Imagine that years of location data from thousands of people are used to train a mobility model. The model learns common commute routes, busy roads, frequently visited neighbourhoods, and broader travel patterns. Later, one person deletes their location history and asks that their information no longer be used.

Removing the original records from a database may be straightforward. But those records have already influenced the model during training. Deleting the data does not automatically reverse what the model learned.

This raises a difficult question: has the information actually disappeared, or does some of its influence remain inside the model? Machine unlearning aims to remove the contribution of selected training data from an existing model without rebuilding it completely. The challenge is to remove the requested influence while preserving useful knowledge learned from the remaining data.

Geographic information makes this harder because observations are connected through space and time. One person’s route may also contribute information about road use, travel times, and neighbourhood activity. Spatio-temporal graph unlearning shows that this problem is beginning to be studied in dynamic, interconnected data.

This article looks at why geographic information is difficult to forget, what GeoAI may need to unlearn, and how we can test whether a model has truly removed the requested influence.

Source: AI- Generated

Deleting Data Is Not the Same as Unlearning

Removing a textbook from a student’s desk does not erase what the student has already learned from it. The same problem appears in machine learning. Deleting a file removes it from storage, but the trained model remains unchanged because its parameters were already influenced by that data.

The most reliable reference is to remove the selected records and train the model again using only the remaining data. This produces a model that was never exposed to the deleted information, but rebuilding a large model can require substantial time and computing resources. Machine unlearning aims to avoid that full rebuild. Exact methods try to reproduce the result expected from retraining, while approximate methods update the existing model more efficiently and accept some difference from the retrained reference.

Some systems can also be designed so that later removal is easier. SISA training divides the training data into isolated shards and trains separate component models. When one record must be removed, only the affected component needs to be retrained rather than the entire system. This can reduce computational cost, but geographic data creates an additional difficulty because nearby observations, repeated routes, and connected locations may not separate cleanly into independent groups.

For GeoAI, the challenge is therefore not only how to update the model. It is deciding what information belongs to the requested deletion and what geographic knowledge should remain.

Why Geographic Information Is Difficult to Forget

Geographic information is hard to unlearn because spatial observations are rarely independent. A single route, image tile, or sensor record usually carries information about a wider system around it.

Take mobility data. One person’s trajectory does not only describe that individual’s movement. It also contributes signals about road connectivity, travel times, popular junctions, and neighbourhood-level traffic rhythms. If a user asks the model to forget their commute, the system should remove that specific contribution without erasing the road network or the broader travel patterns supported by many other users.

The same problem appears in Earth observation. Landscapes are observed through overlapping image tiles, repeated satellite passes, and multiple sensors. Deleting one image does not mean the model has forgotten that place, because nearby tiles often share building patterns, vegetation structure, and land-use context. In other words, geographic learning is spatially connected: the influence of one sample often extends beyond its own border.

This makes geospatial unlearning more complex than deleting a customer profile from a database. A model may be asked to forget a coordinate, a user, a route, an image tile, or even an entire region. These are different tasks, and they require different levels of removal. TraceHiding addresses this problem in mobility data by identifying user-specific trajectory information while preserving movement patterns shared across many users. CallosumNet extends the problem to spatio-temporal graphs, where removing selected information must not break the wider spatial and temporal structure.

Scales of geospatial unlearning, from removing one record to an entire region, with difficulty increasing as spatial dependencies become broader.

What Might GeoAI Need to Forget?

GeoAI may need to forget information for several reasons. The most direct is privacy. Location histories can reveal homes, workplaces, travel routines, and visits to sensitive places. The GDPR provides a right to erasure in certain circumstances, although deleting stored records does not automatically explain how their influence should be removed from a trained model.

Unlearning may also become useful when training data are found to be unreliable. A dataset may contain incorrect coordinates, mislabeled imagery, duplicated observations, sensor errors, or manipulated map information. Removing the source files prevents future use, but it does not reverse their earlier contribution to the model. Data ownership can create a similar problem. If a provider withdraws a dataset or changes its licence, developers may need a way to reduce its influence without rebuilding the complete model.

Regulatory actions show why this distinction matters. In the Everalbum case, the US Federal Trade Commission required the deletion of face embeddings and facial-recognition models developed from certain user data. The WW International and Kurbo settlement similarly required the destruction of algorithms or other work products derived from unlawfully collected children’s data. These cases involved removing complete derived products rather than selectively unlearning individual samples, but they show that deleting raw data may not always be considered sufficient.

For GeoAI, future requests could involve one user, route, image, dataset, or sensitive location. Reliable removal at each of these scales remains an open technical problem.

How Do We Know the Model Really Forgot?

Successful unlearning has two goals. The selected information should no longer influence the model, while useful knowledge learned from the remaining data should stay intact. A model that forgets everything has removed the target information, but it is no longer useful.

The strongest reference is usually a model retrained from the beginning without the selected data. Researchers can compare the unlearned model with this clean reference to see whether their predictions and behaviour are similar. But matching accuracy alone is not enough. Machine-unlearning verification distinguishes between behavioural checks, which examine model outputs, and parametric checks, which inspect changes inside the model. Membership inference tests can also examine whether an auditor can still distinguish forgotten samples from data that were never used for training.

Unlearning can also create new privacy risks. Unlearning inversion shows that differences between the original and updated model parameters may contain information that can be used to reconstruct data intended to be forgotten. UnlearnShield was proposed to reduce this risk while preserving model accuracy and the effect of unlearning.

GeoAI adds another question: what happened around the removed information? If one user’s route is forgotten, predictions for neighbouring roads and communities should not change unnecessarily. If one satellite image is removed, the model should retain valid knowledge supported by overlapping images and nearby observations.

The ideal result removes the requested influence while leaving the rest of the geographic model as unchanged as possible. In GeoAI, proving that one record was forgotten may not be enough. We may also need to understand what changed around it.

Designing GeoAI That Can Forget Responsibly

Most AI systems are designed to retain useful information. Future GeoAI systems may also need to anticipate that some data could later require removal.

The first requirement is better data lineage. Developers need clear records of where training data came from, which geographic areas and time periods they cover, and what licences or permissions apply. Without this information, it may be difficult to identify which parts of a model were influenced by a user, dataset, sensor, or region.

The second requirement is architecture. Models could be designed so that selected data sources or components can be updated without rebuilding the entire system. However, geographic data cannot always be divided into independent regions because neighbouring locations often share roads, environmental patterns, and repeated observations. Any modular design would therefore need to preserve these spatial relationships.

Finally, unlearning must be independently tested. A model should not simply claim that information has been removed. Proof-of-ignorance auditing explores how successful and failed unlearning can be distinguished without always requiring complete retraining as a reference.

GeoAI may eventually need both persistent memory and controlled forgetting. Memory helps models understand how places change, while unlearning provides a way to remove information that should no longer influence them.

The challenge is not only to make machines learn more about the Earth. It is to help them remove what should no longer be remembered, preserve what still matters, and demonstrate the difference.

How do you like this article on machine unlearning? Read more and subscribe to our monthly newsletter!