Toward computing attributions for dimensionality reduction techniques.

Area of Science:

• Computational Biology
• Data Science
• Machine Learning

Background:

• Dimensionality reduction techniques are vital for analyzing complex biological datasets.
• Interpreting the features driving these reductions, such as t-distributed Stochastic Neighbor Embedding (t-SNE), remains a challenge.
• Local feature attribution methods are needed to understand dimensionality reduction outputs.

Purpose of the Study:

• To develop a method for computing local feature attributions for dimensionality reduction.
• To apply this method to the t-SNE algorithm for enhanced biological data analysis.
• To provide an efficient implementation and validation of the feature attribution technique.

Main Methods:

• Utilized gradient-based feature attribution, a technique from supervised classification, adapted for dimensionality reduction.
• Developed an efficient gradient computation implementation for t-SNE.
• Validated the method using synthetic datasets, the MNIST benchmark, and a SARS-CoV-2 sequence dataset.

Main Results:

• The developed method successfully identifies significant features in dimensionality reduction.
• Validation on synthetic and benchmark datasets confirmed the accuracy of feature identification.
• Explanations derived from the method align with domain knowledge in SARS-CoV-2 data analysis.

Conclusions:

• The proposed feature attribution method enhances the interpretability of t-SNE for biological data.
• An efficient Python package, interpretable_tsne, is available for practical application.
• The framework provides a roadmap for applying similar explanation methods to other dimensionality reduction techniques.