Joint dialog sentiment classification and act recognition focus on the simultaneous identification of sentiment and act categories of every utterance from conversations. Current classification-based methods often rely on complex structures to model label dependencies, which can be sensitive to noisy features in dialogue contexts. In light of this, we introduce an end-to-end generative framework, termed Sentiment and Act T5 (SAT5), specifically designed to address this joint task. Our SAT5 framework simply leverages the inherent ability of generative models to adequately model label dependencies without additional structures. Meanwhile, considering the order sensitivity issues in generative models, a set loss mechanism is further incorporated to eliminate these concerns effectively. Additionally, we develop a graph-based feature optimization strategy grounded in the information bottleneck principle, aimed at minimizing the impact of noisy features. Experiments on two public datasets demonstrate that our SAT5 framework significantly outperforms previous models. In-depth analysis further validates the efficacy and rationality of our set loss mechanism and feature optimization strategy.