Multimodal customer intent recognition and profile dynamic update network based on temporal attention and gated fusion

Accurate recognition of customer intent, together with the timely maintenance of user profiles, underpins a range of downstream applications including personalized recommendation, customer segmentation, targeted promotion, customer relationship management (CRM) and customer service quality assurance. Three practical limitations continue to constrain existing systems: multimodal fusion strategies remain insufficiently adaptive to per-sample modality reliability, temporal behavior models tend to collapse short-term intent shifts and long-term preference drift into a single flat sequence, and intent recognition pipelines are typically decoupled from downstream profile update mechanisms, so newly detected signals only propagate to profiles after a batch cycle. To address these gaps in a unified way, this paper proposes the Multimodal Customer Intent Recognition and Profile Dynamic Update Network (MCIP-DUN).

The framework combines a pre-trained language backbone with a cross-modal gated fusion unit that reweights textual, behavioral and interaction-attribute streams per sample, a time-decay-aware temporal attention module that separates short-term tactical shifts from long-term preference evolution, and a gated recurrent profile update module in which the forget gate and update gate are decoupled so that stable dimensions and volatile dimensions can be governed independently. A joint training objective with homoscedastic uncertainty weighting balances the intent classification loss against a bounded-target profile reconstruction loss. On the Alibaba User Behavior Dataset (AUBD) and the Customer Service Dialog Intent Dataset (CSDID), MCIP-DUN reaches 89.46% accuracy and 87.77% macro-F1, exceeding the strongest baseline by 2.59 and 3.25% points; the profile update pipeline achieves 88.39% tag accuracy with an end-to-end persistence lag of roughly 18 min — measured from event ingestion to database refresh, and clearly distinguished from the 14.1 ms model forward latency reported separately.

Ablation studies isolate the individual contribution of each component. It should be noted that the evaluation is restricted to e-commerce and customer-service data; generalization to other verticals such as power-utility customer service is discussed as future work. Multimodal Customer Intent Recognition and Profile Dynamic Update Network Bidirectional Encoder Representations from Transformers Robustly Optimized BERT Pretraining Approach Self-Attentive Sequential Recommendation Term Frequency–Inverse Document Frequency Guangdong Power Grid Co., Ltd.

Zhuhai Power Supply Bureau, Zhuhai, 519000, Guangdong, China Yunlin Yang, Guosheng Lei, Junning Guan, Ya Pan & Wanwen Zheng The authors declare no competing interests. This study did not involve direct interaction with human subjects. All experiments were conducted exclusively on two publicly available, pre-existing, and fully anonymized datasets—the Alibaba User Behavior Dataset (AUBD) and the Customer Service Dialog Intent Dataset (CSDID)—neither of which contains personally identifiable information.

The Research Ethics Committee of Zhuhai Power Supply Bureau of Guangdong Power Grid Co., Ltd. reviewed the study protocol (Reference Number: IRB-2024-08173) and confirmed that formal ethical approval was not required given the use of publicly released anonymized data. The study was conducted in accordance with applicable institutional guidelines and national regulations on data protection. All authors have reviewed the manuscript and consent to its publication.

No identifiable information regarding participants has been included. Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made.