Trust-Aware Social Recommendation with Cosine Similarity, Temporal Decay, and Location Proximity

Abstract

Social recommendation systems leverage interpersonal trust to improve prediction accuracy beyond conventional col- laborative filtering. However, most existing trust-aware models treat all interactions equally, ignoring the fine-grained similarity between users’ preference profiles, the temporal dynamics of evolving tastes, and the geographic context of real-world ordering behaviour. In this paper we propose a Trust-Aware Social Recommendation framework that integrates four complementary sig- nals: (i) co-visitation frequency as a structural trust proxy, (ii) cosine similarity computed on user–item interaction vectors for preference alignment, (iii) an exponential temporal decay function that emphasises recent behaviour, and (iv) geographic loca- tion proximity that captures the tendency of co-located users to share restaurant preferences. We formulate a composite trust score T (u, v) = ?C(u, v) + ?cos(mu, mv) + ?prox(u, v) and introduce a scalable incremental graph-patching evaluation protocol that reduces the per-user cost of Leave-One-Out evaluation from O(U 2) to O(Ur), where Ur ? U. Ablation experiments on a location-biased synthetic restaurant-ordering dataset of 7.5 million orders across 300,000 users and 5,000 restaurants demonstrate consistent improvements: the best trust-based model (Trust + Cosine) achieves a 135% relative gain in NDCG@5 over the popular- ity baseline when evaluated on a stratified sample of 5,000 users. Additional experiments on hyperparameter sensitivity, cold-start users (3–5 orders), and cross-zone user mobility further validate the model’s scalability and practical applicability to large-scale settings.

Introduction

The text discusses the limitations of traditional collaborative filtering (CF) in recommendation systems, particularly issues like cold-start and data sparsity. To overcome these, social recommendation methods incorporate trust relationships between users, improving prediction accuracy. However, existing approaches often fail to jointly consider three key factors: preference similarity, temporal dynamics, and geographic context.

To address these gaps, the paper proposes a unified recommendation framework that combines:

• Co-visitation frequency (shared items),
• Cosine similarity (preference alignment),
• Temporal decay (recency of interactions),
• Geographic proximity (location-based behavior).

These factors are integrated into a single trust score, which is used to build a user social graph and generate personalized recommendations. The system also introduces a scalable evaluation method using incremental graph patching, reducing computational complexity and enabling large-scale deployment.

The framework is tested using a synthetic dataset, created to simulate realistic user behavior with control over location, time, and cold-start scenarios—overcoming limitations of public datasets.

The methodology includes constructing a trust-based user graph, computing multiple trust formulations (basic, cosine-enhanced, time-decayed, and location-aware), and ranking items based on neighbor influence. The system uses additive trust composition for flexibility and efficiency.

Overall, the proposed approach demonstrates that combining multiple signals—social, temporal, and geographic—leads to more accurate and scalable recommendation systems, especially in location-sensitive domains like food delivery.

Conclusion

We presented a trust-aware social recommendation framework that progressively integrates co-visitation frequency, cosine simi- larity, temporal decay, and geographic location proximity into a unified trust formulation. Through a comprehensive experimental evaluation using Leave-One-Out on a large-scale location-biased synthetic dataset of 300 000 users, 5 000 restaurants, and 7.5 mil- lion orders, we demonstrated that: 1) Social trust derived from co-visitation patterns substantially outperforms a popularity baseline at large scale, with Trust + Cosine achieving a 135% relative NDCG@5 improvement. 2) Cosine similarity over user–item interaction vectors is the most effective trust augmentation at large scale, providing normalised preference alignment that handles diverse user activity levels. 3) Temporal decay and geographic proximity show diminishing returns at large scale due to data sparsity, suggesting that component weighting should be scale-adaptive. 4) The cold-start experiment reveals that trust-based strategies require sufficient co-visitation overlap, motivating adaptive threshold approaches for users with limited interaction history. 5) Hyperparameter sensitivity analysis shows stable performance across wide parameter ranges, with ? (cosine weight) being the most impactful and ? (location weight) the most robust. 6) The incremental graph-patching evaluation strategy, combined with evaluation sampling, enables the framework to scale to 300K+ user populations with complete experiment suites running in under 15 minutes.

References

[1] P. Resnick and H. R. Varian, “Recommender systems,” Commun. ACM, vol. 40, no. 3, pp. 56–58, 1997. [2] G. Adomavicius and A. Tuzhilin, “Toward the next generation of recommender systems: A survey of the state-of-the-art and possible extensions,” IEEE Trans. Knowl. Data Eng., vol. 17, no. 6, pp. 734–749, 2005. [3] Y. Koren, R. Bell, and C. Volinsky, “Matrix factorization techniques for recommender systems,” Computer, vol. 42, no. 8, pp. 30–37, 2009. [4] Y. Koren, “Collaborative filtering with temporal dynamics,” Commun. ACM, vol. 53, no. 4, pp. 89–97, 2010. [5] B. Sarwar, G. Karypis, J. Konstan, and J. Riedl, “Item-based collaborative filtering recommendation algorithms,” in Proc. 10th Int. Conf. World Wide Web (WWW), 2001, pp. 285–295. [6] P. Massa and P. Avesani, “Trust-aware recommender systems,” in Proc. ACM Conf. Recommender Systems (RecSys), 2007, pp. 17–24. [7] M. Jamali and M. Ester, “A matrix factorization technique with trust propagation for recommendation in social networks,” in Proc. ACM Conf. Recommender Systems (RecSys), 2010, pp. 135–142. [8] H. Ma, D. Zhou, C. Liu, M. R. Lyu, and I. King, “Recommender systems with social regularization,” in Proc. 4th ACM Int. Conf. Web Search Data Mining (WSDM), 2011, pp. 287–296. [9] B. Yang, Y. Lei, J. Liu, and W. Li, “Social collaborative filtering by trust,” IEEE Trans. Pattern Anal. Mach. Intell., vol. 39, no. 8, pp. 1633–1647, 2017. [10] G. Guo, J. Zhang, and N. Yorke-Smith, “TrustSVD: Collaborative filtering with both the explicit and implicit influence of user trust and of item ratings,” in Proc. AAAI Conf. Artif. Intell., 2015, pp. 123–129. [11] X. He, L. Liao, H. Zhang, L. Nie, X. Hu, and T.-S. Chua, “Neural collaborative filtering,” in Proc. 26th Int. Conf. World Wide Web (WWW), 2017, pp. 173–182. [12] B. Hidasi, A. Karatzoglou, L. Baltrunas, and D. Tikk, “Session-based recommendations with recurrent neural networks,” in Proc. 4th Int. Conf. Learning Representations (ICLR), 2016. [13] J. Li, P. Ren, Z. Chen, Z. Ren, T. Lian, and J. Ma, “Neural attentive session-based recommendation,” in Proc. ACM Int. Conf. Inf. Knowl. Manage. (CIKM), 2017, pp. 1419–1428. [14] X. Wang, Y. Lu, B. Shi, and H. Li, “Temporal-aware trust-based recommendation,” in Proc. ACM Int. Conf. Inf. Knowl. Manage. (CIKM), 2016, pp. 393–402. [15] T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks,” in Proc. 5th Int. Conf. Learn- ing Representations (ICLR), 2017. [16] X. He, K. Deng, X. Wang, Y. Li, Y. Zhang, and M. Wang, “LightGCN: Simplifying and powering graph convolution network for recommendation,” in Proc. 43rd Int. ACM SIGIR Conf. Res. Dev. Inf. Retrieval, 2020, pp. 639–648. [17] M. Ye, P. Yin, W.-C. Lee, and D.-L. Lee, “Exploiting geographical influence for collaborative point-of-interest recommenda- tion,” in Proc. 34th Int. ACM SIGIR Conf. Res. Dev. Inf. Retrieval, 2011, pp. 325–334. [18] J.-D. Zhang, C.-Y. Chow, and Y. Li, “GeoSoCa: Exploiting geographical, social and categorical correlations for point-of- interest recommendations,” in Proc. 38th Int. ACM SIGIR Conf. Res. Dev. Inf. Retrieval, 2015, pp. 443–452. [19] K. Järvelin and J. Kekäläinen, “Cumulated gain-based evaluation of IR techniques,” ACM Trans. Inf. Syst., vol. 20, no. 4, pp. 422–446, 2002. [20] S. Rendle, W. Krichene, L. Zhang, and J. Anderson, “Neural collaborative filtering vs. matrix factorization revisited,” in Proc. 14th ACM Conf. Recommender Systems (RecSys), 2020, pp. 240–248.

Copyright

Copyright © 2026 Manvith Sai Yarlagadda, Harsha Dodda, Pardhasaradhi Pujala, Yusuf Vali Shaik, Dr. Vinod K. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.