Revealing, Analyzing, and Mitigating the Adverse Impact of Shared Accounts on Personalized Recommendations for Single-User Accounts

Shared accounts are not uncommon in some applications, which can breed spurious associations between items that are actually consumed by distinct members. This may adversely affect the recommendations for other single-user accounts, which remains unexplored. To reveal this problem, we present a fair comparison of the recommendation performance for the same subset of single-user accounts under two settings: A controlled setting with only single-user accounts, and an experimental setting with both single-user accounts and shared accounts. The results show that the presence of shared accounts can lead to a performance decrease of up to 10%--40% for other single-user accounts. To gain deeper insights, we define two quantitative metrics to analyze the changes in the proximity of demoted relevant items and promoted irrelevant items to the user's consumed items in the latent space. The results show that the performance degradation can be attributed to the distortion of proximity relations between item embeddings. To mitigate the performance degradation, we leverage the multimodal information of items to construct a reliable item-item semantic graph which is fused with the bipartite behavior graph to counterbalance spurious second-order item-item associations arising from shared accounts. The results show that our mitigation method can effectively alleviate the performance degradation for other single-user accounts caused by shared accounts, recovering or even surpassing the recommendation performance in the controlled setting without shared accounts.