HASTE
A unified training framework for generative recommendation that introduces two additional supervision signals, a hierarchical loss and attribute-specific losses, to explicitly exploit the hierarchical structure of semantic IDs and inject item attribute information. This accelerates training convergence and improves recommendation of long-tail items.
Timeline (last 90 days)
HASTE is a unified generative recommendation training framework that, in addition to the original objective, introduces two types of extra supervision: Hierarchical Loss and Attribute-Specific Losses.
HASTE's hierarchical loss provides a supervision signal at every level of the item hierarchy, making the gradient signal denser and more directional, thereby significantly accelerating training convergence.
HASTE's attribute-specific losses leverage the semantic knowledge already present in the language model, injecting attribute information as a supervision signal into representation learning and improving the representation quality of tail items.
The research team evaluated on a fashion recommendation benchmark and multiple domains of the Amazon dataset, and HASTE brought recommendation performance improvements across different datasets and different item frequency ranges, with the long-tail range benefiting as well.
The experimental conclusions of the HASTE paper come from its own experimental setup, and the magnitude of gains in real business scenarios will vary with data distribution and model scale.
All Facts (5)
HASTE is a unified generative recommendation training framework that, in addition to the original objective, introduces two types of extra supervision: Hierarchical Loss and Attribute-Specific Losses.
50%UnverifiedHASTE's hierarchical loss provides a supervision signal at every level of the item hierarchy, making the gradient signal denser and more directional, thereby significantly accelerating training convergence.
50%UnverifiedHASTE's attribute-specific losses leverage the semantic knowledge already present in the language model, injecting attribute information as a supervision signal into representation learning and improving the representation quality of tail items.
50%UnverifiedThe research team evaluated on a fashion recommendation benchmark and multiple domains of the Amazon dataset, and HASTE brought recommendation performance improvements across different datasets and different item frequency ranges, with the long-tail range benefiting as well.
50%UnverifiedThe experimental conclusions of the HASTE paper come from its own experimental setup, and the magnitude of gains in real business scenarios will vary with data distribution and model scale.
50%