Construction of a Dynamic Evaluation System for Cross-Platform Social Media Marketing Effectiveness Integrating Federated Learning

Existing cross-platform social media marketing evaluation methods are vulnerable to selection bias, attribution error, and privacy risk, which restricts dynamic evaluation accuracy. This paper develops a privacy-preserving dynamic evaluation system for cross-platform social media marketing effectiveness based on federated learning. The system integrates advertisers as label holders and social platforms as feature providers, and applies private set intersection with hash encryption for secure user alignment. A vertical federated neural network combining secure multi-party computation and differential privacy is then used to aggregate hidden-layer outputs securely. A federated causal attribution mechanism based on Shapley values and counterfactual prediction dynamically quantifies the marginal contribution of each platform to conversion. The experimental data cover user behavior logs from Douyin, Weibo, and Xiaohongshu from January to December 2023, involving 40,000 cross-platform users and 5,862 conversion events, supplemented by a public digital marketing dataset for generalization validation. The proposed model achieves an AUC of 0.897 and attribution consistency of 0.766, outperforming FedAvg, FedTime, and FedDP by up to 46.5%. The information leakage risk probability is reduced to 0.184. These results demonstrate that the system improves the accuracy, interpretability, and privacy protection of cross-platform marketing evaluation.