Constructing a Joint Prediction Model for Advertising Click-Through Rate and Conversion Rate Based on a Multi-Task Learning Framework

Click-through rate and conversion rate prediction in online advertising systems are constrained by sample bias and selection bias because conversion-rate training is usually conducted only on clicked samples. To address this problem, this paper constructs a multi-task learning framework for joint CTR and CVR modeling. User, advertisement, and contextual features are first modeled through a shared representation layer. A CTR branch is then designed to generate a stable click signal to assist CVR prediction. A causal bias correction module based on inverse propensity scoring and weighted loss is introduced to correct the distribution discrepancy between click samples and impression samples. Experimental results show that the proposed framework achieves an AUC of 0.912 for joint CTR/CVR prediction, compared with 0.828 for the baseline model. LogLoss decreases to 0.365 and 0.381 for the corresponding tasks, while NDCG@1 for CTR improves to 0.538. Online experiments show CTR and CVR improvement rates of 10.2% and 9.6%, respectively, and revenue increases by 9.9%, demonstrating effective mitigation of sample and selection bias.