Personalized diffusions for top-n recommendation

Abstract

This paper introduces PerDif; a novel framework for learning personalized difusions over item-to-item graphs for top-n recommendation. PerDif learns the teleportation probabilities of a timeinhomogeneous random walk with restarts capturing a user-specifc underlying item exploration process. Such an approach can lead to signifcant improvements in recommendation accuracy, while also providing useful information about the users in the system. Per-user ftting can be performed in parallel and very efciently even in large-scale settings. A comprehensive set of experiments on real-world datasets demonstrate the scalability as well as the qualitative merits of the proposed framework. PerDif achieves high recommendation accuracy, outperforming state-of-the-art competing approaches-including several recently proposed methods relying on deep neural networks. CCS CONCEPTS• Information systems → Recommender systems.

Trust your neighbors: A comprehensive survey of neighborhood-based methods for recommender systems

“…From a user modeling point of view this translates to the implicit assumption that every user explores the itemspace in exactly the same way-overlooking the reality that different users can have different behavioral patterns. The fundamental premise of PerDif [57] is that the latent item exploration behavior of the users can be captured better by user-specific preference propagation mechanisms; thus, leading to improved recommendations.…”

“…Leveraging the special properties of the stochastic matrix G the above non-linear optimization problem can be solved efficiently. In particular, it can be shown [57] that the optimization problem ( 59) is equivalent to minimize…”

“…In particular, the task of finding personalized probabilities for the item exploration process, reduces to that of finding personalized diffusion coefficients ω u over the space of the first K landing probabilities of a walk rooted on φ u (see definition of S u ). Afterwards µ u can be obtained in linear time from ω u upon solving a simple forward recurrence [57]. Taking into account the fact that in recommendation settings K will typically be small and φ u , P sparse, building 'on-the-fly' S u E u row-by-row, and solving the (K + 1)-dimensional convex quadratic problem PerDif free : minimize can be performed very efficiently (typically in a matter of milliseconds even in large scale settings).…”

“…In particular, the dual interpretation of the model parameters (µ u in the item exploration space; and, ω u in the diffusion space) allows utilizing the learned model parameters to identify users for which the model will most likely lead to poor predictions, at training-time-thereby affording preemptive interventions to handle such cases appropriately. This affords a level of transparency that can prove particularly useful in practical settings (for the technical details on how this can be achieved see [57]).…”

Heterogeneous Peer Effects in the Linear Threshold Model

“…Emotions such as happiness (Fowler and Christakis 2008) and hatefulness (Ribeiro et al 2018) have also been observed to spread through social networks. These processes, known as information diffusion, have traditionally been studied in the social sciences (Granovetter 1978), but more recently have motivated applications in viral marketing (Kempe, Kleinberg, and Tardos 2003) and recommender systems (Nikolakopoulos et al 2019). Many models exist that capture the diffusion process, including epidemic models (Kermack and McKendrick 1927), voter models (Clifford and Sudbury 1973), the Independent Cascade (Kempe, Kleinberg, and Tardos 2003), and the Linear Threshold Model (LTM) (Granovetter 1978).…”