Evaluating the Ebb and Flow: An In-depth Analysis of Question-Answering Trends across Diverse Platforms

Tag related analysis: For each platform, we obtain the most frequently occurring tags (top 50) across the questions posted. For $SE$, Java, Design, C# are the most popular tags. For $GD$, Unity, C++, C# and OpenGL are most frequent and for $MO$, United States and taxes are most popular tags. In case of $MA$, $EN$ and $CH$, we do not observe any particularly dominant set of tags. Further, for each portal, we calculate the average number of tags present in $Q_{top}$ and $Q_{bottom}$ questions. In $MA$, $EN$ and $SE$, the average number of tags in $Q_{top}$ and $Q_{bottom}$ are 2.08, 2.03, 2.53, 2.66, 2.17 and 3.02 respectively. In $GD$, $CH$ and $MO$, the average number of tags for $Q_{top}$ and $Q_{bottom}$ are 2.58, 2.28, 3.00, 2.94, 2.54 and 3.23 respectively. We observe that there are lesser number of tags in the questions in $Q_{top}$ than in $Q_{bottom}$. This could possibly indicate lesser tags makes a question topically less confusing and streamlines it more effectively to people who can actually answer them.

Votes received by the question: In order to understand the temporal pattern of votes received by the questions, we count the fraction of questions receiving 1 upvote, 2 upvotes and $\geq 3$ upvotes on three consecutive days, i.e., the same day when the question is posted followed by the two subsequent days. In the same day, $\sim$30% questions of $SE$ platform receive more than 3 votes. But for other platforms most of the questions received one vote. In second day, around $\sim$20% of questions receive one more vote in $SE$, $MO$ and $CH$ platforms. In case of $SE$ and $MO$, there are 10% of questions receiving at least three votes in the second day. In the third day, there are very less number of questions receiving votes across all the platform. This indicates that on all platforms the attention span of users for a particular question dies down within 48 hours from the post of the question. In case of $MA$, $EN$ and $SE$, the average votes for the questions in $Q_{top}$ are 1.99, 5.31 and 15.82 respectively and that for the questions in $Q_{bottom}$ are 2.75, 2.60 and 4.44 respectively. In case of $MO$, $GD$ and $CH$, the average number of votes in $Q_{top}$ are 6.46, 4.63 and 4.21 respectively while that in $Q_{bottom}$ are 2.96, 2.08 and 5.17 respectively. Hence questions in lower response time typically receive more upvotes irrespective of the popularity of the platform.

Reputation score of users: We separately observe the reputation scores of asker and answerer of questions of $Q_{top}$ and $Q_{bottom}$ category. Asker is the user who asked the question in the portal and answerer is the user who answered the question. In $MA$, $EN$ and $SE$, the average reputation scores of askers in $Q_{top}$ are 559.07, 546.27 and 1001.96 respectively. The average reputation scores of askers in $Q_{bottom}$ 705.52, 522.78 and 644.14 respectively. In $MO$, $GD$ and $CH$, the average reputation score of askers in $Q_{top}$ are 585.71, 401.53 and 424.68 respectively. The average reputation score of askers in $Q_{bottom}$ are 540.71, 307.14 and 561.10 respectively. In case of answerers, the average reputations in $Q_{top}$ are 4063.39, 2147.42 and 3209.59 for $MA$, $EN$ and $SE$ respectively and those in $Q_{bottom}$ are 2001.24, 1622.79 and 1846.81 respectively. For $MO$, $GD$ and $CH$, the average reputation of answerer in $Q_{top}$ are 4017.05, 1117.52 and 1948.33 respectively while that in $Q_{bottom}$ are 1965.42, 552.84 and 1397.38 respectively. We observe three things here – (i) across all platforms the reputation scores of askers as well as answerers are typically more for $Q_{top}$ questions (ii) in the popular platforms the reputation scores for both askers and answerers are higher and (iii) answerer reputations are an order of magnitude higher than the askers.

Asker-Answerer graph (AAG): To understand the user interaction patterns on a platform, we build an undirected graph where the nodes are the users and the edges denote if a user (answerer) answers a question posted by another user (asker). For every platform, we build two such AAGs for users who either asked or answered the questions of $Q_{top}$ and $Q_{bottom}$ respectively (See Table 4). Further, we calculate average degree, average clustering coefficient for the whole network. In Figure 1 (e), we observe that the average degree of $Q_{top}$ AAG for all platforms except $CH$ are relatively higher than the average degree of $Q_{bottom}$ AAG. For all the platforms, the average clustering coefficient of $Q_{top}$ AAG are higher than the average clustering coefficient of $Q_{bottom}$ AAG (see Figure 1 (d)). This indicates that there is a higher density of interactions among the asker-answer(s) in the $Q_{top}$ AAG (and exceptionally higher platforms $SE$, $GD$ and $MO$).

Cross platform users: We found around 2739 users who are present in all the six platforms. Fraction of common users posted at least one question is around 29%-30% for $MA$ and $EN$ and for $SE$ it is 18%. However for $CH$, $GD$ and $MO$ platforms, this fraction comparatively lower ($\leq 15\%$). The observations are similar for the number of common users posting an answer – the popular platforms have far larger fractions compared to less popular ones. Next we compute the number of common users who participated in asking/answering a question in $Q_{top}$ and $Q_{bottom}$. The percentage of common users asking one or more questions in case of $MA$, $EN$ and $SE$ are 19.82%, 17.30% and 9.71% respectively in $Q_{top}$ and 20.0%, 16.24% and 9.49% respectively in $Q_{bottom}$. The percentage of common users in these three platforms who answered one or more questions in $Q_{top}$ are 9.3%, 8.72% and 6.17% respectively. In $Q_{bottom}$ these percentages are 12.92%, 7.7% and 6.4% respectively. For $MO$, $GD$ and $CH$, the percentage of common users acting as asker in $Q_{t}op$ are 7.84%, 6.38% and 5.69% respectively and in $Q_{bottom}$ are 8.06%, 6.09% and 5.69% respectively. For these three platforms, the percentage of common users acting as answerer in $Q_{t}op$ are 2.7%, 4.3% and 1.8% respectively and in $Q_{bottom}$ are 4.2%, 4.9% and 2.1% respectively.

Engagement of common users: In order to investigate the involvement of common users (present in $Q_{top}$ and $Q_{bottom}$ as answerers) in the overall platform, we compute the average number of questions/answers posted by them at an aggregate level. For $MA$, $EN$ and $SE$, the average number of questions posted by common users in $Q_{top}$ are 22.68, 8.5 and 3.04 and that in $Q_{bottom}$ are 21.60, 8.74 and 3.57 respectively. For these three platforms, the average number of answers posted by common users in $Q_{top}$ are 228.83, 47.32 and 61.01 and that in $Q_{bottom}$ are 166.7, 52.20 and 56.74 respectively. For $MO$, $GD$ and $CH$, the average number of questions posted by common users in $Q_{top}$ are 11.56, 5.1 and 5.61 respectively. and in $Q_{bottom}$ are 9.02, 5.45 and 7.55 respectively. For these three platforms, the average number of answers posted by common users in $Q_{top}$ are 49.81, 24.74 and 15.55 respectively. The average number of answers posted by common users in $Q_{bottom}$ are 35.11, 21.43 and 14.22 respectively. Thus (i) irrespective of the popularity of the platforms, common users engage more in terms of asking and answering in $Q_{top}$ than in $Q_{bottom}$ and (ii) quite interestingly, the average number of answers posted by the common users in $Q_{top}$ and $Q_{bottom}$ are far more than the questions posted by them which indicates that these users are more engaged in posting answers than questions.

Parameter settings: The parameter settings for each model are given below. For SVC, we consider the value of parameter $C$ in the range of 1-3. The values tried for kernel are poly, rbf and sigmoid. For random forest classifier, the range of the number of estimators is between 160 to 200. The criterion values are gini and entropy. For XGBoost, the range of values for the number of estimators is from 160 to 200. The range of learning rate is 0.1 to 0.3.

Logistic regression (LR): For all datasets, we found the default parameters are the best.
Support vector classifier (SVC): For $CH$, the best value of $C$ is 3, and kernel is set to poly. For other datasets, the best parameter for kernel is rbf. For $EN$ and $GD$, the value of the parameter C is set to 2 and 1 respectively. For the other datasets, the value of C is 3.
Random forest (RF): For $CH$, $EN$, $MO$, $MA$, the best value for criterion is entropy. For $GD$ and $SE$, the criterion is set to gini. For $MO$, the number of estimators is set to 180. For other datasets, the number of estimators is set to 200.
XGBoost: For all the datasets, the best value for the learning rate is 0.1. For $CH$, $GD$, $MO$, the value of the number of estimators is set to 160. For $EN$ and $SE$, the number of estimators is 180. For $MA$, the number of estimators is 200.