PERSONALIZATION · 2026-11-17
Personalization — Paper Daily, 2026-11-17
Today's personalization track asks how a model should spend limited capacity on the signals that actually matter. One paper looks at on-device sequential recommendation and shows that a single compression ratio wastes compute on simple requests and starves complex ones, so it allocates channels and history length by how stable the ranking is. The other returns to feature crossings in tabular models. An ordinary multilayer perceptron buries combinations in hidden layers, and the authors write the product into residual blocks, then check on scientific regression benchmarks whether the model spends its strength on known crossings. For recommender practitioners, the first is about adjusting on-device ranking capacity per request, and the second is a reminder that explicit crossings are not only for lowering error, but also for making it possible to inspect which combinations the model actually captured.
4.09
On-device sequential recommenders often compress every request with the same ratio, so simple interests waste compute and complex ones do not get enough capacity. Pruning and decomposition usually freeze the model and the input at a fixed size, and cannot adjust capacity to how hard the current behavior sequence is.
3.29
When a tabular model must predict an outcome that depends on crossings among several factors, an ordinary multilayer perceptron often buries those crossings in hidden layers. The error may still look acceptable, but it is hard to tell whether the model captured the combinations that actually matter. Factorization machines and cross networks in recommenders also multiply features explicitly. This paper writes the product into residual blocks and uses complex weights to represent magnitude and phase together, which steadies optimization. On the classic Friedman function, root-mean-square error is 53.9% lower than that of an ordinary multilayer perceptron. Its interaction map concentrates 60.7% of total strength on the known pair, compared with 31.0% for the ordinary network. Predictions improve because the model spends its capacity on the real crossings. These comparisons are on scientific regression benchmarks. For click-through or conversion models that also rely on feature crossings, explicit products plus residuals offer a structure one can inspect, rather than leaving combinations entirely to hidden layers.