<p>Conversational recommendation systems (CRS) are a critical application in human–computer interaction aimed at providing users with personalized and efficient recommendations. Although current research has explored using tree structures for reasoning to improve the causal relationships among entities in these systems, two main challenges remain unresolved. First, tree structures are often redundant, making it difficult to construct reasoning branches accurately. Additionally, the graph embeddings derived from knowledge graphs can suffer from information loss and ambiguity during iterative reasoning processes. To address these challenges, we propose the Optimized Path Reasoning System (OPRS) to enhance CRS. Initially, we apply emotional pruning to the tree structure to create a comprehensive preference tree that captures the user’s personalized preferences. Subsequently, we develop a Fourier fusion filtering mechanism that transforms embeddings from the time domain to the frequency domain, incorporating a band-pass filter to reduce noise. This approach aims to recover and compensate for potential information loss in the graph embeddings. Furthermore, we design a multi-path dialogue sampling enhancement module, integrating the concept of a variational autoencoder. By employing specialized Gaussian sampling of latent variables, we generate diverse dialogue paths within a continuous space. This method enables the dialogue system to comprehend better and respond to various scenarios, thereby enhancing the diversity and richness of the interactions. Extensive experiments have demonstrated the effectiveness of the OPRS approach.</p>