Modeling Online User Interactions and their Offline effects on Socio-Technical Platforms

1. Modeling Online User Interactions and their Offline Effects on Socio-Technical Platforms Hitkul PhD Comprehensive - Jan 03, 2023 1

2. Committee Members 2

3. Interactions 3

4. Interactions 💬 3

5. Interactions 💬 📢 3

6. Interactions 💬 💰 📢 3

7. Interactions 💬 💰 📢 ✊ 3

8. Interactions 💬 💰 📢 🙏 ✊ 3

9. Interactions 💬 💰 📢 🙏 ✊ 🤝 3

13. Interactions 💬 💰 📢 🙏 ✊ 🤝 Individual Community Organizations 5

14. Interactions 💬 💰 📢 🙏 ✊ 🤝 Individual Community Organizations 6

22. Interactions 10

23. Interactions Online 10

24. Online Interactions 💬 💰 📢 🙏 ✊ 🤝 11

25. Online Interactions 💰 📢 🙏 ✊ 🤝 11

26. Online Interactions 💰 🙏 ✊ 🤝 11

27. Online Interactions 🙏 ✊ 🤝 11

28. Online Interactions ✊ 🤝 11

29. Online Interactions 🤝 11

30. Online Interactions 11

31. 12

32. 💬 💰 📢 🙏 ✊ 🤝 12

33. 💬 💰 📢 🙏 ✊ 🤝 12

34. 💬 💰 📢 🙏 ✊ 🤝 13

35. 💬 💰 📢 🙏 ✊ 🤝 Over fl ow 13

36. Over fl ow 14

37. Over fl ow - Positive 14

41. Over fl ow 15

42. Over fl ow - Negative 15

50. RESEARCH STATEMENT Quantifying effects of online interactions in of fl ine world and suggest implications 16

51. RESEARCH STATEMENT Quantifying effects of online interactions in of fl ine world and suggest implications 17

52. Our Focus 18

53. Our Focus 19

54. Our Focus Effects of popularity shocks 19

55. Our Focus Effects of popularity shocks Effects of drug disclosures 19

56. Our Focus Effects of popularity shocks Effects of drug disclosures RTO Detection for E-Commerce 19

58. Effect of Popularity Shocks on User Behavior ICWSM’ 22 Omkar Gurjar, Tanmay Bansal, Hitkul, Hemank Lamba, Ponnurangam Kumaraguru 20

59. What is Popularity Shock? 21

60. What is Popularity Shock? Time Views 21

61. What is Popularity Shock? Time Views Past trend First Shock 21

62. Why is Shock Interesting? 22

73. Related Work 25

74. Related Work Social Feedback Content Virality Popularity Shock 25

75. Related Work Social Feedback Content Virality Popularity Shock Positive reinforcement for improvements Widely studied of fl ine and online Few online exceptions 25

76. Related Work Social Feedback Content Virality Popularity Shock Positive reinforcement for improvements Widely studied of fl ine and online Few online exceptions Collaborative platforms - Wikipedia, Github Exogenous shocks Changes in network structures 25

77. Related Work Social Feedback Content Virality Popularity Shock Positive reinforcement for improvements Widely studied of fl ine and online Few online exceptions Collaborative platforms - Wikipedia, Github Exogenous shocks Changes in network structures Predicting Virality Characterizing content Information cascade 25

78. Research Questions Economies of Online Cooperations (Kollock et al. 1999) 26

79. Research Questions Economies of Online Cooperations (Kollock et al. 1999) Do users increase their posting frequency after receiving popularity shocks? 26

80. Research Questions Operant Conditioning (Skinner, 1938) 27

81. Research Questions Operant Conditioning (Skinner, 1938) Do users alter their content in response the Popularity shock? 27

82. Research Questions Longevity in Social Networks 28

83. Research Questions Longevity in Social Networks How long do effect of popularity shocks last? What leads to sustainability of popularity? 28

84. Research Questions Do users increase their posting frequency after receiving popularity shocks? Do users alter their content in response the Popularity shock? How long do effect of popularity shocks last? What leads to sustainability of popularity? 29

85. Data Collection 30

86. Data Collection Crawled pro fi les & fi ltered users with <200 posts 14 Catogries 43 hashtags 21,224 users sampling 33,490 users Authors of posts liked 4k posts each # 30

87. Data Collection Crawled pro fi les & fi ltered users with <200 posts 14 Catogries 43 hashtags 21,224 users sampling 33,490 users Authors of posts liked 4k posts each # 30,969 users 30,969 users 20 million Posts First post: 07/01/2015 Last post: 31/12/2021 30

88. Detecting Shocks 31

91. Detecting Shocks Z-Score 32

92. Detecting Shocks Z-Score Fb’s Prophet 32

93. Detecting Shocks Z-Score Fb’s Prophet Proposed Time Views 32

94. Detecting Shocks Z-Score Fb’s Prophet Proposed Time Views θ η 32

95. Detecting Shocks Z-Score Fb’s Prophet Proposed Time Views θ η 32

97. Detecting Shocks 1,000 Samples 33

98. Detecting Shocks 1,000 Samples 3 Annotators 33

99. Detecting Shocks 1,000 Samples 3 Annotators 0.6 Fleiss Kappa 33

100. Detecting Shocks 1,000 Samples 3 Annotators 0.6 Fleiss Kappa 34

102. Detecting Shocks 23.6% 42.5% 66.6% 35

103. Detecting Shocks Z-Score Fb’s Prophet Proposed 23.6% 42.5% 66.6% 35

109. Effect on Posting Frequency 37

110. Effect on Posting Frequency Regression Discontinuity in Time 37

114. Effect on Posting Frequency Regression Discontinuity in Time Control Variables: Age of User, and Intensity of user 38

115. Effect on Content 39

116. Effect on Content Custom Doc2vec Embeddings 39

121. Effect on Content Custom Doc2vec Embeddings αpre αpost βpre βpost 40

122. Effect on Content Custom Doc2vec Embeddings αpre αpost βpre βpost αpost > αpre 40

123. Effect on Content Custom Doc2vec Embeddings αpre αpost βpre βpost αpost > αpre βpost > βpre 40

124. Effect on Content Custom Doc2vec Embeddings αpre αpost βpre βpost αpost > αpre βpost > βpre βpost > αpost 40

125. Sustainability of Shock 41

126. Sustainability of Shock Survival Analysis / CoX Regression 41

127. Sustainability of Shock Survival Analysis / CoX Regression 42

128. Sustainability of Shock Survival Analysis / CoX Regression Popularity is generally short lived 42

129. Sustainability of Shock Survival Analysis / CoX Regression Popularity is generally short lived 42

130. Sustainability of Shock Survival Analysis / CoX Regression Popularity is generally short lived Post Frequency Post Frequency 42

131. Sustainability of Shock Survival Analysis / CoX Regression Popularity is generally short lived Post Frequency Post Frequency Community Engagement 42

132. Sustainability of Shock Survival Analysis / CoX Regression Popularity is generally short lived Post Frequency Post Frequency Community Engagement Balance Content Similarity 42

133. Implications 43

137. Contributions 44

138. Contributions First paper to study user behaviour after popularity shocks on short video platform 44

139. Contributions First paper to study user behaviour after popularity shocks on short video platform Actionable insights for users to sustain popularity 44

140. Contributions First paper to study user behaviour after popularity shocks on short video platform Actionable insights for users to sustain popularity First dataset to provide posts and accompanying metadata for a short video platform 44

143. Effect of Feedback on Drug Consumption Disclosures on Social Media R&R at a top tier conference Hitkul, Rajiv Ratn Shah, Ponnurangam Kumaraguru 46

144. 47 Drug Consumption Subreddits

145. 48 Drug Consumption Disclosure

146. 48 Drug Consumption Disclosure

147. Related Work 49

148. Related Work Drug on social media Self-harm behavior on social media Causal inference on social data 49

149. Related Work Drug on social media Self-harm behavior on social media Causal inference on social data Addiction detection Trend analysis Forecast drug overdose, abuse 49

150. Related Work Drug on social media Self-harm behavior on social media Causal inference on social data Addiction detection Trend analysis Forecast drug overdose, abuse Simulating RCT using online social data Applications have been shown in weight loss, alcohol consumption, content quality 49

151. Related Work Drug on social media Self-harm behavior on social media Causal inference on social data Addiction detection Trend analysis Forecast drug overdose, abuse Simulating RCT using online social data Applications have been shown in weight loss, alcohol consumption, content quality Self-harm to improve social standing Online challenges e.g. KiKi Demographic features 49

152. Research Questions 50 Impression Management (Goffman 1959) (Leary et al. 1994) (Hogan 2010)

153. Research Questions What is the extent of content indicating of fl ine drug consumption? 50 Impression Management (Goffman 1959) (Leary et al. 1994) (Hogan 2010)

154. Research Questions Primacy Effect (Asch 1946) (Murdock Jr 1962) 51

155. Research Questions Primacy Effect (Asch 1946) (Murdock Jr 1962) How does the community feedback on fi rst drug consumption post affect users' future drug consumption? 51

156. Research Questions Operant Conditioning (Skinner, 1938) 52

157. Research Questions Operant Conditioning (Skinner, 1938) How does continuous positive community feedback affect users' future drug consumption? 52

158. Research Questions What is the extent of content indicating of fl ine drug consumption? How does the community feedback on fi rst drug consumption post affect users' future drug consumption? How does continuous positive community feedback affect users' future drug consumption? Can we use Reddit textual data to classify between drug consumption and non-drug consumption content? 53

159. Data 54

160. Data 54 10 Subreddits

161. Data 54 10 Subreddits 826,905 Posts 6.6M Comments

162. Data 54 493,906 Users 10 Subreddits 826,905 Posts 6.6M Comments

163. Data - User Study 55

164. Data - User Study 55 Anonymous

165. Data - User Study 55 Anonymous Likert Scale

166. Data - User Study 55 45 Participants Anonymous Likert Scale

167. Detecting Drug Disclosure 56

168. Detecting Drug Disclosure Deep Learning Classi fi er 56

170. Detecting Drug Disclosure Deep Learning Classi fi er 57 4,000 Samples

171. Detecting Drug Disclosure Deep Learning Classi fi er 57 4,000 Samples 3 Annotators

172. Detecting Drug Disclosure Deep Learning Classi fi er 57 4,000 Samples 3 Annotators 0.7 Fleiss Kappa

174. Detecting Drug Disclosure Deep Learning Classi fi er 58 Individual subreddit performance equally good

175. 59 Extent of Drug Consumption

176. 59 Extent of Drug Consumption 80% indicate drug consumption (84% in user study) 54% Users Posts Comments 84% indicate drug consumption indicate drug consumption

177. Effect of Community Feedback 60

178. Effect of Community Feedback Propensity Score Matching 60

181. Effect of Community Feedback Propensity Score Matching 61 Treatment Group Control Group

182. Effect of Community Feedback Propensity Score Matching 62 Treatment Group Control Group

183. Effect of Community Feedback Propensity Score Matching 62 Treatment Group Control Group All Pairs

184. Effect of Community Feedback Propensity Score Matching 63 Treatment Group Control Group All Pairs

185. Effect of Community Feedback Propensity Score Matching 63 Treatment Group Control Group All Pairs Propensity matching

186. Effect of Community Feedback Propensity Score Matching 64 Treatment Group Control Group All Pairs Propensity matching

187. Effect of Community Feedback Propensity Score Matching 64 Treatment Group Control Group All Pairs Propensity matching Balance Confounders

188. Effect of Community Feedback Propensity Score Matching 65 Treatment Group Control Group All Pairs Propensity matching Balance Confounders Treatment Effect

190. Effect of Community Feedback Propensity Score Matching 66 Treatment Group Control Group All Pairs Propensity matching Balance Confounders Treatment Effect Treatment Threshold on # of comments or scores 80% receive < 1.1 ± 0.3 comments < 2.9 ± 1.1 score We experiment with [2,6]

191. Effect of Community Feedback Propensity Score Matching 66 Treatment Group Control Group All Pairs Propensity matching Balance Confounders Treatment Effect Treatment Threshold on # of comments or scores 80% receive < 1.1 ± 0.3 comments < 2.9 ± 1.1 score We experiment with [2,6] Text Propensity Content text is a confounder Text based propensity is a developing fi eld Keith et al. 2020, Roberts et al. 2020, Weld et al. 2022 RoBERTa based model

192. Effect of Community Feedback Propensity Score Matching 66 Treatment Group Control Group All Pairs Propensity matching Balance Confounders Treatment Effect Treatment Threshold on # of comments or scores 80% receive < 1.1 ± 0.3 comments < 2.9 ± 1.1 score We experiment with [2,6] Text Propensity Content text is a confounder Text based propensity is a developing fi eld Keith et al. 2020, Roberts et al. 2020, Weld et al. 2022 RoBERTa based model Confounders Past activity, Past feedback SMD <0.25 to ensure balance xtreatment − xcontrol σ2 treatment + σ2 control 2

193. Effect of Community Feedback Propensity Score Matching 66 Treatment Group Control Group All Pairs Propensity matching Balance Confounders Treatment Effect Treatment Threshold on # of comments or scores 80% receive < 1.1 ± 0.3 comments < 2.9 ± 1.1 score We experiment with [2,6] Text Propensity Content text is a confounder Text based propensity is a developing fi eld Keith et al. 2020, Roberts et al. 2020, Weld et al. 2022 RoBERTa based model Confounders Past activity, Past feedback SMD <0.25 to ensure balance xtreatment − xcontrol σ2 treatment + σ2 control 2 Effect Size ∑ N i=1,j=1 (Yi(T = 1) − Yj(T = 0)) * 100 Yj(T = 0) N

195. Effect of Community Feedback Propensity Score Matching 67 Treatment Group Control Group All Pairs Propensity matching Balance Confounders Treatment Effect First and continues positive feedback increased drug consumption by upto 2x

196. Effect of Community Feedback Propensity Score Matching 67 Treatment Group Control Group All Pairs Propensity matching Balance Confounders Treatment Effect First and continues positive feedback increased drug consumption by upto 2x User belief Scores 2.2/5 Comments 2.5/5 (Little to Moderate impact) “Illusion of control” (Lyng 1990)

198. Implications 68 Feedback mechanisms

199. Implications 68 Feedback mechanisms Intervention

200. Implications 68 Feedback mechanisms Intervention Resource

201. Contributions 69

202. Contributions First paper to study effect of community feedback on drug consumption disclosure 69

203. Contributions First paper to study effect of community feedback on drug consumption disclosure Actionable insights for moderators and platforms 69

204. Contributions First paper to study effect of community feedback on drug consumption disclosure Actionable insights for moderators and platforms First dataset to provide drug consumption indicative posts and model checkpoints 69

207. Social Re-Identi fi cation Assisted RTO Detection for E-Commerce Under review at a top tier conference Hitkul, Abinaya, Soham Saha, Satyajit Banerjee, M. Chellaiah, Ponnurangam Kumaraguru In collaboration with 71

208. 72 Return to Origin (RTO)

211. 72 Return to Origin (RTO) Order Cancelled

212. 72 Return to Origin (RTO) Order Return to Origin

213. 72 Return to Origin (RTO) Order Delivered

214. Order Return to Origin 73 Return to Origin (RTO)

215. 74 Why Study RTO?

216. ₹ ₹ ₹ ₹ ₹ ₹ ₹ ₹ ₹ ₹ ₹ ₹ ₹ 74 Why Study RTO?

217. ₹18 Cr. saved for 0.5 increase in recall 74 Why Study RTO?

218. 75 Hypothesis

219. 75 Hypothesis FK user

220. 75 Hypothesis FK user Internal data

221. 75 Hypothesis FK user Internal data Encoded/derived features

222. 75 Hypothesis FK user Internal data Encoded/derived features Model θ Perfromance Metric

223. 75 Hypothesis FK user Internal data Encoded/derived features Model θ Perfromance Metric Corresponding social pro fi le

224. 75 Hypothesis FK user Internal data Encoded/derived features Model θ Perfromance Metric Corresponding social pro fi le Social data

225. 75 Hypothesis FK user Internal data Encoded/derived features Model θ Perfromance Metric Corresponding social pro fi le Social data Encoded/derived features

228. 76 Motivation - Sociology

232. 77 Motivation - CSS

235. 78 Proposed Flow Data collection Data validation Modeling

237. 80 Data Collection

238. 80 Data Collection <Name, Location>

239. 80 Data Collection API <Name, Location>

240. 80 Data Collection API <Name, Location>

241. 80 Data Collection API <Name, Location> Potential PII, platform violations

242. 80 Data Collection API <Name, Location> Potential PII, platform violations Only public data Academic use only

243. 81 Data Collection * November 2021 to April 2022, Only Lifestyle org

244. 81 Data Collection RTO Users * November 2021 to April 2022, Only Lifestyle org

245. 81 Data Collection RTO Users Review Users * November 2021 to April 2022, Only Lifestyle org

246. 81 Data Collection RTO Users Review Users * November 2021 to April 2022, Only Lifestyle org Potential Users

247. 82 Data Collection

248. 82 Data Collection 25,725 rows 6,052 users

249. 82 Data Collection 25,725 rows 6,052 users 7,833 rows 2,460 users Only COD

250. 82 Data Collection 25,725 rows 6,052 users 7,833 rows 2,460 users 6,881 rows 2,121 users Only COD Available publicly

251. 82 Data Collection 25,725 rows 6,052 users 7,833 rows 2,460 users 6,881 rows 2,121 users 3,634 rows 1,091 users Only COD Available publicly Possible social match

252. 83 Data Source 80.89% 30.69% 10.37% 8.96%

253. 83 Data Source 80.89% 30.69% 10.37% 8.96%

256. 86 Re-Identification Validation

257. 86 Re-Identification Validation City intersection to validate matches

259. 87 FK user Re-Identification Validation

260. 87 FK user Re-Identification Validation {Bengaluru, Delhi, Rewari, Jamshedpur}

261. 87 FK user Re-Identification Validation {Bengaluru, Delhi, Rewari, Jamshedpur}

262. 87 FK user Re-Identification Validation {Bengaluru, New Delhi, Tokyo, Jamshedpur} {Bengaluru, Delhi, Rewari, Jamshedpur}

263. 87 FK user Re-Identification Validation {Bengaluru, New Delhi, Tokyo, Jamshedpur} {Bengaluru, Delhi, Rewari, Jamshedpur} ∩

264. 87 FK user Re-Identification Validation {Bengaluru, New Delhi, Tokyo, Jamshedpur} {Bengaluru, Delhi, Rewari, Jamshedpur} ∩ {Bengaluru, Delhi, Jamshedpur}

268. 91 Features - Internal

269. 91 Features - Internal Past trends of 3 months and 1 year Ratio of RTO/Total orders User Product Seller Product category Pincode Street name City Hour of the day Day of the week

270. 92 Features - Social

271. 92 Features - Social

272. 92 Features - Social Is student Is direct link available

273. 92 Features - Social Is student Is direct link available # of jobs

274. 92 Features - Social Is student Is direct link available # of jobs # of educations

275. 92 Features - Social Is student Is direct link available # of jobs # of educations Years of jobs

276. 92 Features - Social Is student Is direct link available # of jobs # of educations Years of jobs Years of education

277. 92 Features - Social Is student Is direct link available # of jobs # of educations Years of jobs Years of education # of skills

278. 92 Features - Social Is student Is direct link available # of jobs # of educations Years of jobs Years of education # of skills # of connections

279. 92 Features - Social Is student Is direct link available # of jobs # of educations Years of jobs Years of education # of skills # of connections # of followers

280. 93 Features - Language

281. 93 Features - Language Data Science Intern Flipkart

282. 93 Features - Language Data Science Intern Flipkart Doctor of Philosophy Indraprastha Institute of Information Technology

283. 93 Features - Language Sentence transformer 387 vector Data Science Intern Flipkart Doctor of Philosophy Indraprastha Institute of Information Technology

284. 93 Features - Language Sentence transformer 387 vector PCA Reduced vector Data Science Intern Flipkart Doctor of Philosophy Indraprastha Institute of Information Technology

285. 94 Evaluation Goodness = P P + N − P − PPred and True P + N − PPred × 104 FPR = PPred − PPred and True PPred

286. 95 Results

287. 95 Results

288. 95 Results

289. 96 Results

291. 98 Challenges & Future Scope Data collection Data validation Modeling

292. 98 Challenges & Future Scope Data collection Data validation Modeling Remove review intersection dependency Scale crawling setup (potentially via VMs) Collect data from platforms apart from LinkedIn

293. 98 Challenges & Future Scope Data collection Data validation Modeling Remove review intersection dependency Scale crawling setup (potentially via VMs) Collect data from platforms apart from LinkedIn Heuristics for edge cases e.g. Chennai vs Madras Add other modalities of validation Language Images Network behaviour

294. 98 Challenges & Future Scope Data collection Data validation Modeling Remove review intersection dependency Scale crawling setup (potentially via VMs) Collect data from platforms apart from LinkedIn Heuristics for edge cases e.g. Chennai vs Madras Add other modalities of validation Language Images Network behaviour More features Finetuned language features Nonlinear dimensionality reduction (UMAP) Categorical embeddings Pre-trained tasks and vectors Transfer learning to other tasks

295. 99 Effects of popularity shocks Effects of drug disclosures RTO Detection for E-Commerce

296. 100 Effects of popularity shocks Effects of drug disclosures RTO Detection for E-Commerce

297. 100 Effects of popularity shocks Effects of drug disclosures RTO Detection for E-Commerce Timeline

298. 101 Timeline

299. 101 Timeline

300. 101 Timeline Effects of drug disclosures RTO Detection for E-Commerce

301. 101 Timeline Effects of drug disclosures RTO Detection for E-Commerce 🦠

302. 101 Timeline Effects of drug disclosures RTO Detection for E-Commerce 🦠 ?

303. 101 Timeline Effects of drug disclosures RTO Detection for E-Commerce 🦠 ?

304. Acknowledgment 102

305. Thank You! Do you have any questions? 103 hitkuli@iiitd.ac.in hitkuljangid @Hitkul_

Modeling Online User Interactions and their Offline effects on Socio-Technical Platforms

You are reading a preview.

Check these out next

Modeling Online User Interactions and their Offline effects on Socio-Technical Platforms

More Related Content

Similar to Modeling Online User Interactions and their Offline effects on Socio-Technical Platforms (20)

More from IIIT Hyderabad (20)

Recently uploaded (20)