Qi, Yunzhe; Zhou, Qinghai; Liu, Boyang; Cui, Can; Zheng, Hua; Suo, Qiuling; Liang, Mingfu; Lov, Kenny; Liu, Xi; Jiang, Shali; Chen, Laming; Chen, Wen-Yen; Jin, Rong; Chang, Shuo; He, Jingrui; Li, Huayu. (2026).听.听Proceedings of the ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, 2-B, 4000鈥�4011.听
Artificial intelligence models used to predict whether people will click on online ads often perform best after one round of training, with additional training causing them to memorize noisy click data and perform worse on new examples. We developed TeMPO, a training framework that uses guidance from a recommendation foundation model to reduce this problem. Instead of learning only from click data, TeMPO also learns from the richer patterns captured by the foundation model, helping the model generalize better during repeated training. Theoretical analysis and experiments on public datasets and Meta鈥檚 large-scale advertising system showed that TeMPO allows additional rounds of training to improve rather than reduce performance. This approach could enable more effective multi-pass training of high-capacity recommendation models while maintaining their ability to generalize to new data.
