Customer Preference Learning with Zero-Initialized Distance and Embedding Fusion (ZIP-EF)

Abstract
Recommender systems play a key role in online shopping and constitute a major component of user interaction on e-commerce platforms. Classical recommendation approaches rely on fixed similarity metrics, such as cosine similarity or Euclidean distance between textual and visual embeddings, which limits their ability to adapt to dynamic and personalized user preferences. This paper proposes a multimodal recommendation framework in which the distance function between user and product embeddings is fully learnable. Textual and visual embeddings are combined using element-wise learnable fusion, and a bidirectional multilayer perceptron with GELU activation projects embeddings into a preference-oriented latent space. An interaction layer extracts four complementary representations, including user features, item features, their element-wise difference, and their element-wise product, following neural collaborative filtering paradigms. The metric layer is initialized to exactly reproduce cosine similarity, ensuring stable performance at initialization, and is subsequently adapted through pairwise preference learning. User click behavior is transformed into preference pairs, enabling the model to assign higher similarity to preferred items and lower similarity to ignored ones. Experimental results demonstrate that the proposed approach achieves stable and progressive personalization, providing greater adaptability to real user behavior than fixed-metric recommendation methods.
Keywords
Multimodal Recommendation
Learnable Similarity Metric
Pairwise Preference Learning
Element-wise Fusion
User Preference Modeling
Metric Learning
E-commerce Recommendation
[1]    C. J. C. Burges, “From RankNet to LambdaRank to LambdaMART: An overview,” Microsoft Research, Redmond, WA, USA, Tech. Rep., 2010.
[2]    C. J. C. Burges, T. Shaked, E. Renshaw, A. Lazier, M. Deeds, N. Hamilton, and G. Hullender, “Learning to rank using gradient descent,” in Proc. 22nd Int. Conf. Mach. Learn. (ICML), 2005, pp. 89–96.
[3]    X. Chen, Y. Li, C. Li, M. Zhang, S. Ma, and S. Wang, “A survey on reinforcement learning for recommender systems,” IEEE Trans. Neural Netw. Learn. Syst., vol. 32, no. 6, pp. 2142–2165, Jun. 2021.
[4]    P. Covington, J. Adams, and E. Sargin, “Deep neural networks for YouTube recommendations,” in Proc. 10th ACM Conf. Recommender Syst. (RecSys), 2016, pp. 191–198.
[5]    C. Gao, P. Zhao, Y. Li, Z. Zhang, and Y. Yu, “Multimodal representation learning for recommendation: A survey,” ACM Comput. Surv., vol. 55, no. 13, pp. 1–38, 2023.
[6]    R. He and J. McAuley, “VBPR: Visual Bayesian personalized ranking from implicit feedback,” in Proc. 30th AAAI Conf. Artif. Intell. (AAAI), 2016, pp. 144–150.
[7]    X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua, “Neural collaborative filtering,” in Proc. 26th Int. World Wide Web Conf. (WWW), 2017, pp. 173–182.
[8]    X. He, K. Zhang, M.-Y. Kan, and T.-S. Chua, “LightGCN: Simplifying and powering graph convolution network for recommendation,” in Proc. 43rd Int. ACM SIGIR Conf. Res. Develop. Inf. Retrieval (SIGIR), 2020, pp. 639–648.
[9]    P.-S. Huang, X. He, J. Gao, L. Deng, A. Acero, and L. Heck, “Learning deep structured semantic models for web search using clickthrough data,” in Proc. 22nd ACM Int. Conf. Inf. Knowl. Manage. (CIKM), 2013, pp. 2333–2338.
[10] W.-C. Kang and J. McAuley, “Self-attentive sequential recommendation,” in Proc. IEEE Int. Conf. Data Mining (ICDM), 2018, pp. 197–206.
[11] L. Ouyang et al., “Training language models to follow human instructions with human feedback,” arXiv preprint arXiv:2203.02155, 2022.
[12] S. Rendle, C. Freudenthaler, Z. Gantner, and L. Schmidt-Thieme, “BPR: Bayesian personalized ranking from implicit feedback,” in Proc. 25th Conf. Uncertainty Artif. Intell. (UAI), 2009, pp. 452–461.
[13]    Y. Wei, A. Zhang, X. Zhang, and S. Feng, “MMGCN: Multi-modal graph convolution network for personalized recommendation,” in Proc. 28th Int. Joint Conf. Artif. Intell. (IJCAI), 2019, pp. 3372–3378.
[14]  K. Zhou, H. Wang, W. X. Zhao, and J.-R. Wen, “MMRec: Multimodal recommendation via multimodal alignment,” in Proc. 45th Int. ACM SIGIR Conf. Res. Develop. Inf. Retrieval (SIGIR), 2022, pp. 1052–1062.
[15] L. Zou, S. Zhang, L. Li, S. Liu, and J. Li, “Reinforcement learning to optimize long-term user engagement in recommender systems,” in Proc. 28th Int. Joint Conf. Artif. Intell. (IJCAI), 2019, pp. 4308–4314.
Marketing Science and Technology Journal
Volume 3, Issue 4
Winter 2026
Pages 96-119