Abstract
Understanding users' interactions with highly subjective content—like artistic images—is challenging due to the complex semantics that guide our preferences. In this paper, we build large-scale recommender systems to model the dynamics of a vibrant digital art community, Behance, consisting of tens of millions of interactions (clicks and 'appreciates'). Our main contributions are modeling rich visual content, temporal dynamics (visual consistency within sessions), and social dynamics (preferences toward both art styles and specific artists).
Introduction
Recommending art differs significantly from recommending utilitarian goods. Preferences are driven by visual style, emotional resonance, and artist reputation. Traditional collaborative filtering methods often fail in this domain due to the extreme "cold-start" problem—where new art is uploaded without historical interaction data—and the complex social and temporal patterns inherent in art consumption.
Methodology
The proposed Vista (VIsually, Socially, and Temporally-aware) framework integrates three key dimensions to overcome these challenges:
- Social Dynamics: Vista explicitly models the relationship between the viewer and the creator (owner). By learning latent vectors for artists, the system captures "brand loyalty," allowing it to recommend new works by a known artist even before they receive any interactions.
- Temporal Consistency: Building on Factorized Personalized Markov Chains (FPMC), Vista incorporates higher-order Markov chains to capture short-term browsing "streaks." It uses a personalized decaying weight to account for varying user behaviors in maintaining visual consistency across a session.
- Visual Integration (Vista+): To address the cold-start problem, Vista+ utilizes a pre-trained VGG-16 CNN to extract deep visual features from project images. These features are projected into the model's latent space, enabling content-aware recommendations for brand-new items.
Experimental Results
The models were evaluated on two large-scale datasets from Behance: "Appreciates" (11.8M interactions) and "Clicks" (48.1M interactions).
Quantitative Performance
Vista and Vista+ consistently outperformed state-of-the-art baselines, including BPR-MF, FPMC, and VBPR. The performance gains were particularly pronounced in cold-start scenarios, where Vista+ improved recommendation accuracy by up to 30% for items with five or fewer interactions.
Qualitative Analysis
To understand what the model learns, the researchers visualized the latent space using t-SNE. The results show that the model naturally clusters items by genre—such as typography, culinary arts, and fashion—without any explicit category labels. This indicates that the combination of user behavior and visual features effectively captures the underlying semantics of the art.
Conclusion
Vista demonstrates that effective artistic recommendation requires a multi-faceted approach. By combining visual content analysis with social and temporal modeling, the framework successfully navigates the subjectivity of art, providing a robust solution for both established and cold-start items in creative communities.
Summary updated