Personalized Sentiment Classification Based on Latent Individuality of Microblog Users

Kaisong Song, Shi Feng, Wei Gao, Daling Wang, Ge Yu, Kam-Fai WongIJCAI 2015

Standard sentiment analysis often fails because it ignores the "who" behind the "what." For instance, a fitness enthusiast and a sick patient both saying "I lost five pounds" express fundamentally different sentiments. Traditional "one-size-fits-all" algorithms treat language as universal, assuming a word's polarity is fixed regardless of the speaker's background, occupation, or personality. This research introduces a personalized classification model that leverages Latent Factor Models (LFM) and social network data to capture these individual nuances, significantly outperforming traditional methods.

The Challenge of Personalization

Personalizing sentiment analysis for microblogging platforms like Twitter or Sina Weibo presents two major hurdles:

  1. Individuality: The same phrase can have opposite polarities depending on the user’s background. A "finance" professional and an "education" professional might use the same vocabulary to describe very different emotional states.
  2. Data Sparsity: Most users do not post frequently enough for a machine learning model to build a robust, individual classifier. If a user has only posted a handful of times, there isn't enough "signal" to understand their unique linguistic style.

Methodology: Extending the Latent Factor Model

To address these issues, the researchers proposed an extensible version of the Latent Factor Model (LFM), a technique commonly used in recommendation systems to map users and items into a shared latent space.

The Extended LFM decomposes the user-microblog matrix into latent factors representing interested aspects and individual traits.

As shown in Figure 1, the core innovation lies in the Word-Level Decomposition. Instead of treating a whole post as a single item in a matrix—which results in extreme sparsity—the model decomposes posts into individual words. This maps users and words into a shared "latent factor space," where the dimensions represent personal traits or specific areas of interest.

Furthermore, the model moves beyond a simple "bag of words" approach by using dependency parsing. This identifies "Sentiment Units" (e.g., "not good") and "Topic Units" (e.g., "work overtime"), allowing the model to capture the syntactic relationships between words rather than just their frequency.

Solving the Sparsity Problem

The transition from a User-Post matrix to a User-Word matrix was a critical breakthrough for scalability. In a standard User-Post matrix, the data is incredibly sparse because any single user only interacts with a tiny fraction of all possible posts.

Comparison of non-zero item rates between word-level and post-level data across Weibo and Twitter datasets.

By shifting to a word-level analysis, the researchers increased the "non-zero" data density from roughly 0.3% to nearly 3% (as illustrated in Figure 6). This increased density makes matrix factorization significantly more effective, providing the model with enough data points to actually learn a user's latent profile.

To further combat sparsity, the model incorporates Social Regularization. Based on the principle of homophily—the idea that people tend to follow others with similar interests—the model uses "following" relations to "borrow" information from a user's social circle. If a user's own data is sparse, the model nudges their latent profile toward the profiles of their social peers.

Key Results and Insights

The model was evaluated on two real-world datasets: Sina Weibo (43,250 posts) and Twitter (48,563 posts). Using the G-mean metric to account for class imbalance, the results were striking:

  • Superior Accuracy: The full model outperformed standard SVMs by massive margins. On Weibo, the G-mean jumped from 0.587 (SVM) to 0.735. On Twitter, it rose from 0.619 to 0.840.
  • Latent vs. Explicit Personalization: The LFM approach even outperformed "domain-based" models that group users by explicit categories like occupation. This suggests that "latent individuality" captured through behavior is more nuanced than simple demographic labels.
  • The Power of Context: Ablation studies confirmed that both dependency parsing and social relations are essential components, proving that linguistic structure and social context are complementary in understanding sentiment.

Conclusion

By mapping users to a latent space and leveraging social connections, this research demonstrates that knowing who wrote a post is often as important as the words themselves. While the model's performance is currently tied to the richness of the social graph and the accuracy of linguistic parsers, it represents a significant step toward sentiment analysis that truly understands the individual.

Summary updated

BibTeX
@inproceedings{song2015personalized,
  author = {Kaisong Song and Shi Feng and Wei Gao and Daling Wang and Ge Yu and Kam-Fai Wong},
  title = {{Personalized Sentiment Classification Based on Latent Individuality of Microblog Users}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2015},
}
Figure