Диссертация (1136614), страница 21
Текст из файла (страница 21)
The identified academic component of users’interest explains as much variation in their learning outcomes as the socioeconomic status that is measured by thePISA index of economic, social and cultural status (ESCS)(OECD 2013). While the obtained model does not allowreconstruction of raw scores reliably, it is able to discriminate between low- and high-performing students with ahigh degree of accuracy. Unlike previous research that focused on predicting academic achievements of students fromone university (Kassarnig et al. 2017; Lian et al.
2016;Wang et al. 2015), we were able to use data on a representative sample of Russian students.It should be taken into account that PISA is not an optimal instrument for the measurement of individual performance because its plausible values contain random errorvariance components. Instead, PISA was designed to provide accurate summary statistics about the population of interest within countries and about correlations between keyvariables (Jerrim et al.
2017). That fact might affect the performance of our model.We also find a significant digital divide among students.Subscribers to the World Arts and Culture (WAC) pagedemonstrate results that are much higher than results of theirpeers subscribed to Love Horoscope. The gap is roughlyequivalent to two years of formal schooling. The large variety of public page names (from Cool Gags to Science andMathReadingScience538 (4.6)521 (4.2)509 (4.0)507 (4.1)505 (3.9)530 (4.5)502 (4.1)507 (4.0)479 (4.3)492 (3.9)532 (4.3)516 (3.8)508 (3.9)504 (4.0)503 (3.7)473 (3.3)471 (5.1)467 (4.9)462 (5.1)450 (5.3)449 (3.4)447 (5.1)444 (5.1)450 (5.3)442 (5.8)472 (3.2)471 (4.7)465 (4.9)460 (5.0)453 (5.2)Technology) masks the fact that their content is rather homogenous.
Most of the pages, including WAC, post predominantly funny pictures and videos. Thus, there is little reasonto believe that subscription to these pages might significantlyaffect the academic performance of students. However, thesubscription to pages, the names of which suggest a relationto science and art, might serve a signaling purpose and playa role in students’ self-identification. Here, WAC provides anillustrative example because its tagline reads “Only intellectuals. Only hardcore”.The rate of adoption of large-scale datasets enabled by social media platforms is slower in education research than insome other areas of social science. This slow rate of adoption might be partly explained by the fact that one of themain variables of interest (i.e.
academic performance) isgenerally not available in these datasets to the researchers.The ability to infer the academic performance of social media users, even if imperfectly, might open new possibilitiesfor educational research. It also raises ethical concerns aboutthe potential use of such data. It has already been reportedthat some universities used VK data to target potential entrants by identifying high school graduates with “a keen interest to humanities” and planned to include prediction of“intelligence, creativity, and motivation” in the upcomingyear (Tomsk State University 2017). While screening candidates in the hiring process has already been discussed in theliterature (Landers and Schmidt 2016; Drouin et al.
2015;McDonald, Thompson, and O’Connor 2016), these newpractices require additional attention.MethodsSampleWe use data from the Russian Longitudinal Panel Study ofEducational and Occupational Trajectories (TrEC) (Kurakin2014). The study tracks 4,399 students from 42 Russian regions who took the PISA test in 2012. In 2016, publiclyavailable information from the social networking site VKwas collected for 3,483 TrEC participants who provided informed consent for the use of this data for research purposes.Table 3: Model performance in discrimination between different proficiency levels in reading measured as the area under theROC curve.Level 1 Level 2 Level 3 Level 4 Level 5 Level 6Bellow level 1 0.5980.7150.8400.9220.9540.992Level 10.6310.7750.8800.9260.980Level 20.6530.7850.8450.931Level 30.6550.7300.861Level 40.5790.739Level 50.674Note that while the initial sample was representative of 9thgrade high school Russian students in 2012, the social network data is not necessarily representative.VK DataVK is the largest European social networking site, with morethan 100 million active users.
It provides an applicationprogramming interface (API) that allows systematic downloading of public information from users’ profiles, includingtheir subscriptions to various public pages. The users fromour sample are subscribed to 73,389 different public pages.The median number of users’ subscription is 54. The maximum value is 1,000, because this is the largest number ofpages that could be returned by the VK API.
We excludepages that have less than ten subscribers and users with lessthan ten subscriptions, resulting in a sample of 2,637 usersand 4,485 pages.PISAInstead of a single point estimate of students’ abilities, PISAprovides five “plausible values” that are described in detailin its technical report (OECD 2014b). We deal with thesevalues in the following way. First, we build five linear regression models for each of the plausible values and averagetheir coefficients to obtain the final model. We then measurethe performance of the resulting model for each of the fiveplausible values and report its average performance.ModelFrom information about users’ subscriptions to public pages,we construct matrix A 2,637 x 4,486 so that Aij = 1 if useri is subscribed to public page j, and Aij = 0 otherwise.
Wethen use singular value decomposition of A to extract tenmain components representing user’s interests. These components were then used in a linear regression model to predict users’ PISA scores. The 10-fold cross-validation wasused to control for model overfitting. This approach is similar to the one used for Facebook data (Kosinski, Stillwell,and Graepel 2013) and makes the comparison of results possible. The use of non-linear models or increasing the numberof components do not provide any substantial increase in themodel’s performance.AcknowledgmentsThe data of the Russian panel study “Trajectories in Education and Career” (TrEC http://trec.hse.ru/) are presented inthis work. Support from the Basic Research Program of theNational Research University Higher School of Economicsis gratefully acknowledged.ReferencesAn, J., and Weber, I.
2016. # greysanatomy vs.# yankees:Demographics and hashtag use on twitter. In Tenth International AAAI Conference on Web and Social Media.Breakspear, S. 2014. How does PISA shape education policy making? Why how we measure learning determines whatcounts in education. In Centre for Strategic Education Seminar Series Paper, volume 40.Büchi, M.; Just, N.; and Latzer, M. 2016. Modelingthe second-level digital divide: A five-country study of social differences in internet use.
New media & Society18(11):2703–2722.DiMaggio, P., and Hargittai, E. 2001. From the digital divideto digital inequality: Studying internet use as penetration increases. Princeton: Center for Arts and Cultural Policy Studies, Woodrow Wilson School, Princeton University4(1):4–2.Drouin, M.; OConnor, K. W.; Schmidt, G. B.; and Miller,D. A. 2015. Facebook fired: Legal perspectives and youngadults opinions on the use of social media in hiring and firingdecisions. Computers in Human Behavior 46:123–128.Egelund, N.
2008. The value of international comparativestudies of achievement–a danish perspective. Assessment inEducation: Principles, Policy & Practice 15(3):245–251.Ertl, H. 2006. Educational standards and the changing discourse on education: The reception and consequences ofthe pisa study in germany. Oxford Review of Education32(5):619–634.Gebru, T.; Krause, J.; Wang, Y.; Chen, D.; Deng, J.; Aiden,E. L.; and Fei-Fei, L. 2017. Using deep learning and GoogleStreet View to estimate the demographic makeup of neighborhoods across the United States. Proceedings of the National Academy of Sciences 114(50):13108–13113.Jerrim, J.; Lopez-Agudo, L.
![](https://s.studizba.com/z.php?f=/templates/si/i/noavatar.png&w=100&h=100&t=1"){kind=avatar}
