Improving Recommendation Quality by Merging Collaborative Filtering and Social Relationships

Matrix Factorization techniques have been successfully applied to raise the quality of suggestions generated by Collaborative Filtering Systems (CFSs). Traditional CFSs based on Matrix Factorization operate on the ratings provided by users and have been recently extended to incorporate demographic aspects such as age and gender. In this paper we propose to merge CFS based on Matrix Factorization and information regarding social friendships in order to provide users with more accurate suggestions and rankings on items of their interest. The proposed approach has been evaluated on a real-life online social network; the experimental results show an improvement against existing CFSs. A detailed comparison with related literature is also present.

💡 Research Summary

The paper addresses a well‑known limitation of collaborative filtering (CF) systems that rely on matrix factorization (MF): the sparsity of the user‑item rating matrix often leads to poor similarity estimates, especially when two users have few or no co‑rated items. While recent works have incorporated demographic attributes or explicit trust relationships to mitigate this problem, trust is typically asymmetric and requires users to explicitly declare whom they trust, which limits its applicability.

To overcome these drawbacks, the authors propose a novel framework that integrates social friendship information directly into the MF model. They define a social rating network (SRN) as a quadruple ⟨U, I, φ, ψ⟩ where φ maps user‑item pairs to observed ratings (or a missing symbol) and ψ maps user‑user pairs to a binary indicator of a social tie (friendship). The tie can be symmetric or asymmetric, and the set of neighbors Nₓ for a user uₓ consists of all users v such that ψ(uₓ, v)=1.

The core idea is to enforce that friends be close to each other in the latent space. Starting from the classic MF objective

L = ½‖R – P Qᵀ‖²_F + λ (‖P‖²_F + ‖Q‖²_F)

they add a social regularization term weighted by a hyper‑parameter μ:

L′ = L + μ ∑ₓ ∑_{y∈Nₓ} ‖pₓ – p_y‖ .

Thus, the optimization simultaneously minimizes rating reconstruction error, controls model complexity (λ), and penalizes large distances between the latent vectors of socially linked users (μ).

Deriving the gradients yields:

∂L′/∂pₓ = Σ_j δ_{xj}(pₓ·q_j – R_{xj}) q_j + λ pₓ + μ Σ_{y∈Nₓ}(pₓ – p_y)

∂L′/∂q_j = Σ_x δ_{xj}(pₓ·q_j – R_{xj}) pₓ + λ q_j

where δ_{xj}=1 if rating R_{xj} is observed, otherwise 0. The authors solve these equations using stochastic gradient descent (or batch gradient descent) with a learning rate β, iterating until convergence or a preset error threshold ε.

For empirical validation, they built a small social platform called “Cofe” where 37 university students could rate movies and form friendship links. The dataset contains 297 ratings. They split the data into training and test sets, evaluate using Root Mean Square Error (RMSE), and compare three methods: (1) a naive baseline that predicts the global average rating, (2) standard Non‑negative Matrix Factorization (NMF) without social information, and (3) the proposed socially‑aware MF. Hyper‑parameters λ and μ are tuned via cross‑validation. Results show that the proposed method consistently achieves lower RMSE than both baselines, with the most pronounced gains for users who have many friends, confirming that social proximity improves prediction accuracy.

The paper’s contributions are threefold: (i) it demonstrates that explicit social ties, which are easier to obtain than trust statements, can be seamlessly incorporated into MF; (ii) it provides a simple yet effective regularization term that encourages socially connected users to occupy neighboring points in the latent space; (iii) it retains the computational simplicity of standard MF, requiring only modest modifications to the gradient updates.

However, several limitations are acknowledged. The binary representation of friendships ignores varying tie strengths (e.g., interaction frequency, closeness), potentially discarding useful signal. The experimental dataset is very small and may not reflect the challenges of large‑scale recommender systems, raising questions about scalability and generalization. Moreover, the paper does not compare against more recent graph‑based or social‑MF models (e.g., SocialMF, TrustSVD, Graph Convolutional Networks), which limits the assessment of relative performance. The tuning procedure for μ and λ is described only at a high level, leaving reproducibility concerns.

In conclusion, the study proposes a conceptually clear and technically sound method for merging collaborative filtering with social relationships, showing measurable improvements on a real‑world dataset. Future work should explore weighted social ties, test the approach on larger, more diverse datasets, and benchmark against state‑of‑the‑art social recommendation algorithms to fully establish its practical value.