Can the tool be adapted for multi-label classfication?

[lcarnini]

Hi, we are looking for a tool to assiging categories to incoming customer request. Since categories can be added at later times, this looks like the perfect approach for the usecase.
The question is if we can use this for assisgning multiple arguments either using the predicted probablities and taking the ones with simlar value and/or above a threshold or adapting this using the multilabel classification task for the model ( if even possibile )

Thanks in advance!

[codelion]

Yes, the adaptive classifier is well-suited for your use case! I have created an example Google colab notebook to show how it can be done.

The MultiLabelClassifier class wraps the base AdaptiveClassifier and adds multi-label support through:

• A probability threshold to determine which labels to include
• A minimum probability difference to group similar-confidence predictions

For training:

• Each text example can have multiple labels (stored as a set)
• The training data is flattened into single-label examples internally
• This allows the base classifier to learn the relationship between text and each label independently

For prediction:
The classifier returns multiple labels when their probabilities are:

• Above the specified threshold (default 0.3)
• Close to each other (within min_probability_diff, default 0.1)

You can adjust the thresholds based on your needs:

• Increase threshold for stricter label assignment
• Decrease min_probability_diff to allow more labels with similar probabilities

You can do something similar for your use-case:

# Initialize with suitable thresholds
classifier = MultiLabelClassifier(
    threshold=0.1,  # Lower threshold allows more categories
    min_probability_diff=0.05  # Similar probabilities grouped together
)

# Train with your existing categorized requests
classifier.add_examples(historical_requests, historical_categories)

# Use for new requests
predictions = classifier.predict("New customer request text")