Token-Weighted Multi-Target Learning for Generative Recommenders with Curriculum Learning

TL;DR

Generative recommender systems with semantic IDs suffer from token-level information imbalance when trained with standard next-token likelihood. The authors propose two information-gain token-weighting strategies—Front-Greater, which emphasizes early tokens based on prefix-based semantic dispersion, and Frequency weighting, which upweights rare tokens using the effective number of samples—and integrate them in a curriculum-guided multi-target framework that also includes the traditional likelihood objective. The three objectives are combined with learnable, normalized weights and a curriculum that shifts emphasis from coarse, frequent tokens to fine-grained rare tokens during training, enabling stable optimization. Across four benchmark datasets and alternative ID constructions (including PQ), the method yields consistent, significant gains on head and tail items, demonstrating robustness, generalization, and practical impact for generative recommendations with semantic IDs, while preserving model stability through curriculum-guided objective balancing. $L_{total} = \alpha_{fg} L_{fg} + \alpha_{fr} L_{fr} + \alpha_{or} L_{or}$, with curriculum-dependent $\alpha_j'$ guiding the learning dynamics.

Abstract

Generative recommender systems have recently attracted attention by formulating next-item prediction as an autoregressive sequence generation task. However, most existing methods optimize standard next-token likelihood and implicitly treat all tokens as equally informative, which is misaligned with semantic-ID-based generation. Accordingly, we propose two complementary information-gain-based token-weighting strategies tailored to generative recommendation with semantic IDs. Front-Greater Weighting captures conditional semantic information gain by prioritizing early tokens that most effectively reduce candidate-item uncertainty given their prefixes and encode coarse semantics. Frequency Weighting models marginal information gain under long-tailed item and token distributions, upweighting rare tokens to counteract popularity bias. Beyond individual strategies, we introduce a multi-target learning framework with curriculum learning that jointly optimizes the two token-weighted objectives alongside standard likelihood, enabling stable optimization and adaptive emphasis across training stages. Extensive experiments on benchmark datasets show that our method consistently outperforms strong baselines and existing token-weighting approaches, with improved robustness, strong generalization across different semantic-ID constructions, and substantial gains on both head and tail items. Code is available at https://github.com/CHIUWEINING/Token-Weighted-Multi-Target-Learning-for-Generative-Recommenders-with-Curriculum-Learning.

Figures (5)

• Figure 1: Average purity gain per token. For each token, purity gain is defined as the increase in prefix purity when the token is appended to a prefix. Purity is computed as $1 - \frac{H}{H_{\max}}$, where $H$ denotes the entropy of item frequencies under the prefix and $H_{\max} = \log(n)$, with $n$ being the number of items sharing the prefix. The reported value is the average purity gain of each token across all its occurrences. Most tokens contribute little purity gain, while a small subset plays a dominant role in identifying items.
• Figure 2: Comparison between the empirical item-occurrence distribution (ground-truth labels) and TIGER’s predicted item distribution on the Musical Instruments (Amazon'23) dataset amazon23. We split item frequency into ten quantiles (deciles) and count how many items fall into each quantile. TIGER exhibits a strong tendency to over-recommend popular items while under-representing rare items, reflecting the inherent bias in generative recommenders under long-tailed distributions.
• Figure 3: Head--tail evaluation based on item frequency in the test set. Items are split into head (top 50%) and tail (bottom 50%) groups by interaction frequency. Our method consistently outperforms TIGER on both sets across all datasets, with particularly strong improvements on tail items, highlighting the effectiveness of frequency-based token weighting in addressing long-tailed recommendation challenges.
• Figure 4: Performance of our method on different values of $c$, which controls how quickly the curriculum transitions during training.
• Figure 5: The variation in weights across different objectives on Yelp dataset with $c=2e-5$.

Theorems & Definitions (2)

• lemma 1: Prefix Tokens Reduce Semantic Dispersion
• lemma 2: Token Frequency Induces Diminishing Marginal Information Gain