Abstract
Systems and methods are described for automatically discovering user intent categories in conversion rate prediction using differentiable discrete routing. A routing network produces logits over K expert prediction heads based on interaction features such as click type, page type, and position. During training, a Gumbel-Softmax reparameterization generates differentiable approximations of discrete expert assignments, optionally with temperature annealing to transition from soft sharing to near-discrete specialization. A load balancing loss, such as a KL-divergence between an average routing distribution and a uniform distribution, discourages expert collapse. During inference, routing may collapse to a hard argmax such that exactly one expert head is evaluated per impression. After training, the routing function may be probed on synthetic inputs to export an interpretable taxonomy mapping feature combinations to discovered intent categories for downstream use.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Automatic Discovery of User Intent Categories in Conversion Rate Prediction Using Differentiable Discrete Routing", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10637