Pay-per-Question: Towards Targeted Q&A with Payments Steve Jan , Chun Wang, Qing Zhang , Gang Wang
Online Question & Answer Services • Web search engines – But they can not give users the customized answers • Online Q&A Service – Quora: 1M questions / month – Stack Overflow: 200K questions / month Are they good enough? 2
Scenario 1 • I got a traffic ticket the other day Q: What are the tips to fight this ticket on court? Ask friends Post it online Ask lawyers They may not Should I trust Too expensive understand the law them? Too slow 3
Scenario 2 • I feel mildly sick the other day Q: What were causing my headache and nausea? Ask friends Post it online Ask doctors Too expensive They may not Should I trust Too slow understand medic them? 4
Another Option • I can directly ask some experts: – Convenient: use my smart-phone – Trustworthy: certified domain experts – Targeted: ask experts who I can trust – Cheaper: cheaper than making a real appointment Targeted Q&A apps are for the rescue! 5
Targeted Q&A Apps • Social network: connect users to certified experts • Targeted: ask an question to a specific user • Pay a small amount of money to ask questions • Very popular Fenda Whale Campfire Yam 2,000,000 USD revenue 10,000,000 registered users Launched 500,000 paid questions 2016 July May June 6
How Fenda Works Price $10/Q What were causing my headache and nausea? Pay 10 USD Verified Give answer via audio Asker MD. Yang $7 cents $7 cents 14 cents USD to Listen 10% Income 10% Income Received: 7 cents * 1000 listeners = 70 USD Fenda has a unique monetary incentive model! Paid: Received*0.1 + question fee = 7 + 10 = 17 USD 1,000 Listeners Final Profit: Received – Paid = 70 - 17 = 53 USD 7
This Study • Research questions – How does monetary incentive affect Q&A? – Are there any manipulative behavior from users? – How does the pricing strategy affect users’ engagement? • Data driven analysis – Collect over 200K paid questions from two websites o Fenda (China), Whale (US) 8
Outline • Introduction • User behavior in targeted Q&A apps – Role of experts – Impact of monetary incentive • Manipulative behavior • Pricing strategy 9
Datasets Dataset #Questions Time Coverage #Users #Experts Fenda 212,000 05/16 – 07/16 30% 88,540 4,370 Whale 9,200 09/16 – 03/17 - 1,419 118 • Collect Fenda and Whale – Using open API with slow speed – Using data set of Whale as a comparison – Coverage is around 30% 1 • Experts are verified manually by websites – People can also ask questions to the normal users 10 1 Li Xuanmin. 2016. Putting a price on knowledge. http://www.globaltimes.cn/content/997510.shtml.
Role of Experts 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% # Users Revenue ($) # Answered Questions Normal Users Experts Experts are extremely important • Experts: 5% of total of users, contribute 95% of How does targeted Q&A service retain the experts? revenue and answer 82% of questions 11
Motivations on Q&A Service • Answerers are motivated to answer their questions by – Intrinsic reward (e.g. helping people) – Social reward (e.g. respect from others) – Extrinsic reward (e.g. money) • Targeted Q&A service primary use extrinsic reward • Existing research suggested extrinsic reward 1 may leads to – Less response delay – High answer quality Are they true on Fenda and Whale? 1 Haiyi Zhu, Sauvik Das, Yiqun Cao, Shuang Yu, Aniket Kittur, and Robert Kraut. A Market in Your Social Network: The Effects of 12 Extrinsic Rewards on Friendsourcing and Relationships. In Proc. of CHI (2016).
Short Response Time? • Fenda & Whale Average Response Time Hours 70 – Short response time 60 – But not shorter than Yahoo 50 answers 40 30 • Yahoo answers: large number 20 of potential answerers 10 0 • Fenda & Whale: target one Yahoo Fenda Whale Google Stack specific answerer 1 2 3 Answers Answers Overflow Targeted Q&A service are faster than most of crowdsourcing service 13
High Answer Quality? • High answer quality: Other people are also interested; Willing to pay for the answers • 56% of answers: Have at least one listener • Among these listened answers, 71% can make profits for askers (listening income > question fee) • Good question: good chance for making profits Majority of targeted Q&A service questions are high quality 14
Outline • Introduction • User behavior in targeted Q&A apps • Manipulative behavior – Bounty Hunters – Collaborative Users • Pricing strategy 15
Manipulative Behavior • Bounty hunters: users ask lots of questions for $ • Several types of experts to ask questions to: Type 3 Experts Type 1 Experts Type 2 Experts Many listeners Few listeners Many listeners Bounty hunter High price High price Low price No guarantee High chance High chance not earning Earning earning 16
Manipulative Behavior • Bounty hunters: users ask lots of questions for $ • Several types of experts to ask questions to: Ratio of Questions to Experts This outlier Average Asked 1300 Asked 2.5 questions questions Earned around Can not earn $200 ($-1.95) # of Questions that the user asked Act as spam to experts 17
Manipulative Behavior 2 • Collaborative users: Answerer and asker work together exclusively Increase Cheap questions perception of popularity Ask questions to User-2: Asker User-1: Answerer draw attentions Expensive questions 18 Other askers
Manipulative Behavior 2 • Collaborative users: Answerer and asker work # of Questions by same person User-1 answered 435 questions. User-2 asked User-1 307 questions. Together earn $950 Fake perception of popularity # of Answered questions 19
Outline • Introduction • User behavior in targeted Q&A apps • Manipulative behavior • Pricing strategy – How do users set their price of questions? – How does the pricing strategy affect their income and engagement? 20
Dynamic Pricing • Answerers can adjust their price dynamically • Examples: Too many askers Too few askers $3 $2 $1 User 1 Without change $10 User 2 • How many common pricing strategies are there? 21
Pricing Strategy • Cluster pricing history into groups • Construct 9 features based on the pricing history • For example: Top 3 features based on Chi-square id Feature Name Description 1 Price Change Frequency # of price change / # answers 2 Price Up Frequency # price up / # answers 3 Price Down Frequency # price down / # answers … … … Applied hierarchical clustering algorithm on these features User 1: [2/3, 1/3, 1/3, …] $3 $2 $1 22 User 1
Three Different Strategies • Got 3 groups (strategies) with highest modularity Group 1 frequently price up and down active users Improve users’ incomes and engagement inactive users Group 2 rarely price up and down Hurt users’ incomes and engagement Group 3 celebrities mostly price up Hurt users’ incomes and engagement 23
Conclusion • Targeted Q&A service – Short response time – High answer and question quality – Some manipulative behavior • Future Q&A work – Crowdsourcing v.s. Targeted – Add more dataset 24
Thank You 25
Reference • [1] Dan Wu and Daqing He. 2014. Comparing IPL2 and Yahoo! Answers: A Case Study of Digital Reference and Community Based Question Answering. In Proc. of Iconf. • [2] Benjamin Edelman. 2011. Earnings And Ratings At Google Answers. Economic Inquiry 50, 2 (2011), 309–320. • [3] Lena Mamykina, Bella Manoim, Manas Mittal, George Hripcsak, and Björn Hartmann. 2011. Design lessons from the fastest Q&A site in the west. In Proc. of CHI. 26
Recommend
More recommend
