When Kafka is used in production environments, a large number of parameters are provided to facilitate user configuration for specific application environments in order to obtain better performance. However, configuring Kafka's parameters requires in-depth knowledge of the user, which is far beyond the ability of the average user and prevents Kafka from obtaining better performance. To address this problem, we propose an ACT-SAGAN method that adds a self-attention mechanism to the generative adversarial network model to capture the associations between hidden structures in good configuration combinations and configuration parameters, which uses these hidden structures and associations to generate better configuration combinations to improve Kafka's performance. Experimental results show that the algorithm improves Kafka's throughput and reduces latency after deployment for the configuration combinations generated by Kafka.