RRM: Decoding the "Internal Influence" of Reviews for Precision Social Recommendation

TL;DR

While most social recommenders look at who you follow, the Review-based Recommendation Model (RRM) looks at how your friends write. By quantifying "Sentimental Deviation" (how polarized a user's opinions are) and "Review Reliability" (the depth of their feedback), and fusing these with an attention mechanism, this model improves rating prediction accuracy by up to 6.7% over existing SOTA methods.

The Missing Link: Why Traditional Social Recommendation Fails

Most social recommender systems operate on a simple "trust" premise: if User A follows User B, they share interests. However, this relies on explicit social graphs or tags that are often sparse or non-existent.

The authors identify a critical gap: Internal Factors. They argue that a reviewer’s influence isn't just about popularity—it's about their psychological impact on the reader. For instance:

• The Polarized Critic: Someone who gives varied, strong opinions (high sentimental deviation) is often more influential than someone who gives "3 stars" to everything.
• The Detailed Guide: A user who writes long, feature-rich reviews is more "reliable" than a user who writes one-word summaries.

Methodology: From Sentiment to Social Latent Space

The RRM architecture processes raw text into a mathematical influence score through three main stages:

1. Sentiment Analysis via Word2Vec + SVM/SVR

Instead of simple lexicon-based matching, the authors use word2vec to capture the semantic nuance of reviews. They transform reviews into feature vectors and use Support Vector Regression (SVR) to predict a continuous sentiment score ($E_{u,i}$).

2. Quantifying Internal Influence

The model introduces two novel metrics:

• Sentimental Deviation ($D(E_u)$): Calculated as the variance of a user's sentiment scores across items. High variance indicates a user who is discerning and polarized, making their reviews more "accessible" and impactful.
• Reliability ($r_u$): Modeled as an exponential distribution where review length is correlated with the number of product features mentioned.

Fig 1: The RRM framework integrating sentiment extraction and matrix factorization.

3. The Attention Mechanism

Not all influence factors are equal. The paper uses an Attention Mechanism to automatically learn the weights ($\alpha ,\beta ,\gamma$) for Sentimental Deviation, Reliability, and Popularity. This allows the model to adapt to different datasets (e.g., valuing detail more in "Home Services" than in "Nightlife").

Experimental Results: SOTA Performance

Testing on the Yelp dataset across 8 categories (from Restaurants to Pets), RRM consistently outperformed baselines like CircleCon and PRM.

Key Insight: The Power of Polarization

One of the most fascinating findings (shown in the figure below) is that $\alpha$ (Sentimental Deviation) usually received a high positive weight, while $\gamma$ (Popularity) was often lower or even negative in its relative gradient. This suggests that a user's distinct "voice" is more predictive of their influence than the sheer number of friends they have.

Fig 2: The relevance between weight learning and iterations, showing the dominance of sentimental factors.

Critical Analysis & Takeaways

The RRM model successfully shifts the focus from "tangible" social actions to "untouchable" internal influence.

Strengths:

• Implicit Signal Mining: Extracts value from unstructured text reviews that are usually discarded by matrix factorization models.
• Scalability: The attention mechanism allows the model to "ignore" non-beneficial factors automatically.

Limitations:

• Text Length Dependency: The authors admit their Word2Vec-based sentiment analysis is optimized for short texts (like Yelp) and underperforms on long-form movie reviews compared to NBSVM.
• Computational Overhead: Performing SVR and Word2Vec on millions of reviews adds significant preprocessing latency compared to simple CF.

Future Outlook: This research paves the way for "Psychological Recommendations," where the system understands not just what you like, but whose style of critique moves you to make a purchase.