SASRec sequential recommender

A self-attentive sequential recommender with multi-head attention, causal masking and residual feed-forward blocks for next-item ranking, tuned with hyperparameter sweeps and selected by offline metrics.

Sparse interaction histories make next-item prediction unstable; the goal was a compact model that holds up under sparsity.

Swept dropout 0.05 to 0.5, attention heads 1 to 2 and hidden dimensions 5 to 50, visualized stability across five variants and picked the best checkpoint by nDCG and hit rate.

1. 01

   nDCG@10 of 0.3166 and Hit@10 of 0.4913 at epoch 14 with a two-block, 50-dimensional configuration.