Coresets and Coreset Trees¤
A Coreset is a weighted subset of samples from a larger dataset, selected in such a way that the selected samples maintain the statistical properties and corner cases of the full dataset such that training an algorithm on the Coreset will yield the same results as training that same algorithm on the full dataset.
The DataHeroes library can be used to build a Coreset as well as to build and maintain a more complex Coreset Structure, known as a Coreset Tree. Once the Coreset or Coreset tree is built, various data science operations can be performed on it, such as training a model, updating labels and removing samples.
Unlike a Coreset which is built over the entire dataset in one iteration, a Coreset Tree is comprised of multiple Coresets, each built separately for a chunk (batch) of the dataset (
chunk_size) and then combined iteratively in a tree-like manner.
In the illustration below, the entire dataset was split into 8 chunks, of 10K instances each (X1, …, X8) and a Coreset of up to 1K samples (
coreset_size) was built for each chunk separately (C1, C2, C4, …, C12). Every pair of Coresets is then combined into a new Coreset, in a tree-like hierarchy, with the root being the minimal subset, called the root Coreset (C15).
The Coreset Tree data structure has several advantages over a single Coreset computed over the entire dataset:
- A single Coreset requires the entire dataset to fit into the device’s memory and is therefore limited in the size of datasets it can handle. A Coreset Tree, does not have this limitation as it processes the dataset in chunks and can therefore handle any dataset size by splitting the data into the appropriate number of chunks.
- A Coreset Tree can be computed much faster since the Coresets’ computation can be distributed across a cluster of devices or processors.
- In a Coreset Tree, additional data can be added to the original dataset without requiring re-computation of the entire Coreset Tree. A Coreset just needs to be built for the additional data and then that Coreset is added to the Coreset Tree, while updating only the necessary Coreset nodes on the path to the root Coreset. This makes it a great structure to use for Model retraining in production.
- Similarly, updating the target or features of existing instances or removing instances, does not require re‑computation of the entire Coreset Tree, updating the Coreset nodes on the path to the root Coreset will suffice. E.g.: If we updated the target for some instances in X6, all we need to do to keep the Coreset Tree updated, is to update the Coresets C9, C10, C14 and C15, while the rest of the Coresets remain unchanged.
The main disadvantage of the Coreset Tree versus a single Coreset computed over the entire dataset, is that each level of the Coreset tree increases the approximation error, therefore leading to a slightly less accurate root Coreset. However, this error can be controlled and decreased easily, by increasing the
Building a Coreset or Coreset Tree¤
The class representing the Coreset Tree in the DataHeroes library is named CoresetTreeService. While the
coreset_size can be passed as parameters to the class, the default behavior is for them to be automatically calculated by the class when the Coreset is built based on the quadruplet: ‘number of instances’, ‘number of features’ (deduced from the dataset), ‘number of classes’ (if passed, otherwise deduced from the dataset) and the ‘memory’ (if passed, otherwise deduced based on the device’s memory). Based on this quadruplet the class will also decide whether to build an actual Coreset Tree or to build a single Coreset over the entire dataset. In the case of a single Coreset, every time additional data is added to the class using the partial_build function, the class will assess whether it is worth moving to an actual Coreset Tree, and will automatically convert the structure as necessary.
Data Cleaning Use Case¤
You can use a Coreset property referred to as Importance (or Sensitivity) to systematically identify potential errors and anomalies in your data. When building a Coreset, every instance in the data is assigned an Importance value, which indicates how important it is to the final machine learning model. Instances that receive a high Importance value in the Coreset computation, require attention as they usually indicate a labeling error, anomaly, out-of-distribution problem or other data-related issue. This version allows you to find the important samples in the dataset, regardless of what algorithm you use to train your model.
To review data instances based on their importance, first build a Coreset using any
build() function, while setting the parameter
optimized_for to 'cleaning', then use the
get_cleaning_samples() function to get the samples with the highest importance, for the classes of interest.
When you find incorrectly labeled samples, use the
update_targets() function to update their labels or remove the samples using the
remove_samples() function. Any such change to the samples will automatically update the Coreset data structure to reflect the changes.
Should you prefer to suppress these updates until you finish all your changes, set the force_do_nothing flag to True when calling the
remove_samples() functions and call the
update_dirty() function to update the Coreset data structure when you’re ready. More advanced cleaning techniques can be applied by using filters with the
When cleaning, it is important to remember that both the train and test dataset should be cleaned to maintain a high quality, non-biased test.
Training and Hyperparameter Tuning Use Case¤
Using our much smaller Coreset structure, you can train or tune your model orders of magnitude faster and consume significantly less compute resources and energy, compared to using the full dataset.
Use the DataHeroes’ library
fit() function to fit a model on the Coreset and the
predict() function to run predictions on the model. Alternatively, if you prefer to use other libraries for training, use the
get_coreset() function to retrieve a numpy array or pandas dataframe version of the Coreset which can be used with other libraries.
To check the quality of your Coreset, you can fit a model and compare its predictions to predictions from a model built using your full dataset. To decrease the approximation error in the case of a Coreset Tree, requesting level 1 or 2 instead of the default level 0 when calling
get_coreset() will return the requested level from the Coreset Tree.
To hyperparameter tune your model, use the library’s
grid_search() function, which works in a similar manner to Scikit-learn’s GridSearchCV class, only dramatically faster. Note that calling
get_coreset() and passing its output to GridSearchCV would yield sub-optimal results, as GridSearchCV would split the returned Coreset into folds. When a Coreset is split, it is no longer a Coreset and won’t maintain the statistical properties of your dataset. The library’s
grid_search() function overcomes this obstacle by obtaining the individual nodes of a certain level of the Coreset tree and using them as folds. By default,
grid_search() uses cross validation as it’s validation method, but one can select hold-out as an alternative validation method. It is also possible to use the
holdout_validate() functions directly.
The current version provides Coresets optimized for training with the following algorithms: all decision tree classification-based problems (including XGBoost, LightGBM, CatBoost, Scikit-learn and others), logistic regression, linear regression, K-Means, PCA and SVD. To build the Coreset or Coreset Tree use any build() function, while setting the parameter
Model retraining Use Case¤
You can use the Coreset Tree structure to update models in production when new data comes in, by updating the Coreset Tree with the new data and training the model on the Coreset, without having to re-train the model on the full dataset.
After building the Coreset Tree using one of the
build() functions, you can add additional data to the Coreset Tree using the
partial_build() function or one of its sibling functions
partial_build_from_file(), and retrain the model as explained above.
- Create a free account on https://dataheroes.ai/getting-started/.
- Install the library on your device by running:
pip install dataheroes.
- Activate your account by executing the following code once (from each device you’re using):
- Check out our documentation and examples available here.
The DataHeroes library has dependency on other libraries. Please note that when installing the DataHeroes library, older versions of other libraries you may be using, may get automatically updated. The table below shows the minimum version required from each library the dataheroes library depends on.