TL;DR
This paper analyzes how users' interests change over time on Stack Overflow, revealing that general public interest shifts follow a power-law distribution, unlike scientists' exponential patterns. This research offers insights into user behavior that can inform community management and recommendation systems.
What Problem Does This Paper Address?
While the evolution of scientific interests has been extensively studied, there's a notable lack of research on how interests change within general public online communities, such as question-and-answer platforms. Understanding these patterns is crucial for comprehending user engagement, designing effective recommendation algorithms, and fostering the healthy growth of these communities. This paper aims to fill this gap by investigating interest change patterns on Stack Overflow, a massive programming Q&A community.
Methods
The researchers developed a method to quantify user interest changes by analyzing their posting history.
- Tag Network Analysis: They first constructed a tag network where nodes represent programming tags and edges connect tags that co-occur in questions. This network was then partitioned into communities using the Infomap algorithm to represent distinct technical domains.
- Topic Vectorization: For each user, they created a "topic vector" representing their interest profile. This vector was derived from the sequence of questions asked, where each question's topic was abstracted by its tags. The topic vector was then normalized based on the frequency of tags within a user's recent questions.
- Interest Change Quantification: To measure interest change, they calculated the "interest change" ($J$) between the user's initial set of $m$ questions and their final set of $m$ questions. This was done by computing the complementary cosine similarity between their initial and final topic vectors. A $J$ value of 0 indicates no change, while $J=1$ signifies a complete shift in interest.
- Feature Analysis: They investigated the influence of three key features on interest change:
- Heterogeneity: How diverse a user's interests are.
- Recency: The tendency to engage with topics similar to recent ones.
- Proximity: The tendency to explore new domains that are similar to current interests. They employed experimental manipulations (e.g., removing heterogeneity, recency, or proximity from user sequences) to assess their impact on the interest change distribution.
Datasets
The study utilized a publicly available dataset from Stack Exchange, specifically focusing on the Stack Overflow Q&A community. The dataset spanned from July 2008 to September 2016 and included information on posts, questions, answers, tags, posting dates, and user reputations. For their analysis, they focused on "active users" who had asked more than 50 questions (totaling 31,303 users) and tags that occurred at least 50 times (totaling 19,978 tags).
Key Insights and Ideas
- Power-Law vs. Exponential: The most striking finding is that general users' interest changes on Stack Overflow follow a power-law distribution, meaning most users exhibit small interest shifts, but a rare few undergo dramatic changes. This contrasts sharply with scientists' interest changes, which tend to follow an exponential distribution, suggesting scientists are more likely to gradually shift their focus.
- Heterogeneity Matters: A user's diverse range of interests (heterogeneity) plays a role in limiting drastic interest changes. When this heterogeneity is removed, the distribution of interest changes shifts, indicating that a broader interest base can lead to more conservative exploration.
- Recency Drives Exploration: The tendency to engage with recent topics (recency) actually encourages users to explore new domains, leading to a broader variety of interest changes. Removing recency makes users less likely to shift their interests significantly.
- Proximity's Dual Role: Proximity, the tendency to explore topics similar to current ones, has a complex effect. It can limit both very small and very large interest changes, suggesting it can both anchor users to familiar areas and guide them towards related new ones.
- Concentration for Reputation: A significant negative correlation exists between user reputation and interest change. This implies that users who focus on specific areas (i.e., have lower interest change) are more likely to gain higher reputations, highlighting the value of concentration for success on the platform.
Key Results
- The distribution of user interest change ($J$) on Stack Overflow follows a power-law distribution ($P \sim J^{-0.608}$), indicating that most users have minor interest shifts, but a small proportion experiences significant changes.
- When heterogeneity is removed from user interest sequences, the distribution of $J$ changes significantly, showing a sharply rising trend followed by a slow fall, different from the original power-law decrease.
- Excluding recency transforms the interest change distribution from a power-law to an exponential distribution, with a more significant decrease in users exhibiting extremely large interest changes compared to scientists.
- Excluding proximity reduces the proportion of users with both small ($J < 0.3$) and large ($J > 0.7$) interest changes, fitting a normal distribution.
- Average short-term interest change ($J_s$) increases linearly with the time a user spends on the community.
- User reputation is negatively correlated with interest change ($J$). Active users with low $J$ have higher reputations than inactive users with low $J$.
Limitations and Future Work
- Timescale Ignored: The current study focuses on the sequence of interests but does not explicitly consider the timescale of interest changes, particularly rapid "explosive" interest emergence. Future work could incorporate timescale analysis.
- Tag Network Division: The method for dividing the tag network into communities might be influenced by hypernym-hyponym relationships. Future research could refine this division by considering such linguistic structures for more accurate topic representation.
- Generalizability: While Stack Overflow is a large Q&A community, further research could explore interest patterns in other types of online communities and their relationship with different contribution types.
- Community Management: The authors suggest that community managers could leverage these findings to design better recommendation systems by focusing more on users' current interests rather than solely on their historical data.
Summary updated