal classification workflows rather than generic forecasting or regression workflows, keeping the tooling focused on repeated-measures tabular data.

Auto-Sklong is the companion AutoML project for longitudinal classification. AutoML means Automated Machine Learning: instead of choosing a pipeline and tuning its hyperparameters by hand, the system searches automatically across candidate longitudinal pipelines and their hyperparameters.

In practice, this is a combinatorial-optimisation problem, because the system must explore many possible combinations of data-preparation methods, feature-selection options, class-balancing choices, classifiers, and hyperparameter settings.

In other words, Sklong is the library for building and running longitudinal workflows, whilst Auto-Sklong is the separate project for automating model and pipeline selection on top of that ecosystem.

Since this FAQ belongs to the Sklong repository, the full Auto-Sklong documentation and source code live separately:

• Auto-Sklong documentation
• Auto-Sklong on GitHub

You can, and MerWav-Time(-) is exactly that baseline. The issue is that a standard learner will then treat repeated measurements as ordinary independent columns, without understanding that cholesterol_w1 and cholesterol_w4 are the same underlying variable observed at different time points.

This matters because no single method dominates across all longitudinal datasets, and methods that preserve or explicitly exploit temporal structure often outperform plain flattening. Sklong exists so you can compare that baseline against more longitudinal-data-aware alternatives instead of assuming flattening is always enough.

The main setup assumed in Sklong is this one: features are observed across multiple waves, whilst the class is observed at a single reference wave, often the most recent one. In other words, the library is primarily designed for predicting one target from repeated measurements collected earlier or alongside that final reference point.

If your task has labels at every wave, is mainly sequence forecasting, or is centred on very long numerical series, that is a different problem family from the one Sklong currently targets.

Sklong prefers the wide format: one row per subject, with wave-specific measurements represented as columns such as smoke_w1, smoke_w2, and smoke_w3. This makes train/test splitting safer, reduces leakage risk, and works naturally with the metadata used throughout the library.

This also reflects the design of Sklong: rather than hard-coding dataset-specific column names, the library uses a wide-table representation together with lightweight metadata describing which columns form longitudinal groups and which columns are static.

If your data is currently in long format, it is best to pivot it first. The longitudinal data format tutorial and the advanced uneven-wave setup tutorial walk through that process.

A wave is an ordered visit or time point in your study. A features_group is the ordered list of columns belonging to the same longitudinal variable across waves. non_longitudinal_features are the columns that do not follow that temporal pattern, such as static demographics.

These metadata are central to Sklong. They let estimators and preprocessors reason about temporal structure and recency without hard-coding dataset-specific column names. This is consistent with the design choice in Sklong, where temporal grouping is represented explicitly through metadata rather than through bespoke code for each dataset. The temporal dependency tutorial explains how to define them in practice.

Keep the dataset in wide format, include the maximum number of waves you want to model, and leave missing visits as NaN when a subject simply does not have that observation. If an entire wave column does not exist for a feature, pad that position with -1 inside features_group so the temporal ordering still remains explicit.

In short:

• Use NaN for missing values within an existing wave column.
• Use -1 only when a whole wave column is absent.
• Keep one row per subject throughout.

The advanced temporal setup tutorial covers this pattern in detail.

Use the data-transformation pathway when you want to convert longitudinal data into a form that standard tabular learners can consume. This is usually the easiest path for baselines, quick experiments, or when you want to stay close to familiar scikit-learn workflows.

Use the algorithm-adaptation pathway when you want the downstream model to use temporal structure directly, especially with longitudinal-data-aware feature selection and estimators. The comparison between both pathways is strongly motivated because neither one dominates across all datasets. If you are unsure, compare both manually or let Auto-Sklong search across them for you.

These methods correspond to the main longitudinal pathways used in Sklong and Auto-Sklong.

• MerWav-Time(-) flattens all wave-specific measurements into a standard tabular dataset and treats the multiple values of an original feature at different time points as distinct features.
• AggrFunc aggregates each longitudinal feature across waves using a statistic such as the mean or median, yielding a single-wave representation.
• SepWav trains one classifier per wave and combines those predictions afterwards.
• MerWav-Time(+) keeps time indices explicit so that longitudinal-data-aware feature selection and estimators can operate on the temporal structure directly.

The first three belong to the data-transformation family. MerWav-Time(+) belongs to the algorithm-adaptation family. See the API pages for Merge Waves: Drop Indices, Aggregation Function, Separate Waves, and Merge Waves: Keep Indices.

Yes. Sklong supports both binary and multiclass classification for its main classification workflows. That includes the lexicographic estimators, NestedTrees, and SepWav variants.

The main differences are in how you interpret predict_proba, classes_, and evaluation metrics such as AUPRC. The binary vs. multiclass tutorial shows the same workflow in both settings.

Use class weighting when your target classes are imbalanced and missing the minority class would be costly. Class weighting is a simple way to rebalance learning pressure without changing the dataset structure itself.

Sklong is currently focused on supervised longitudinal classification. It is not a general-purpose framework for every temporal machine-learning problem.

The main current boundaries are:

• no heavy regression support yet, although it is wanted,
• no longitudinal neural-network family integrated into the core workflow yet, although it is wanted,
• no attempt to cover full time-series forecasting problems, since aeon and similar toolkits are already well shaped for that space,
• an expectation that temporal structure is provided explicitly through wide-format data and feature-group metadata.