[Paper Review] TribeFlow: Mining & Predicting User Trajectories

Flávio Figueiredo, Bruno Ribeiro+2 more|arXiv (Cornell University)|Nov 3, 2015

Music and Audio Processing59 references4 citations

TL;DR

TribeFlow is a scalable, parameter-free method that models non-stationary, transient, and time-heterogeneous user trajectories by decomposing them into short random walks over latent, stationary, and time-homogeneous environments. It achieves up to 900% higher accuracy and 413× faster performance than state-of-the-art methods like MultiLME and FPMC in next-item prediction tasks.

ABSTRACT

Which song will Smith listen to next? Which restaurant will Alice go to tomorrow? Which product will John click next? These applications have in common the prediction of user trajectories that are in a constant state of flux over a hidden network (e.g. website links, geographic location). What users are doing now may be unrelated to what they will be doing in an hour from now. Mindful of these challenges we propose TribeFlow, a method designed to cope with the complex challenges of learning personalized predictive models of non-stationary, transient, and time-heterogeneous user trajectories. TribeFlow is a general method that can perform next product recommendation, next song recommendation, next location prediction, and general arbitrary-length user trajectory prediction without domain-specific knowledge. TribeFlow is more accurate and up to 413x faster than top competitors.

Motivation & Objective

• To address the challenge of predicting user trajectories in non-stationary, transient, and time-heterogeneous environments.
• To develop a general-purpose method for next-item prediction without requiring domain-specific features or time-of-day effects.
• To enable accurate and scalable personalized recommendations across diverse applications such as music, product, and location prediction.
• To overcome limitations of tensor decomposition and Markov-based models in capturing asynchronous, evolving user behavior.

Proposed method

• TribeFlow decomposes user trajectories into short random walks over latent, stationary, and time-homogeneous item-item graphs (latent environments).
• It uses a nonparametric hierarchical Bayesian model to infer latent environments and user preferences without parameter tuning.
• The model captures user behavior through weighted jumps between latent environments, enabling dynamic adaptation to changing interests.
• It performs next-item prediction by computing a posterior over active latent environments based on user history and simulating random walks.
• The method is fully parallelizable, enabling high-speed inference even on large-scale datasets.
• It integrates user history to infer current latent environment and predict the next item via random walk sampling.

Experimental results

Research questions

• RQ1 Can a general method predict user trajectories across diverse domains without domain-specific features or time-of-day effects?
• RQ2 How well can a model capture non-stationary, transient, and time-heterogeneous behavior in user trajectories compared to tensor-based and Markov-based methods?
• RQ3 What is the trade-off between accuracy and scalability in next-item prediction for large-scale user trajectory data?
• RQ4 Can latent environments be effectively inferred from raw user trajectories alone, without timestamps or external features?

Key findings

• TribeFlow achieves up to 900% higher accuracy than state-of-the-art methods like MultiLME, PRLME, and FPMC in next-item prediction tasks.
• It is up to 413× faster than the fastest competitor, MultiLME, and completes tasks in hours instead of days.
• On synthetic data with asynchronous users, TribeFlow perfectly recovers underlying Markov chain structures, while tensor decomposition fails to identify user-specific patterns.
• TribeFlow outperforms TM-LDA and tensor factorization methods in capturing latent user behavior, especially in transient and non-stationary settings.
• The method is robust without parameter tuning, demonstrating strong generalization across datasets including YooChoose, LFM-1k, and FourSquare.
• In sub-sampled datasets, TribeFlow is at least 23% more accurate than its competitors, confirming consistent superiority in predictive performance.