ORION: Integrated Runtime Modelling for Predicting Deep Learning Training Time

ICPE 2026

ORION predicts deep learning training time by jointly modeling GPU computation, CPU preprocessing, data loading, and storage I/O. By capturing both compute-bound and I/O-bound bottlenecks, it provides more accurate hardware-aware training-time estimates and improves prediction error by 44.36% over a GPU-focused baseline.

CAPE: Generalized Convergence Prediction Across Architectures Without Full Training

TMLR 2025

CAPE predicts how many epochs a deep learning model will need to converge before running full training. It uses a small initialization-time probe to extract features such as initial loss, gradient norm, NTK trace, parameter count, batch size, and learning rate, then predicts convergence across MLPs, CNNs, RNNs, and Transformers. CAPE achieves strong accuracy, with a 0.89 Pearson correlation in cross-fold evaluation, helping reduce trial-and-error in model selection and training planning.

PreNeT: Leveraging Computational Features to Predict Deep Neural Network Training Time

ICPE 2025

PreNeT predicts deep neural network training time using layer-specific computational, memory, and hardware features. It supports diverse layer types, including attention, embedding, convolution, dense, normalization, and recurrent layers, and can estimate training time on unseen GPU configurations. PreNeT improves prediction accuracy by up to 72% compared to prior training-time prediction baselines.

Neural Embedding Features for Point-of-Interest Recommendation

ASONAM 2019

Neural Embedding Features for Point-of-Interest Recommendation proposes a venue recommendation method that learns users and venues in the same neural embedding space using users’ check-in sequences. It derives user, venue, and user–venue interaction features from these embeddings and integrates them into a feature-based matrix factorization model. The approach outperforms several POI recommendation baselines, showing that embedding-based similar users and venues are strong signals for future check-ins.

Point-of-Interest Recommendation Using Heterogeneous Link Prediction

EDBT 2018

Point-of-Interest Recommendation Using Heterogeneous Link Prediction frames venue recommendation as a link prediction problem over a heterogeneous graph containing users, venues, categories, regions, and social relationships. The method uses meta-paths to capture relationships such as friends visiting venues, venues in the same category, and venues in the same region, then ranks recommended POIs using a lightweight classifier. It improves precision and F1-score over several stronger POI recommendation baselines while remaining computationally simple.