Impact of alleviating misinformation: an impulsive buying-aware model for sequential recommendation
Hongchen Wu, Xiaochang Fang, Hongxuan Li, Jie Sun, Jing Jing, Lin Zhang, Yihong Meng, Zhaorong Jing, Huaxiang Zhang · Expert Systems with Applications · 2025
Sequential recommendation systems are often misled by large-scale traffic data containing misinformation, which can trigger impulsive buying and reduce prediction accuracy. However, most existing methods either focus solely on long-term user preferences or assume a smooth evolution of user intentions, resulting in inadequate modeling and poor performance. This paper proposes an in-depth propaganda strategy by proposing an impulsive buying-aware model based on users’ long-term and short-term preference representation for sequential recommendation (INSPEQ), aiming to mitigate misinformation effects and enhance recommendation quality. INSPEQ first distinguishes long- and short-term preferences by extracting item attributes embedded in both intrinsic and extrinsic knowledge through relation paths. For long-term behavior modeling, a PATR-GRU network with apGRUs is used to learn stable user preferences by leveraging persistent item attributes and encoding temporal dependencies between adjacent and nonadjacent items along knowledge sequences, capturing both direct and indirect temporal influences. To capture short-term user preferences, a modified self-attention mechanism is introduced, enhanced with time-aware positional encoding, enabling the model to reflect recent shifts in user behavior more effectively within dynamic sessions. These dual representations are then adaptively fused via an MLP-based gated mechanism, assigning dynamic weights based on user impulsivity levels to flexibly balance stability and recency in decision-making. Extensive experiments on three real-world datasets demonstrate that INSPEQ consistently outperforms nineteen state-of-the-art methods. Specifically, it achieves up to +10.1 % in nDCG@5 and +11.4 % in HitRatio@5 over the strongest baselines, highlighting the effectiveness of jointly modeling preference dynamics while alleviating misinformation effects in practical recommendation environments.