Event2Vec – interpretable embeddings for event sequences (github.com)

Event2Vector, a novel framework for learning representations of discrete event sequences, has been announced, introducing a simple, additive recurrent structure that results in composable and interpretable embeddings. The core innovation, termed the Linear Additive Hypothesis, posits that the representation of an entire event sequence can be effectively modeled as the vector sum of its individual event embeddings. This allows for intuitive vector arithmetic, facilitating the easy composition and decomposition of event trajectories. The model is available in both Euclidean and hyperbolic geometric variants, with the hyperbolic model specifically designed for hierarchical data structures, proving advantageous for embedding tree-like patterns with minimal distortion. The significance of Event2Vector lies in its user-friendly API and integration capabilities, mirroring scikit-learn’s design to seamlessly slot into existing NLP pipelines. Features like padded batching enable efficient processing of variable-length sequences, significantly accelerating training on large datasets. By offering convenient methods such as fit, transform, and most_similar for nearest-neighbor lookups, Event2Vector empowers practitioners to leverage sophisticated event sequence modeling without extensive overhead. This release could greatly enhance applications in many areas, from natural language processing to complex event tracking in various domains, highlighting the growing imperative for interpretable AI models.