Introduction
Traditional recommendation systems often struggle with the "one-size-fits-all" problem. On platforms like TripAdvisor, a business traveler seeking a quiet room and a family looking for a central pool receive the same generic rankings. This paper introduces a method to personalize these rankings by mining review texts to build "Topic Profiles" for users, effectively overcoming the chronic data sparsity of traditional numeric ratings. By applying intelligent data filters and a hybrid switching mechanism based on user "opinionatedness," the proposed system achieves a significant improvement in prediction accuracy over standard benchmarks.
Addressing Data Sparsity through Topic Profiling
Personalization is notoriously difficult because of data sparsity; in typical datasets, fewer than 10% of users have co-rated the same items. This makes traditional Collaborative Filtering (CF)—which relies on shared rating history—largely ineffective. While review text is a goldmine of preference data, it is often noisy.
The authors propose Topic Profile Collaborative Filtering (TPCF). Instead of matching users who rated the same hotel, TPCF matches users who care about the same topics (e.g., cleanliness, price, location). Users are mapped to a set of $k$ topics based on either Absolute Importance (frequency of mention) or Relative Importance (mentions compared to the average user).
Methodology and Applicability Filters
To ensure recommendations are based on high-quality data, the authors introduce three "Applicability Filters":
- Relevance: Only reviews sharing a minimum number of topics ($\alpha$) with the user's profile are considered.
- Density: Personalization is only triggered if there are at least $\gamma_0$ relevant reviews available for a product.
- Freshness: Recent reviews are prioritized to account for temporal changes in product quality.
The Role of Opinionatedness
A key innovation is the identification of "non-opinionated" users—often referred to as the "Kenny" persona—who leave only positive reviews. If a user never expresses a negative opinion, their preferences are indistinguishable from noise. The system utilizes a hybrid approach: if a user is "opinionated" (having left negative feedback on preferred topics), TPCF is used; otherwise, the system falls back to a non-personalized global average.
Experimental Results
The study utilized a large-scale crawl of 435,102 TripAdvisor reviews, focusing on "top contributors" to ensure robust profiles. The results demonstrate that personalization is a double-edged sword; applying it indiscriminately can actually degrade performance if data is sparse. However, when the density filter ($\gamma_0$) is sufficiently high, the gains are substantial.
As shown in the performance metrics, when at least 15 relevant reviews are available, TPCF outperforms the non-personalized benchmark by 33.6% in Mean Absolute Error (MAE) and 34.5% in Root Mean Square Error (RMSE). Furthermore, the system showed a significant improvement in ranking correlation (Kendall’s $\tau$), which the authors argue is a more practical metric for real-world utility than raw error scores.
Conclusion and Future Work
By using text as a "sparsity-killer," TPCF successfully identifies user neighbors even without shared item history. The hybrid approach ensures that the system only personalizes when it has enough high-quality data to trust the result. Future work will focus on improving the reliability of aspect-based sentiment analysis and addressing the "cold start" problem for users with fewer reviews.
Summary updated