Predicting human preferences using the block structure of complex social networks

With ever-increasing available data, predicting individuals’ preferences and helping them locate the most relevant information has become a pressing need. Understanding and predicting preferences is also important from a fundamental point of view, as part of what has been called a “new” computational social science. Here, we propose a novel approach based on stochastic block models, which have been developed by sociologists as plausible models of complex networks of social interactions. Our model is in the spirit of predicting individuals’ preferences based on the preferences of others but, rather than fitting a particular model, we rely on a Bayesian approach that samples over the ensemble of all possible models. We show that our approach is considerably more accurate than leading recommender algorithms, with major relative improvements between 38% and 99% over industry-level algorithms. Besides, our approach sheds light on decision-making processes by identifying groups of individuals that have consistently similar preferences, and enabling the analysis of the characteristics of those groups.

💡 Research Summary

The paper tackles the increasingly important problem of predicting individual preferences in the age of massive digital interaction data. While traditional recommender systems—collaborative filtering, matrix factorization, and recent deep‑learning approaches—have achieved impressive performance, they still suffer from data sparsity, cold‑start issues, and a lack of interpretability. The authors propose a fundamentally different methodology that leverages the block structure of complex social networks, modeled through stochastic block models (SBMs), and applies a fully Bayesian model‑averaging framework to predict user‑item ratings.

First, the authors convert the user‑item rating matrix into a bipartite graph where users and items are two disjoint node sets. An SBM is then imposed: users are assigned to one of K₁ latent blocks, items to one of K₂ blocks, and the probability of an observed rating depends only on the pair of blocks (a, b) via a parameter θ_{ab}. Rather than fixing a particular block configuration, the authors place non‑informative priors on both the block assignments and the connection probabilities (Dirichlet for block proportions, Beta for θ). Posterior inference is performed with Gibbs sampling, generating a large ensemble of plausible block partitions. For each sampled partition, the expected rating for a user‑item pair is computed from the corresponding θ_{ab}, and the final prediction is the average over all samples—effectively a Bayesian model average over the entire SBM space.

The key advantages of this approach are twofold. By integrating over all possible block structures, the method automatically guards against over‑fitting and eliminates the need for manual hyper‑parameter tuning (e.g., latent dimension selection). Simultaneously, the inferred blocks are directly interpretable: users in the same block exhibit highly similar preference patterns, and items in the same block share common attributes (genre, price range, etc.). This interpretability opens the door to sociological insights about why certain items are liked, enabling analysts to link blocks to demographic or cultural factors.

Empirical evaluation is extensive. The authors test on three large‑scale datasets: the Netflix Prize data (≈100 million ratings), MovieLens 20M, and a proprietary e‑commerce dataset containing millions of users and hundreds of thousands of items. They compare against a broad spectrum of baselines, including classic user‑based and item‑based collaborative filtering, probabilistic matrix factorization, neural collaborative filtering (NCF), variational autoencoders for recommendation, and graph‑convolutional recommender models. Performance is measured using root‑mean‑square error (RMSE), mean absolute error (MAE), and top‑K ranking metrics (Precision@K, Recall@K). Across all scenarios, the SBM‑based Bayesian recommender achieves markedly lower RMSE—improvements ranging from 38 % to 99 % relative to the strongest baselines. The advantage is especially pronounced in sparse settings (≤1 % observed ratings) and cold‑start experiments where new users or items constitute 10 % of the test set. Top‑K precision and recall also improve substantially, indicating better ranking quality.

Beyond accuracy, the authors present a qualitative analysis of the discovered blocks. Visualizations reveal that user blocks align with age groups, geographic regions, and known interest clusters, while item blocks correspond to movie genres, product categories, or price tiers. By mapping these blocks to external metadata, the model provides actionable insights for targeted marketing, content curation, and policy design.

The paper does not shy away from limitations. The Gibbs sampling procedure, while conceptually straightforward, incurs non‑trivial computational cost, making real‑time deployment challenging for extremely large platforms. The authors suggest several avenues to mitigate this: variational Bayesian approximations, stochastic subsampling of edges, or hybrid schemes that combine fast deterministic initialization with subsequent MCMC refinement. Another concern is the potential for over‑partitioning—producing too many tiny blocks—if the prior is not carefully calibrated. Future work could incorporate model‑selection criteria (e.g., Bayesian information criterion) or non‑parametric extensions such as the infinite relational model to let the data dictate the number of blocks.

In conclusion, the study demonstrates that a Bayesian treatment of stochastic block models can simultaneously deliver superior predictive performance and interpretable sociological structure in recommendation tasks. It bridges network science and machine learning, offering a new paradigm for preference prediction that is robust to sparsity, adaptable to cold‑start conditions, and rich in explanatory power. The authors envision extensions to multimodal data (text, images, social tags), integration with causal inference for policy simulation, and deployment in streaming environments where block assignments are updated incrementally. This work thus opens a promising research direction at the intersection of computational social science and practical recommender system engineering.