TL;DR
This paper introduces a novel recommendation system that leverages social sentiment from external platforms like Facebook, moving beyond traditional e-commerce site reviews. It models user opinions and product features in a polynomial space, using asymmetric similarity and advantage metrics to provide more nuanced recommendations, even for job applications.
What Problem Does This Paper Address?
Traditional recommendation systems often rely on structured data within e-commerce platforms (like Amazon's review system). However, this approach misses out on the vast, unstructured, and opinion-rich data available on social networking services (SNS) like Facebook. These external platforms offer a more dynamic and diverse source of user sentiment. The challenge lies in effectively extracting and utilizing this "social sentiment" from semi-structured SNS data, which is often less organized than dedicated review systems, to generate relevant product and even job recommendations.
Methods
The core of the proposed system lies in its unique approach to modeling social sentiment and product features.
- Polynomial Vector Space: User opinions and product features are represented as polynomials in a finite-dimensional linear space. This allows for a flexible representation of complex relationships.
- Asymmetric Similarity Measurement: Unlike traditional symmetric measures (where A is as similar to B as B is to A), this system uses an asymmetric measure. This is crucial because user preferences are often directional (e.g., someone looking for an HDTV cable is likely to already own an HDTV, but not vice-versa). The formula
<<p,q)) = exp((p-q,p-q))is introduced, wherepandqare polynomial representations of user opinions on different walls. - Asymmetric Advantage Measurement: Beyond similarity, the system incorporates an "advantage" measurement (
<|p,q|>)) to capture directional user preference and brand loyalty. This helps to avoid recommending products that might be technically similar but contextually inappropriate (e.g., recommending Dell products to someone unhappy with their Dell laptop). - Computational Efficiency: The paper highlights mathematical properties of their similarity and advantage measures that significantly reduce computational overhead, making the system scalable for large datasets. This includes leveraging orthogonal bases and properties of polynomial operations.
- Relevance and Advantage Combination: The final recommendation score is a weighted combination of a "relevance" score (derived from the similarity measure) and an "advantage" score.
Datasets
- FB-8000: A large dataset comprising 8094 Facebook public walls, synchronized in real-time. It includes over 196 million unique users, 360 million comments, and 1.83 billion likes. This dataset is used for training and evaluating the core recommendation system.
- AMZN-900: A dataset from Amazon.com, containing 914 products, over 330,000 unique reviewers, and over 510,000 reviews/ratings. This dataset represents a traditional, structured review system and is used for comparison.
- LNKD-2000: A dataset from LinkedIn, containing 571 company profiles, over 1.5 million user profiles, 3.1 million comments, and 1.1 million likes. This dataset is used to demonstrate the system's applicability to job recommendation.
Key Insights and Ideas
- Leveraging External SNS is Key: The paper strongly argues that external SNS data is a richer and more valuable source for recommendations than solely relying on internal e-commerce review systems.
- Asymmetry Matters in Social Contexts: The concept of asymmetric similarity and advantage is crucial for capturing real-world user behavior and preferences, where relationships are not always reciprocal.
- Polynomials as a Powerful Representation: Using polynomials to model user sentiment and product features provides a mathematically convenient and expressive way to capture complex opinions and their nuances.
- Bridging Product and Job Recommendations: The system's adaptability to job recommendation tasks demonstrates its broad applicability beyond just product suggestions, highlighting the universality of social sentiment analysis.
- Computational Efficiency is Achievable: The paper provides a clear path to making complex social sentiment analysis computationally feasible for large-scale applications through clever mathematical derivations.
Key Results
- The proposed "AdvR" method (combining relevance and advantage) performs comparably to or better than "FeatureR" (a method using structured review data) on the AMZN-900 dataset, demonstrating the effectiveness of using external SNS data.
- Figures 4 and 5 show that the AdvR method achieves competitive MRR-IA and NDCG-IA scores, often increasing with larger k-values (number of recommendations considered), indicating robust performance.
- Case studies, like the Amazon Kindle recommendations, show that the system surfaces relevant products and avoids less successful ones, while also identifying competitive products like NOOK through "breadth-first" expansion.
- Experiments on the LNKD-2000 dataset for job recommendations show promising results, with the system identifying relevant companies and fields for specific focal companies like IBM, ESPN, and Groupon.
Limitations and Future Work
- Privacy Concerns: The authors acknowledge that providing personalized results would require non-trivial privacy-preserving experiment design.
- High-Level Meanings: Future work aims to incorporate more sophisticated methods for understanding high-level semantic meanings beyond simple text sentiment, potentially leading to even more insightful recommendations.
- Spam User Filtering: While some spam filtering was applied, the paper notes that more robust methods could further improve data quality.
- Generalization: While the system shows promise in generalizing to job recommendations, further exploration of its applicability to other domains is suggested.
Summary updated