E-Learning Course Recommendation Based on Employee Interaction History and Implicit Preferences: A Protocol-Aware Comparison of Sequential Recommendation Architectures

Large corporate e-learning catalogs make relevant courses hard for employees to find. This single-organization case study compares six sequential models for next-course recommendation from interaction histories: an Encoder-Decoder LSTM with Temporal Attention, its no-attention and Keras multi-head-attention variants, and GRU4Rec and SASRec, each with and without the same temporal-attention pooling. Each model family receives the same 30-trial Optuna search and is evaluated over ten training seeds on 4,553 (default protocol) and 4,551 (strict, train-history-only protocol) temporally isolated test samples. Under equal tuning, the six backbones are statistically indistinguishable on default-protocol Hit@10 (92.6% to 92.8%). On the strict protocol, SASRec outperforms the LSTM by 0.75 percentage points (95% CI 0.35 to 1.16, Holm-adjusted p = 0.004), but only under a dual-protocol tuning objective (E7b) adopted after inspecting the test results of a single-protocol objective (E7); under E7 the advantage becomes a non-significant 0.15-point deficit. Temporal attention produces no Holm-significant Hit@10 change on any backbone. All models outperform a first-order Markov-chain baseline (by 6.0 to 6.1 points, default; 33 to 34 points, strict) and a popularity baseline. A fixed five-course curriculum followed by 27.3% of test users inflates every raw score; excluding these users leaves every comparison intact. A forward-in-time audit, run on validation users and not a retraining of the main models, gives an absolute Hit@10 of only about 0.51. The 4,795 users (49.96%) with fewer than four interactions were excluded, so the findings do not cover cold-start recommendation, and offline ranking metrics do not measure learning or workforce outcomes.