Linking Smart Cities Datasets with Human Computation - the case of UrbanMatch
0 likes969 views
This document summarizes a study that links datasets from smart cities using human computation through a game called UrbanMatch. UrbanMatch links points of interest (POIs) from OpenStreetMap to representative photos from Wikimedia Commons and social media by having online players group photos with POIs. The game achieves high accuracy (>99%) and throughput (485 links/hour) in linking photos to POIs, showing the effectiveness of citizen computation games. However, engagement needs improvement as games averaged only 3 minutes. Overall, the study shows how human capabilities can be leveraged through games to link geospatial datasets.
![The Record Linkage Problem [Fellegi&Sunter,1969]
Γ=A×B Comparison space
B Γ= 𝐴× 𝐵
Purpose is to split Γ in
A M (set of matches) and
U (set of non-matches)
Each link γ has a score 𝑠 𝛾
𝑠 𝛾 = 𝑃 𝛾 ∈ Γ 𝑀 /𝑃(𝛾 ∈ Γ|𝑈)
B
Partitioning based on
lower/upper thresholds:
A 𝑠 𝛾 > 𝑈𝑃𝑃𝐸𝑅 ⇒ 𝛾 ∈ 𝑀
𝑠 𝛾 < 𝐿𝑂𝑊𝐸𝑅 ⇒ 𝛾 ∈ 𝑈
∈M 𝐿𝑂𝑊𝐸𝑅 ≤ 𝑠 𝛾 ≤ 𝑈𝑃𝑃𝐸𝑅 ⇒ 𝛾 ∈ 𝐶
∈U (candidate links to be assessed by experts)
∈C
UrbanMatch - ISWC 2012 4 Boston, 13th November 2012](https://image.slidesharecdn.com/urbanmatch-iswc-2012-121113151540-phpapp01/85/Linking-Smart-Cities-Datasets-with-Human-Computation-the-case-of-UrbanMatch-4-320.jpg)
![Link score in UrbanMatch
Effect of players' evidences on the score sγ
If there is a positive evidence (an UrbanMatch player says that the
link γ is correct), sγ is increased at most by Kpos
If there is a negative evidence (an UrbanMatch player says that the
link γ is incorrect), sγ is decreased at most by Kneg
Each evidence is weighted with a factor that represents UrbanMatch
player's reliability
𝜀
− 𝑝 2
𝑟𝑝 = 𝑒
where 𝜀 𝑝 is the number of errors made by the player, thus 𝑟 𝑝 ∈ [0. . 1]
(reliability is 1 if the player makes no mistake, 0.6 if he makes 1 error,
almost zero if he makes 6 mistakes or more)
For each link γ in C, its score varies as follows
Positive evidence: 𝑠′ 𝛾 = 𝑠 𝛾 + 𝐾 𝑝𝑜𝑠 ∗ 𝑟 𝑝
Negative evidence: 𝑠′ 𝛾 = 𝑠 𝛾 − 𝐾 𝑛𝑒𝑔 ∗ 𝑟 𝑝
UrbanMatch - ISWC 2012 10 Boston, 13th November 2012](https://image.slidesharecdn.com/urbanmatch-iswc-2012-121113151540-phpapp01/85/Linking-Smart-Cities-Datasets-with-Human-Computation-the-case-of-UrbanMatch-10-320.jpg)
Linking Smart Cities Datasets with Human Computation - the case of UrbanMatch
- 1. Linking Smart Cities Datasets with Human Computation: the case of UrbanMatch Irene Celino, Simone Contessa, Marta Corubolo, Daniele Dell’Aglio, Emanuele Della Valle, Stefano Fumeo and Thorsten Krüger CEFRIEL – Politecnico di Milano – SIEMENS UrbanMatch - ISWC 2012
- 2. Citizen Computation Human Computation Citizen Science exploiting human capabilities exploiting volunteers to solve computational tasks to collect scientific data or difficult for machines to conduct experiments "in the world" Citizen Computation exploiting human capabilities to contribute to a mixed computational system by living "in the world" UrbanMatch - ISWC 2012 2 Boston, 13th November 2012
- 3. UrbanMatch: Citizen Computation GWAP Photos from social media Places & POIs from OpenStreetMap Wikimedia Commons ? e.g. Duomo e.g. equestrian statue ? ? Which photos actually represent the urban POIs? Can we link POIs to their most representative photos? UrbanMatch - ISWC 2012 3 Boston, 13th November 2012
- 4. The Record Linkage Problem [Fellegi&Sunter,1969] Γ=A×B Comparison space B Γ= 𝐴× 𝐵 Purpose is to split Γ in A M (set of matches) and U (set of non-matches) Each link γ has a score 𝑠 𝛾 𝑠 𝛾 = 𝑃 𝛾 ∈ Γ 𝑀 /𝑃(𝛾 ∈ Γ|𝑈) B Partitioning based on lower/upper thresholds: A 𝑠 𝛾 > 𝑈𝑃𝑃𝐸𝑅 ⇒ 𝛾 ∈ 𝑀 𝑠 𝛾 < 𝐿𝑂𝑊𝐸𝑅 ⇒ 𝛾 ∈ 𝑈 ∈M 𝐿𝑂𝑊𝐸𝑅 ≤ 𝑠 𝛾 ≤ 𝑈𝑃𝑃𝐸𝑅 ⇒ 𝛾 ∈ 𝐶 ∈U (candidate links to be assessed by experts) ∈C UrbanMatch - ISWC 2012 4 Boston, 13th November 2012
- 5. UrbanMatch Linkage Problem B (photos) A contains POIs A (POIs) B contains photos Piazza del Duomo Γ contains POI-photo links "Playable places" group Piazza della POIs partitioning of Γ Scala (problem space reduction) Input data in UrbanMatch: UrbanMatch links 34 POIs in 14 playable places 11,089 photos 196 links in M 382 links in U 37,413 links in C UrbanMatch - ISWC 2012 5 Boston, 13th November 2012
- 6. Showing links to UrbanMatch players The POI-photo links to be evaluated could be shown directly to players... Santa Maria correct? delle Grazie ??? ...but the player could not know the name of the POI On the other hand, the player can easily "see" (Santa Maria correct? (Sant'Ambrogio) delle Grazie) NO! thus we use a "sure" photo as proxy for the POI, thus the UrbanMatch game is designed as a photo-pairing challenge UrbanMatch - ISWC 2012 6 Boston, 13th November 2012
- 7. UrbanMatch level definition Set B of photos Set A of POIs a b c POI 1 d (Duomo) h POI 2 (statue) g f e Photos d and g depict POI 1 (and do not depict POI 2) Photos c and h depict POI 2 (and do not depict POI 1) Photos b and f do not depict POI 1 nor POI 2 (distractors) Photos a and e maybe depict POI 1 and/or POI 2 UrbanMatch - ISWC 2012 7 Boston, 13th November 2012
- 8. Feedbacks to UrbanMatch players If the selected pair d is correct (two sure photos) or plausible, UrbanMatch gives a positive feedback e b If the selected pair is wrong (at least one selected photo is a "distractor" option), UrbanMatch gives a negative feedback f (and counts the player's mistake) UrbanMatch - ISWC 2012 8 Boston, 13th November 2012
- 9. UrbanMatch: gameplay Video at: http://bit.ly/um-video UrbanMatch - ISWC 2012 9 Boston, 13th November 2012
- 10. Link score in UrbanMatch Effect of players' evidences on the score sγ If there is a positive evidence (an UrbanMatch player says that the link γ is correct), sγ is increased at most by Kpos If there is a negative evidence (an UrbanMatch player says that the link γ is incorrect), sγ is decreased at most by Kneg Each evidence is weighted with a factor that represents UrbanMatch player's reliability 𝜀 − 𝑝 2 𝑟𝑝 = 𝑒 where 𝜀 𝑝 is the number of errors made by the player, thus 𝑟 𝑝 ∈ [0. . 1] (reliability is 1 if the player makes no mistake, 0.6 if he makes 1 error, almost zero if he makes 6 mistakes or more) For each link γ in C, its score varies as follows Positive evidence: 𝑠′ 𝛾 = 𝑠 𝛾 + 𝐾 𝑝𝑜𝑠 ∗ 𝑟 𝑝 Negative evidence: 𝑠′ 𝛾 = 𝑠 𝛾 − 𝐾 𝑛𝑒𝑔 ∗ 𝑟 𝑝 UrbanMatch - ISWC 2012 10 Boston, 13th November 2012
- 11. Evaluating UrbanMatch precision Accuracy definition (metrics specific to Data Quality) candidate links correctly assessed to game ability to correctly be either trustable or incorrect assess candidate links 𝐶𝑀 − 𝐹𝑃 + (𝐶𝑈 − 𝐹𝑁) 𝐴𝑐𝑐𝑢𝑟𝑎𝑐𝑦 = 𝐶𝑀 + 𝐶𝑈 candidate links assessed to be either trustable or incorrect CM = right links, i.e. candidate links moved from C to M (because 𝑠 𝛾 > 𝑈𝑃𝑃𝐸𝑅) CU = wrong links, i.e. candidate links moved from C to U (because 𝑠 𝛾 < 𝐿𝑂𝑊𝐸𝑅) FP = false positives (candidate links assessed as trustable but actually incorrect) FN = false negatives (candidate links assessed as incorrect but actually trustable) accuracy requires ground truth computed only on a manually analyzed set of links for evaluation's sake ! UrbanMatch - ISWC 2012 11 Boston, 13th November 2012
- 12. Evaluating UrbanMatch as a GWAP Throughput and ALP definition (metrics specific to Games with a Purpose) game ability to assess 𝑆𝑜𝑙𝑣𝑒𝑑𝑃𝑟𝑜𝑏𝑙𝑒𝑚𝑠 candidate links 𝑇ℎ𝑟𝑜𝑢𝑔ℎ𝑝𝑢𝑡 = 𝑃𝑙𝑎𝑦𝑒𝑑𝑇𝑖𝑚𝑒 game ability to 𝑃𝑙𝑎𝑦𝑒𝑑𝑇𝑖𝑚𝑒 engage players 𝐴𝐿𝑃 = 𝐴𝑐𝑡𝑖𝑣𝑒𝑃𝑙𝑎𝑦𝑒𝑟𝑠 where 𝑆𝑜𝑙𝑣𝑒𝑑𝑃𝑟𝑜𝑏𝑙𝑒𝑚𝑠 = 𝐶𝑀 + 𝐶𝑈 in which: CM = right links, i.e. candidate links moved from C to M (because 𝑠 𝛾 > 𝑈𝑃𝑃𝐸𝑅) CU = wrong links, i.e. candidate links moved from C to U (because 𝑠 𝛾 < 𝐿𝑂𝑊𝐸𝑅) UrbanMatch - ISWC 2012 12 Boston, 13th November 2012
- 13. Setting the thresholds for link score 𝒔 𝜸 Accuracy and Throughput as a function of 𝐿𝑂𝑊𝐸𝑅 LOWER 0.05 0.10 0.15 0.20 CU 321 348 1152 1216 FN 4 5 7 8 Accuracy 98.75 % 98.56 % 99.39 % 99.34 % Throughput 108.08 117.17 387.87 409.42 Accuracy and Throughput as a function of 𝑈𝑃𝑃𝐸𝑅 UPPER 0.60 0.65 0.70 0.75 0.80 0.85 0.90 CM 227 225 68 65 61 60 49 FP 48 47 4 4 3 3 1 Accuracy 78.85 % 79.11 % 94.11 % 95.38 % 95.08 % 95.00 % 97.95 % Throughput 76.43 75.75 22.89 21.88 20.53 20.20 16.49 UrbanMatch - ISWC 2012 13 Boston, 13th November 2012
- 14. UrbanMatch evaluation results Accuracy, Throughput and ALP results Evaluation data 54 unique players, 290 games (781 game levels) 2006 input links, 1284 assessed (trustable/incorrect) upper threshold 0.70, lower threshold 0.20 Accuracy (on the manually-checked subset) 99.06% very good result Throughput and ALP (on all processed links) 485 links/h very good result 3m 17s one-time game, needs improvement UrbanMatch - ISWC 2012 14 Boston, 13th November 2012
- 15. Evaluation UrbanMatch as a whole User-based evaluation: actual "engagement" Based on game evaluation literature and integrated with our research-specific questions Questionnaire at http://bit.ly/um-survey Findings: UrbanMatch - ISWC 2012 15 Boston, 13th November 2012
- 16. Conclusions Citizen Computation games are effective to leverage linking capabilities of "humans" Game is to be designed to keep in consideration users' capabilities and usage context: (lack of) background knowledge small mobile-phone screen Different tasks to be solved need different game "missions" Linking task pairing mission Verifying task rating/selecting mission Collecting task selecting/editing mission Advertisement try out Urbanopoly game here in Boston, to collect and assess geo-spatial linked data participant to Semantic Web Challenge http://bit.ly/urbanopoly UrbanMatch - ISWC 2012 16 Boston, 13th November 2012
- 17. Thanks for your attention! Any question? Irene Celino – CEFRIEL, ICT Institute Politecnico di Milano email: Irene.Celino@cefriel.it – web: http://swa.cefriel.it UrbanMatch - ISWC 2012