RSSP: Bridging Social Graphs and Emotional Consistency for Robust Sentiment Analysis

TL;DR

Social media sentiment analysis is notoriously difficult due to the "noise" of short-form text and the scarcity of labeled data. The Robust Semi-supervised Sentiment Propagation (RSSP) framework tackles this by using social network relations (who follows whom) and a novel "emotional consistency" check. It prevents the common pitfall of "negative transfer" by intelligently weighting predictions based on how well different classifiers agree on specific regions of the data.

Problem & Motivation: Beyond the Text

Traditional sentiment analysis looks at words. But on Twitter, users don't just post in vacuums; they exist in a Social Network. The principle of homophily suggests that connected users often share similar opinions.

Current limitations in the field include:

1. Data Scarcity: Sentiment labels are expensive to get.
2. Domain Sensitivity: Lexicons (lists of "good" vs "bad" words) fail when slang or context changes.
3. The Noise Trap: Distant supervision (using emoticons as labels) provides quantity but lacks quality, often confusing the model (Negative Transfer).

The author's insight: If we can measure the consistency between a graph-based prediction (social) and a text-based prediction (lexicon), we can identify which data points are reliable even without manual labels.

Methodology: The Three-Phase RSSP Framework

The core of the paper is a beautifully structured pipeline that moves from raw graph data to a precise, weighted classifier.

Phase 1: Hybrid Label Propagation

The authors don't just look at text similarity. They build a complex similarity matrix $W$ that adds three components:

• Textual Similarity: Gaussian kernel on message features.
• User Consistency: Are these two tweets from the same person?
• Social Connection: Are the authors of these tweets friends/followers?

Labels are then propagated from a small set of ground truth to the entire graph by minimizing a loss function that favors "smoothness" (neighbors should have same labels).

Phase 2: Emotional Clustering Consistency

This is the "secret sauce." The model trains two preliminary classifiers: one on "Gold" (labeled) data and one on "Noisy" (distant supervision) data. By comparing their predictions on different "regions" (clusters) of the unlabeled data, the system calculates a Relevance Score.

Figure 1: Visualizing how consistency determines relevance. If the social-graph labels and the noisy-text labels agree on a region, that region's data is highly relevant for training.

Phase 3: Final Target Learning

The final classifier is learned in a Reproducing Kernel Hilbert Space (RKHS). It aims to satisfy three goals simultaneously:

1. Minimize error on the small labeled set.
2. Control model complexity (Regularization).
3. Stay close to the "Reference Predictions" (the weighted consensus from Phase 2).

Experiments & Results: Dominating the Baselines

The framework was tested on the STS (Stanford Twitter Sentiment) and OMD (Obama-McCain Debate) datasets.

SOTA Comparison

RSSP proved superior to classic SVMs, Logistic Regression, and even other graph-based methods like SANT.

• Average Improvement: RSSP outperformed SANT by 5.5% in accuracy.
• Weak Supervision Performance: Even with only 10% of training data, RSSP maintained high accuracy, while standard models like SVM struggled significantly.

Table 1: Accuracy results on STS dataset across different training sizes. RSSP consistently maintains a lead.

Critical Analysis & Conclusion

Takeaway

The paper successfully argues that social context is not just "extra" data—it is a grounding mechanism. By measuring consistency between different signal sources (social vs. lexicon), we can filter out the noise inherent in Twitter data.

Limitations

• Graph Dependency: The method relies on having access to a follower/friend graph, which is becoming increasingly restricted by platform APIs (e.g., changes in Twitter/X API access).
• Computation: Large-scale label propagation and kernel-based learning can be computationally expensive as the number of tweets ($n$) grows.

Future Work

The relevance-weighting mechanism in RSSP is essentially a "filtering" layer that could be adapted for Multi-modal Sentiment Analysis (combining text, images, and social graphs) or even real-time trend monitoring where "noisy" signals are the only available data.