Hybrid feedback fusion with multi‑head cross-attention for top‑N recommendation in E-learning systems

Learner ratings in massive open online courses are sparse and, where they exist, overwhelmingly positive, while behavioural traces such as opening, replaying and returning to a resource are far more plentiful and are usually discarded. We set out to test whether fusing the two channels through cross-attention improves recommendation on the Mandarine Academy Recommender System (MARS) corpus. We report the model, and five results about measurement that we consider more consequential than the model. First, the hybrid construction leaks. Because 99.4% of held-out test resources also appear in the same learner’s implicit log, an unguarded frequency-recency-session signal returns the evaluation target to the model’s own input; at the originally submitted configuration, closing the leak with four explicit guards lowers our own HR@10 from 0.910 to 0.635 under the conventional sampled protocol and from 0.526 to 0.206 under full-catalogue ranking. Second, the tuning budget decides the comparison: giving every tunable model the same twelve-trial validation search raised nine of the ten learned baselines, by up to 0.077 in full-catalogue HR@10, and moved our model from fourth of thirteen to eighth of twelve. Third, the sampled protocol that dominates this literature, one positive against ninety-nine sampled negatives, orders models differently from full-catalogue ranking (Spearman 0.664, p = 0.018). Fourth, rating-prediction error carries no usable information about ranking quality here (Spearman − 0.077 against full-catalogue HR@10, p = 0.81), and a constant predictor returning the training-set mean rating achieves RMSE 3.320, lower than that of every model. Fifth, a silent column-name mismatch had left most of our own content-feature matrix empty while its shape remained correct; correcting it reversed two conclusions we had drawn. Under the guarded protocol, with random tie-breaking and Holm-corrected paired tests, Mult-VAE attains the best full-catalogue HR@10 at 0.269 ± 0.003 and HAH-Rec reaches 0.214 ± 0.004, separated from six of the seven models above it. Its ablation shows that content metadata, not the implicit channel, supplies what gain there is, and that cross-attention adds nothing measurable. Text representations of the resources match the curated metadata at best: TF-IDF matches it, a sentence encoder trained for paraphrase similarity approaches it, and averaged word vectors and a transformer encoder fine-tuned for syntax fall below popularity ranking as content-only models and cost HAH-Rec accuracy inside it. The attention weights are weakly correlated with the measured importance of the resources they weight. What survives for the architecture is narrow: its content tower recommends resources withheld from training entirely, where popularity ranking scores zero, though a pure content model on the same metadata does so at least as well. We release the guarded pipeline so that the leakage and tuning effects we describe can be checked for on other corpora.