Event2Vec
A sequence event embedding method that represents history as a rolling sum of event vectors (additive embedding), allowing any linear score to be decomposed exactly into the contributions of individual events. It supports multiple geometric spaces such as hyperbolic space and spherical space.
Timeline (last 90 days)
Event2Vec defines a history h_t as the rolling sum of event vectors: h_t = h_{t-1} + e_{x_t}, with norm clipping performed at each step.
Event2Vec performs next-event prediction via softmax(W h_t + b), and its history representation is additive.
Event2Vec training is self-supervised, requires no manual labels, targets next-event prediction, and adds a reconstruction term requiring that h_t - h_{t-1} can reconstruct e_{x_t}.
Since the history is a sum, any linear scoring model built on it can be exactly decomposed into the independent contribution of each event: score = w·h + b = Σ_i (e_i·w) + b.
The theoretical details of Event2Vec (including how multiple loss functions work together to ensure the arithmetic properties hold during training) are written up in a NeurReps@NeurIPS Workshop paper.
The Event2Vec paper's number is arXiv 2509.12188.
Event2Vec's code is open-sourced under the MIT license, can be installed via pip install event2vector, uses a PyTorch backend, and supports CPU/CUDA/MPS.
Event2Vec provides a scikit-learn-style fit/transform API and a gensim-compatible most_similar interface.
The Event2Vec library provides several non-commutative or geometrized alternative spaces: hyperbolic space (Möbius addition), the κ-stereographic model, spherical space (SO(n) rotations), and the Heisenberg group.
Event2Vec's starting point was root cause analysis of failure sequences at the Advanced Light Source (Berkeley Lab), where the authors originally used Word2Vec on event logs.
2 more timeline events
All Facts (12)
Event2Vec defines a history h_t as the rolling sum of event vectors: h_t = h_{t-1} + e_{x_t}, with norm clipping performed at each step.
50%UnverifiedEvent2Vec performs next-event prediction via softmax(W h_t + b), and its history representation is additive.
50%UnverifiedEvent2Vec training is self-supervised, requires no manual labels, targets next-event prediction, and adds a reconstruction term requiring that h_t - h_{t-1} can reconstruct e_{x_t}.
50%UnverifiedSince the history is a sum, any linear scoring model built on it can be exactly decomposed into the independent contribution of each event: score = w·h + b = Σ_i (e_i·w) + b.
50%UnverifiedThe theoretical details of Event2Vec (including how multiple loss functions work together to ensure the arithmetic properties hold during training) are written up in a NeurReps@NeurIPS Workshop paper.
50%UnverifiedThe Event2Vec paper's number is arXiv 2509.12188.
50%UnverifiedEvent2Vec's code is open-sourced under the MIT license, can be installed via pip install event2vector, uses a PyTorch backend, and supports CPU/CUDA/MPS.
50%UnverifiedEvent2Vec provides a scikit-learn-style fit/transform API and a gensim-compatible most_similar interface.
50%UnverifiedThe Event2Vec library provides several non-commutative or geometrized alternative spaces: hyperbolic space (Möbius addition), the κ-stereographic model, spherical space (SO(n) rotations), and the Heisenberg group.
50%UnverifiedEvent2Vec's starting point was root cause analysis of failure sequences at the Advanced Light Source (Berkeley Lab), where the authors originally used Word2Vec on event logs.
50%UnverifiedIn Event2Vec, norm clipping, whenever the history vector is updated, if its L2 norm exceeds the threshold τ, shrinks it back to the sphere according to h_t ← h_t · τ / max(τ, ‖h_t‖).
50%UnverifiedEvent2Vec does not attempt to replace large sequence models, but rather draws a boundary between interpretability and sequence modeling, and is suitable for teams that need event-level attribution under linear scoring.
50%