Lottery4CVR: Neuron-Connection Level Sharing for Multi-task Learning in Video Conversion Rate Prediction

Xuanji Xiao, Jimmy Chen, Yuzhen Liu, Xing Yao, Pei Liu, Chaosheng Fan

Abstract: As a fundamental task of industrial ranking systems, conversion rate (CVR) prediction is suffering from data sparsity problems. Most conventional CVR modeling leverages Click-through rate (CTR) &CVR multitask learning because CTR involves far more samples than CVR. However, typical coarse-grained layer-level sharing methods may introduce conflicts and lead to performance degradation, since not every neuron or neuron connection in one layer should be shared between CVR and CTR tasks. This is because users may have different fine-grained content feature preferences between deep consumption and click behaviors, represented by CVR and CTR, respectively. To address this sharing &conflict problem, we propose a neuron-connection level knowledge sharing. We start with an over-parameterized base network from which CVR and CTR extract their own subnetworks. The subnetworks have partially overlapped neuron connections which correspond to the sharing knowledge, and the task-specific neuron connections are utilized to alleviate the conflict problem. As far as we know, this is the first time that a neuron-connection level sharing is proposed in CVR modeling. Experiments on the Tencent video platform demonstrate the superiority of the method, which has been deployed serving major traffic. (The source code is available at https://github.com/xuanjixiao/onerec/tree/main/lt4rec).