Residual connections improve click-through rate and conversion rate prediction performance

The prediction of click-through rate (CTR) and conversion rate (CVR) are crucial tasks in online advertising and recommendation systems. As the learning models become more complex with increasing depth, it has become increasingly challenging to predict CTR and CVR accurately. This paper addresses the challenges associated with the increasing depth in CTR and CVR prediction models by introducing the integration of residual connections into the models. The experiments we conducted involve using five different CTR or CVR prediction models together with residual connections on benchmark datasets from both Avazu and Criteo, and the company dataset. The results demonstrate that residual connections can effectively improve CTR and CVR prediction models, with an increase in AUC by \(1.4\%\) $1.4\%$ , a decrease in loss by \(3.5\%\) $3.5\%$ , and an increase in F1 by \(20.2\%\) $20.2\%$ . The results also show that we can safely increase the depth and the size of the network without a need to optimize or a decrease in the performance.