TL;DR
This paper introduces a "Buzz of the Day" recommender system that identifies trending queries on an e-commerce platform to surface timely, novel, and serendipitous product recommendations. By focusing on short-lived query bursts, it aims to combat the staleness of traditional popularity-based recommenders and drive user engagement through curated daily highlights.
What Problem Does This Paper Address?
Traditional recommender systems often rely on long-term popularity or collaborative filtering, which can lead to recommendations that lack freshness and novelty. Products recommended based on sustained popularity might not reflect current user interests or emerging trends. This can result in a less engaging user experience, where users are shown the same popular items repeatedly. The authors address this by proposing a system that capitalizes on short-term "buzz" – sudden surges in user query interest – to provide dynamic and surprising recommendations. This approach aims to increase user curiosity, improve site activity, and offer a more engaging discovery experience.
Methods
The core of the system is the identification and selection of "buzz queries." These are queries that exhibit a significant burst in trend, detected using a bi-state automaton-based generative model (as described in [1]). The system then ranks these burst queries based on their occurrence and the amount of matching inventory available on eBay. The top-ranked query is chosen as the "Buzz of the Day."
Once a buzz query is selected, the system constructs a "semantic space" around it. This involves building a semantic query network using a combination of three similarity measures:
- Textual similarity (Ts): Based on the literal text of queries.
- Session similarity (Ss): Queries that appear together in user sessions.
- Semantic similarity (Ks): Derived from historical transaction data using a kernel function.
These similarity measures are linearly combined with learned coefficients (α, β, γ) to create a composite similarity score (Cs). This network allows the system to identify related products, competing products, similar personalities, and current trends associated with the buzz query. Recommendations are then made from products that fall within this semantic neighborhood.
The system also incorporates user feedback through simple "thumbs up" and "thumbs down" emoticons, and a "Surprise Me" button to access past buzz recommendations.
Datasets
The primary dataset used is the stream of queries flowing into the eBay US site. For evaluation, the authors created a website (http://www.buzzaday.com) and invited 100 users to provide feedback on recommendations. A total of 113 votes were collected for approximately 68 unique recommendations.
Key Insights and Ideas
- Leveraging Query Bursts for Novelty: The fundamental idea is that sudden spikes in user queries (buzz) are strong indicators of current, albeit temporary, user interest that can be exploited for recommendations.
- Combating Recommendation Staleness: By focusing on transient trends, the system offers an alternative to static, popularity-driven recommendations, enhancing the novelty and serendipity of product discovery.
- Semantic Neighborhoods for Richer Context: Building a semantic network around a buzz query allows for more nuanced recommendations, going beyond direct matches to related concepts and entities.
- Category Sensitivity to Buzz: The classification of buzz recommendations into eBay's taxonomy reveals that certain categories (e.g., "Sports Memorabilia," "Cell Phones") are more susceptible to buzz than others, which has implications for targeted marketing.
- User Feedback as a Signal: While simple, the thumbs up/down feedback provides a direct signal of user reception, allowing for analysis of what types of buzz recommendations resonate and which do not.
Key Results
- Positive User Reception: Out of 113 votes, 83 were "thumbs up" and 30 were "thumbs down," indicating a generally positive reception to the buzz-based recommendations (approximately 73% positive).
- Category Performance: "Tickets" and "Sports Memorabilia, Cards & Fan Shop" categories received a significant portion of positively voted recommendations.
- Classification Insights: "Sports Memorabilia, Cards & Fan Shop," "Cell Phones & PDAs," "Clothing, Shoes & Accessories," and "Tickets" accounted for about 43% of buzz recommendations, suggesting higher sensitivity to external trends.
- Negative Vote Analysis: Negative votes were sometimes attributed to recommendations that didn't align with user expectations for seasonal events (e.g., Valentine's Day, April Fool's Day) or lacked sufficient context.
Limitations and Future Work
- Personalization: The current system does not personalize buzz recommendations based on individual user history. Future work aims to incorporate user history to tailor recommendations.
- Community Feedback: The authors plan to integrate community reaction and feedback more deeply to refine the recommendation process.
- Contextual Understanding: The analysis of negative votes suggests that a deeper understanding of user context and expectations around specific buzz events is needed.
- External Buzz Triggers: While the system uses query bursts, explicitly linking them to external news and inventory information is mentioned as a key part of identifying buzz queries. The paper could benefit from a more detailed discussion on how external signals are integrated beyond just query trends.
Summary updated