DisenReason: Behavior Disentanglement and Latent Reasoning for Shared-Account Sequential Recommendation

A research team led by Jiawei Cheng and Min Gao from multiple Chinese universities has introduced DisenReason, a breakthrough AI model designed to solve the persistent problem of shared-account recommendations on streaming and e-commerce platforms. Traditional methods assume a fixed number of users per shared account, severely limiting accuracy when families or groups use services like Netflix, Amazon Prime, or Spotify under one login. DisenReason fundamentally reframes the problem by shifting from 'inferring preferences behind a user' to 'inferring the users behind an account,' using a novel two-stage approach that first disentangles collective account behavior before reasoning about individual latent users.

The technical innovation lies in DisenReason's frequency-domain behavior disentanglement stage, which creates a unified account representation, followed by a latent user reasoning stage that dynamically infers how many people are actually using the account. This allows the model to adapt to real-world scenarios where sharing patterns are fluid and unpredictable. In rigorous testing across four benchmark datasets, DisenReason consistently outperformed all state-of-the-art baselines, achieving relative improvements of up to 12.56% in MRR@5 (Mean Reciprocal Rank) and 6.06% in Recall@20. The research, detailed in a 17-page arXiv paper, represents a significant advance in sequential recommendation systems, with immediate applications for platforms struggling with accurate personalization in shared environments.