PaperlayerView exact revision on arXiv

Source-linked AI summary

Computational Socioeconomics

Jian Gao, Yi-Cheng Zhang, Tao Zhou

arXiv:1905.06166v1physics.soc-phecon.GNq-fin.EC

TL;DR

Socioeconomic research must address complex human behavior and gaps in traditional data, which can be small, biased, costly, or untimely. This review presents Computational Socioeconomics as an approach combining large-scale real-world data, computational analysis, and small-sample information to infer population-level socioeconomic attributes. It surveys findings and applications drawn from these data, while stressing that their coverage and regional validity are limited.

  • Problem

    Complex, heterogeneous social processes are difficult to test and predict, and conventional surveys and censuses can be small, biased, costly, or untimely.

    Source: §1 · pp. 3–4

  • Method

    The review synthesizes computational socioeconomics research using large-scale data and computational methods, including small-sample training to infer hard-to-obtain attributes across population-scale data.

    Source: §1 · p. 5

  • Results

    The review presents examples in which large-scale real data and advanced analysis yield socioeconomic insights that traditional methods may not readily obtain.

    Source: §6 · p. 94

  • Takeaways & Limitations

    The authors expect the data- and AI-driven methodological perspective to become mainstream and affect research across social science and other qualitative disciplines.

    Source: §1 · p. 5

    Large digital datasets are not random population samples, and findings may not transfer across regions with different religious, cultural, or political contexts.

    Source: §6 · p. 94

Abstract

from arXiv · show

Uncovering the structure of socioeconomic systems and timely estimation of socioeconomic status are significant for economic development. The understanding of socioeconomic processes provides foundations to quantify global economic development, to map regional industrial structure, and to infer individual socioeconomic status. In this review, we will make a brief manifesto about a new interdisciplinary research field named Computational Socioeconomics, followed by detailed introduction about data resources, computational tools, data-driven methods, theoretical models and novel applications at multiple resolutions, including the quantification of global economic inequality and complexity, the map of regional industrial structure and urban perception, the estimation of individual socioeconomic status and demographic, and the real-time monitoring of emergent events. This review, together with pioneering works we have highlighted, will draw increasing interdisciplinary attentions and induce a methodological shift in future socioeconomic studies.

1. Introduction

Social science faces unusually complex, uncertain subjects and limited means to test theories or predict outcomes, while conventional data sources are often small, biased, costly, or untimely. The review frames Computational Socioeconomics as a quantitative approach using large-scale real data and computational methods, including models that infer hard-to-obtain population-wide attributes from accessible data and small labeled samples.

1. Introduction

  • Human heterogeneity, unstable interacting factors, and limited controlled experimentation make quantitative testing and prediction especially difficult in social science.The passage contrasts this with physical science and notes that incorrect predictions may be attributed to unknown factors rather than flaws in theories.Source: §1 · p. 3
  • Questionnaires can be small and socially biased, while censuses are resource-intensive and often untimely or infeasible in poor regions.Digital sensing and online platforms now record social, economic, and human-activity data at larger scale, higher resolution, and near real time.Source: §1 · pp. 3–4
  • Computational Socioeconomics studies socioeconomic development with quantitative methods based on large-scale real data, focusing on economic problems tied to social processes and social problems tied to economic development.The authors describe it as an immature but burgeoning interdisciplinary field.Source: §1 · p. 4
  • A small labeled sample can train a model to infer hard-to-get attributes from accessible population-scale data and extend those estimates across the population.The review gives household income inferred from mobile-phone data as an example, while noting inferred values are imperfect and algorithm-dependent.Source: §1 · p. 5, Figure 1 · p. 5
  • The review gathers dispersed research, organizes it by study objects and data, and aims to unify descriptions and terminology in this emerging field.Its planned coverage spans macro, urban, and individual scales, emergency management, and data resources including remote sensing, mobile phones, and social media.Source: §1 · pp. 5–6

2.1. World development and poverty mapping

This section reviews how satellite imagery, mobile-phone records, and combined data are used to map poverty and infer socioeconomic conditions across scales.

2.1.1. Remote sensing observes poverty

  • Satellite night lights have been used to proxy economic activity and map poverty, but uniformly dark imagery limits their effectiveness in less-developed areas.Night-time light measures correlate with GDP and electricity use; daytime imagery and deep learning offer an approach for areas with uniformly dark night-time imagery.Source: §2.1.1 · pp. 6–8
  • CNN-extracted daytime satellite-image features explained up to 75% of average household asset variation across five African countries.The method transfers knowledge from ImageNet object recognition to predict night-light intensity, then uses that knowledge for poverty prediction; it can also reconstruct survey-based regional poverty indicators with high accuracy.Source: §2.1.1 · p. 8, Figure 3 · p. 8

2.1.2. Mobile phones reveal socioeconomic status

  • In Rwanda, call-detail records and surveys supported wealth predictions for 1.5 million users, whose district-level estimates strongly correlated with government ground-truth data.The approach used feature generation followed by elastic-net selection of a parsimonious model.Source: §2.1.2 · p. 10, Figure 4 · p. 10

2.1.3. Combined data for better inference

  • Combined CDR-RS models achieved highest R2 = 0.78, versus RS-only R2 = 0.71 and CDRs-only R2 = 0.70, for Bangladesh living-standard maps.In Rwanda, a four-feature model explained 76% of MPI variance across 295 sectors.Source: §2.1.3 · pp. 10–11, Figure 5 · p. 11
  • Across six international networks, global multiplex degree performed consistently well across 14 socioeconomic indicators and had the strongest negative correlation with HDI ranking (Spearman’s rank correlation ρ ≈0.8).The networks represent trade, postal, migration, international flights, IP, and digital communications.Source: §2.1.3 · pp. 11–12

2.2. Economic complexity and fitness of nations

This section reviews trade-based measures of economic complexity and country fitness, and how they relate to development, growth prospects, and economic dynamics. It also surveys applications to inequality, industrialization, and economic forecasting.

2.2.1. Product space and economic complexity

  • Countries with initially sophisticated export baskets tend to achieve stronger subsequent economic growth, according to analysis of more than 5,000 products and 124 countries.EXPY represents the sophistication of a country's export basket.Source: §2.2.1 · p. 12
  • The product space has a core-periphery structure, and countries tend to develop products related to their current exports, making development path-dependent [77, 81].Most Sub-Saharan African countries studied were concentrated in unsophisticated, poorly connected products, constraining structural transformation.Source: §2.2.1 · p. 13
  • The Method of Reflections uses country-product trade links to derive the Economic Complexity Index, which is correlated with income and predictive of future growth.Its iterative framework relates countries through shared products and accounts for product ubiquity and country diversity.Source: §2.2.1 · p. 14

2.2.2. Fitness index and economic dynamics

  • The Fitness-Complexity Method uses coupled nonlinear maps and outperforms the Method of Reflections in capturing trade-network structure and quantifying countries' growth potential.It links country fitness to product complexity, with product complexity bounded by the fitness of exporting countries.Source: §2.2.2 · p. 15, §2.2.3 · p. 16
  • Countries in the Fitness-GDPpc plane occupy laminar, more predictable dynamics or chaotic, less predictable dynamics, where regression-based prediction is unsuitable.A selective-analogue forecast of five-year GDP growth outperformed the IMF five-year GDPpc forecast by more than 25% in accuracy.Source: §2.2.2 · p. 16, Figure 7 · p. 16

2.2.3. Variant indices and development analysis

  • When γ ≳2, MFCM’s product-complexity ranking is perfectly correlated with FCM’s but volatile and very noise-sensitive; on reliable, high-quality data, MFCM can produce better country and product rankings.As γ →∞, MEM is a special case of MFCM; it reproduces the nested country-product network structure but is highly sensitive to data noise.Source: §2.2.3 · p. 17
  • FCM is unstable to small network perturbations, and adding a product exported by one country may reorganize the fitness landscape.The cited analysis attributes this to nonlinear iteration amplifying low-fitness countries' effects on product complexity.Source: §2.2.3 · p. 17
  • Applications connect product diversification to development paths, service-export complexity to growth prospects, and complex exports to lower income inequality.CIRD treats development as a one-dimensional process, with approximately −2 as a poverty-trap exit threshold [114]; wage inequality also follows a persistent pattern along industrialization.Source: §2.2.3 · pp. 17–19

2.3. Spatial demography and culture evolution

This section examines how geospatial and digital data and computational methods support population mapping, migration estimation, and the study of cultural evolution. It covers language and food-related practices, their geographic and shared cultural influences, and links to demographic and health outcomes.

2.3.1. World population distribution

  • Combining telecommunications, satellite, and census data, Douglass et al. found total out-call volume correlated about 0.68 with grid-level population, and building land cover and 10am outgoing calls were top predictors.Source: §2.3.1 · p. 21
  • A framework integrating census data, Baidu points of interest, and real-time Tencent user densities estimated building-level population and achieved high correlation with official census data.Source: §2.3.1 · p. 21

2.3.2. International migration

  • International migration flows were about 0.6% of world population per five-year period since 1995, with regional patterns including predominantly intra-African migration and long-distance flows toward higher-income countries.Source: §2.3.2 · p. 22
  • Digital records enabled migration estimates by age and gender from Yahoo! email, mobility mapping from Yahoo! and Skype, and professional-worker migration trends from LinkedIn.The Yahoo! age profiles were consistent with official data, female mobility grew faster, and the share of professional migrants to the US declined from 2000 to 2012 while Asia was a major destination.Source: §2.3.2 · p. 22
  • Twitter-based migration research identified spatially cohesive mobility regions, predicted turning points in OECD out-migration trends, and found a clustered, disassortative, modular, small-world migration network.Migration-network link weights followed a power law with a stable exponent of about 1.3.Source: §2.3.2 · p. 23

2.3.3. Culture evolution

  • Digitized books revealed cultural and linguistic trends, while historical mobility analysis identified two European cultural regimes and location-based social-network data produced survey-matching cultural maps.Chinese cultural-value words associated with adaptive individualism increased in frequency from 1970 to 2008.Source: §2.3.3 · pp. 23–24
  • Baby-name studies found long-running power-law frequency distributions and higher turnover for female names, while language studies examined linguistic change, scaling laws, and language-specific departures from them [182, 183, 188–201].For Chinese texts, character growth can progress from linear to logarithmic to saturation, and long texts can show a two-layer rank-frequency structure.Source: §2.3.3 · pp. 24–25
  • Vocabulary and spelling shifted from British to American English over two centuries, suggesting cultural influence gradually shifted from the UK to the US.The study examined geotagged tweets and Google Books datasets for books published in the US and UK.Source: §2.3.3 · p. 26
  • Western cuisines tend to share ingredients with compound flavors, while East Asian cuisines tend to avoid such pairings.A flavor network was used to capture ingredient combinations in recipes.Source: §2.3.3 · p. 26
  • Geographical proximity matters more than climate proximity for Chinese regional cuisine similarity, and an evolution model reproduces the real dataset’s similar tendency.The study analyzed online recipes.Source: §2.3.3 · p. 26
  • Ingredient preferences vary less by region than recipe preferences, weekday and weekend preferences differ, and tweeted food calories correlate 0.77 with state obesity rates.The studies used recipe-platform logs and dining tweets from 0.21 million US users; users in higher-educated areas tweeted about less-caloric food, and food names plus demographics predicted county obesity and diabetes statistics.Source: §2.3.3 · pp. 26–27
  • Across 31 European food cultures, shared internal states correlate positively with shared culinary practices, and neighbouring countries tend to have similar practices.Cultural relations were mined from descriptions and popularity of culinary practices across Wikipedia language communities.Source: §2.3.3 · p. 27

3.1. Economic activity and socioeconomic status

This section surveys how nighttime lights, very high resolution imagery, mobile-phone records, and social-media data can reveal economic activity and socioeconomic status at regional and urban scales. It also covers how social-media signals relate to regional socioeconomic deprivation and landscape values, including their validation against census data and links to socioeconomic, geographic, and accessibility factors.

3.1.1. Nighttime lights reflect economic activity

  • VIIRS nighttime lights correlate with prefecture GDP (R2 = 0.8935) and county GDP (R2 = 0.9243), yet at micro-level they proxy population density better than economic activity [221, 219].The micro-level relationship with economic activity is overestimated in large urban areas, underestimated in rural areas, and stronger for radiance than saturated light [219].Source: §3.1.1 · p. 28

3.1.2. Very high resolution imagery maps poverty

  • VHR slum maps face context-dependent slum conditions, but line-detection lacunarity reached 0.8333 accuracy at Λ ∈[1.10, 1.15], and satellite mapping in Accra achieved 94.3% against field maps.The Accra accuracy dropped when compared with two census-derived slum maps.Source: §3.1.2 · pp. 28–30

3.1.3. Mobile phones track socioeconomic levels

  • UK census-based IMD rank correlated with mobile-phone social diversity (r = 0.73), spatial diversity (r = 0.58), and composite diversity (r = 0.78), supporting regional socioeconomic inference [248].In a Latin American country, multivariate regression including mobile-phone usage variables predicted census-based socioeconomic level (R2 = 0.83).Source: §3.1.3 · p. 31

3.1.4. Social media reveals socioeconomic status

  • Twitter-derived socioeconomic profiles reflected optimism and business, networking, and leisure in affluent neighborhoods; urban structure shaped activity, with mid-income users potentially traveling least and affluent users more internationally.Source: §3.1.4 · p. 33
  • Registered Weibo users correlated with city GDP (r = 0.88; ρ = 0.90), while combined online-information and talent-mobility network features explained up to about 84% of regional GDP variance.Weibo activity also helped identify resource-dependent cities whose economies later suffered losses after market-price fluctuations.Source: §3.1.4 · p. 33
  • Cross-state long ties correlated with GDP (r = 0.921), patents (r = 0.788), and startups (r = 0.892), while Dutch prosperity tracked geographic rather than topological network diversity [272, 273].Source: §3.1.4 · p. 34
  • Using Foursquare and OpenStreetMap data, Venerandi et al. defined offering advantage to identify distinctive neighborhood elements and built fine-scale deprivation classifiers verified against census-based IMD.The method mined deprivation from urban elements in the physical environment in the UK.Source: §3.1.4 · p. 34
  • Panoramio, Flickr, and Instagram data showed similar landscape-value patterns, with a significant portion of variation explained by accessibility, population density, income, mountainous terrain, and proximity to water.This analysis was conducted by van Zanten et al..Source: §3.1.4 · p. 34

3.2. Industrial structure and development path

This section reviews methods for mapping regional industrial structure and understanding diversification, emphasizing economic relatedness, learning across industries and regions, and development strategies. It presents regional industrial change as path-dependent while also discussing ways economies may pursue less related activities.

3.2.1. Economic structure and relatedness

  • Weibo user numbers predicted GDP better than population and average GDP measures across 282 Chinese cities, and deviations from the fitted line helped estimate regional economic structure.The analysis associated cities below the fitted line with tertiary industries and those above it with secondary industry.Source: §3.2.1 · p. 35
  • China’s industry space links 70 industries and shows core-periphery and dumbbell structures; Beijing and Shanghai gained internet and financial services, while Hebei and Zhejiang gained manufacturing [281].A separate analysis found that coastal provinces tend to be industrially sophisticated and provincial complexity rankings remained relatively stable.Source: §3.2.1 · p. 35
  • Regions with related industries tend to diversify and grow more: Spanish provinces with varied related industries had higher economic growth, while related knowledge spillovers are geographically bounded [284, 285, 286].Relatedness also predicts industrial entry and plant survival in Swedish regions.Source: §3.2.1 · p. 36, §3.2.3 · p. 39

3.2.2. Collective learning in economic development

  • Neighboring-country specialization is associated with a 65% higher probability of exporting a new product after regressions discount product relatedness [297]; US states also show a higher probability of new-industry development when a neighboring state specializes in it.The US-state analysis also found that sharing a border raises export similarity by 0.43 standard deviations.Source: §3.2.2 · p. 37
  • The probability of a province developing a new industry within five years rises with related-industry density and neighboring-province density; their joint effects are significant but show diminishing returns [281].Figure 19 plots the marginal probabilities for each density and the joint probability across both densities.Source: §3.2.2 · p. 38, Figure 19 · p. 38

3.2.3. Development paths and strategies

  • Simulations identify strategies that balance core and peripheral industries in initial activation, and nearby and distant regions in creating spatial links, to maximize learning effects in diversification.The simulations use Brazilian industry and regional networks.Source: §3.2.3 · p. 40
  • Strategic diffusion targets core products during an early, narrow time window to minimize the time needed to develop all products; countries diversify near the modeled optimum but target core products later.Their timing nevertheless leaves room to reduce total development time.Source: §3.2.3 · pp. 40–41
  • Developing regions can jump farther into core industries by relying on internal innovation and extra-regional links, while unrelated product entry is associated with higher human capital and intermediate development [329].Countries entering more unrelated products experienced a significant increase in economic growth [329].Source: §3.2.3 · p. 41

3.3. Urban scalings and perception

This section reviews how urban scaling laws and computational methods use diverse data to examine relationships between city population and socioeconomic metrics, urban structure, economic activity, and environmental perception. It covers scaling patterns and explanations, functional-area mapping, housing markets, neighborhood change, and analyses of urban life and socioeconomic patterns.

3.3.1. Scaling laws for cities

  • Urban metrics fall into three scaling classes: β ≈1 for individual needs, β ≈1.2 for social currencies, and β ≈0.8 for infrastructure.These patterns were observed across urban systems in China, Germany, and the USA.Source: §3.3.1 · p. 42, Figure 21 · p. 42
  • Models explain scaling through social interaction and geography: output can scale as β = 4/3, while distance-dependent activity yields superlinear, sublinear, or linear scaling.In the geographical network account, activity increasing with distance produces superlinear scaling, decreasing with distance produces sublinear scaling, and sufficiently strong geographical constraints produce linear scaling.Source: §3.3.1 · p. 43
  • The scale-adjusted metropolitan indicator measures city performance independently of population size, while urban productivity shows systematic dependence on population.The indicator is based on deviations from general scaling laws and supports comparisons without population-size bias.Source: §3.3.1 · p. 45

3.3.2. Unfolding urban functional areas

  • Nighttime-light and satellite data map urban expansion, revealing distinct growth trajectories in China and India and improving estimates of urban land area.China’s cities expanded built-up infrastructure in height and extent with increased population density, whereas Indian cities primarily expanded outward at lower density.Source: §3.3.2 · p. 45
  • Mobile-phone, points-of-interest, and mobility data identify functional regions, and Yuan et al.’s DRoF method outperformed baselines using only POIs or human mobility.Mobile interaction networks revealed communities aligned with county and prefecture boundaries, while DRoF combines mobility patterns and POIs to infer regional functions.Source: §3.3.2 · p. 46
  • Urban villages discounted nearby house prices by 2.5%, and their removal increased prices by over 3.3%; restaurant activity captured 20-40% of appreciation near new subway stations.The estimates concern Beijing, China.Source: §3.3.2 · pp. 47–48

3.3.3. Perceiving urban environment

  • Crowdsourced image perceptions varied more in Boston and NYC than in Linz and Salzburg, helped explain NYC violent crime at zip-code resolution, and predicted perceptions correlated with official crime and wealth statistics.The image-based measures captured perceptions of safety, class, uniqueness, and wealth.Source: §3.3.3 · p. 48
  • R2 = 0.5676 was achieved by Streetscore for predicting safety Q-scores; it scored about 1 million street-view images to map perceptions across 21 US cities.The model used generic image features and ratings from the Place Pulse survey, with 7,000 participants, to predict perceived street safety.Source: §3.3.3 · pp. 48–49
  • Theft-rate predictors were on average 33% more accurate than humans, and greenery was the most positive visual cue associated with beauty, quiet, and happiness.Quercia et al. translated London street ratings into measures of those three qualities; Arietta et al. linked visual elements to city attributes including crime rates, population density, and tree presence.Source: §3.3.3 · p. 49
  • RSS-CNN outperformed SS-CNN on perceptual-attribute prediction, and models trained on one visual attribute could predict others with fair accuracy.Other image-based work quantified scenic beauty from over 200,000 Great Britain images and used satellite imagery to classify urban features and compare neighborhoods across cities.Source: §3.3.3 · p. 50
  • Street-view changes from 2007 to 2014 supported three urban-change theories: education and density predict improvement, proximity predicts spillovers, and better initial appearances predict larger gains.The analysis used over one million image cutouts from five US cities; related work found Yelp-recorded Starbucks entries indicative of house-price growth and urban form explanatory of up to 70% of IMD variance in six UK cities.Source: §3.3.3 · pp. 50–51

3.3.4. Urban computing for better lives

  • Studies used urban data to test links between physical environments and vitality, and to examine segregation, mobility, religious connections, and wealth-based neighborhood change.Research tested Jacobs’s four conditions for urban vitality in Seoul and six Italian cities; other studies found restricted cross-income interactions, highly segregative religious networks, distinct native and non-native check-in patterns, and persistent wealth segregation that could shift toward mixed-wealth distributions when some disallowed moves were allowed.Source: §3.3.4 · pp. 51–53
  • City-level vehicle-based estimates correlated with the American Community Survey at r = 0.82 for median household income, r = 0.87 for Asians, and r = 0.70 for graduate degrees.The method detected vehicles in 50 million street-view images from 200 US cities and also produced estimates at zip-code resolution, including r = 0.77 for the percentage of Asians in Seattle.Source: §3.3.4 · p. 53
  • Public facilities scaled with population density at α ≈ 2/3, while private facilities scaled at α ≈ 1; amenity networks also enabled neighborhood-scale recommendations for missing facilities.The proposed model’s predicted optimal positions for public and private facilities on real US maps agreed well with empirical data.Source: §3.3.4 · p. 54
  • Redirecting 5% of individual card transactions reduced neighborhood spatial inequality by more than 80% while improving total travel distance and spatial mixing.The redistribution rewired shopping trips to businesses in the same category but different neighborhoods while preserving human-mobility properties.Source: §3.3.4 · p. 54

4.1. Individual socioeconomic level

This section reviews how mobile and financial records, social-network structure, and human mobility patterns can be used to estimate individual socioeconomic level. The studies cover both behavioral correlates of status and predictive approaches using large-scale data.

4.1.1. Mobile phone and credit card usage

  • Using 38 mobile-phone features, a model classified socioeconomic levels for about 0.5 million citizens with accuracy up to 80%.The authors describe the method as a cost-effective complement to traditional socioeconomic estimation techniques.Source: §4.1.1 · p. 54
  • Outgoing purchase diversity correlated with socioeconomic status at r = 0.77 in a European country and r = 0.53 in a Latin American country.A separate Spanish transaction model predicted regional GDP per capita with accuracy 0.729 using seven principal components.Source: §4.1.1 · pp. 55–56

4.1.2. Social profile and network structure

  • Foursquare visits and Twitter messages supported social-class estimation, while Twitter features predicted income with accuracy 0.633.The Foursquare/Twitter classification models had average F1 values from 0.57 to 0.73 across classification segments.Source: §4.1.2 · p. 56
  • Network measures correlated with individual economic status: k-shell R² = 0.96 and CI (l = 2) R² = 0.93; a call-based Bayesian income classifier achieved AUC = 0.71 at τ = 0.4.The correlation for CI increased when individual ages were included.Source: §4.1.2 · pp. 57–58
  • Ego-network structural diversity positively correlated with income after controlling for education, occupation, age, and gender [475], while game-network position diversity correlated with economic output.In the gaming study, class economic outputs increased with cluster position diversity.Source: §4.1.2 · pp. 58–59

4.1.3. Human mobility pattern

  • Across Colombian socioeconomic groups, poorer residents covered larger urban areas sparsely using few, cheaper modes, whereas the richest covered smaller areas densely using costly modes.The same study found the poorest group had the smallest clustering coefficient and tended to avoid transport-mode overlap.Source: §4.1.3 · p. 59
  • Mobility measures predicted socioeconomic indicators: mobility variables outperformed consumption variables, and time-series models predicted total assets with training R² about 0.68.Across municipalities, mobility diversity showed the strongest correlations among mobility and social measures, including r = 0.49 with per capita income and r = −0.43 with deprivation [492, 494].Source: §4.1.3 · pp. 60–61
  • In Barcelona and Madrid, passengers’ income was negatively correlated with commuting distance and transit frequency, linking financial transactions and transport modes to status estimation.A separate transaction study found spatio-temporal traits improved demographic models by 30% to 49% in predicting future financial difficulties.Source: §4.1.3 · pp. 61–62

4.2. Employment and performance

This section reviews how digital traces and network data support unemployment estimation and the analysis of individual, group, and scientific performance. It also covers behavioral features used to predict student performance and examine its links with campus routines.

4.2.1. Search queries indicate unemployment

  • GI-augmented models had MSE 29% lower one step ahead and 40% lower three steps ahead than traditional models in US unemployment forecasting.The comparison came from an out-of-sample evaluation of more than 500 time-series forecasting models.Source: §4.2.1 · p. 62
  • Search queries predicted unemployment across countries and age groups, but their usefulness varied by population, language, and out-of-sample period [504, 512, 519, 522].GT improved nowcasting in three of four south-western European countries; one Korean youth model explained about 80% of unemployment-rate variation, while a VAR outperformed ARIMA for Italian youth unemployment.Source: §4.2.1 · pp. 62–64

4.2.2. Other sources relevant to unemployment

  • In Spain, Twitter-derived behavioral measures predicted regional unemployment, with R2 = 0.62 for people below 25 and R2 = 0.52 for those aged 25–44.Unemployment was associated with Twitter penetration, tweet timing, misspellings, and less diverse communication patterns.Source: §4.2.2 · p. 64
  • Employee network features were associated with turnover and promotion, and action-network features predicted both outcomes more strongly than social-network features.Turnover was negatively correlated with several centrality measures, while high in-degree was associated with promotion and PageRank and LeaderRank were positively correlated with promotion.Source: §4.2.2 · p. 65
  • Mobile-phone records tracked employment shocks and predicted individual and regional unemployment, including province-level rates one quarter ahead with r = 0.85.For laid-off users, total calls fell 51%, outgoing calls fell 54%, unique towers visited fell 17%, and radius of gyration fell 20%; individual employment prediction accuracy averaged 0.675 across professions and reached 0.735 for identifying unemployed users.Source: §4.2.2 · p. 66

4.2.3. Individual and group performance

  • Network structure and team communication were linked to performance, with cooperative call-center teams answering about 40% more phones than control teams.Studies also associated advice-network centrality and informal or multiplex network structure with individual performance, and identified energy, engagement, and exploration as team communication dynamics related to performance.Source: §4.2.3 · p. 67
  • Scientific performance research links team structure and collaboration to research outcomes and proposes measures and models for assessing individual impact.The section describes the h-index and discipline-normalized h_f index, a model of sustained high-impact publishing, and findings that team publications tend to receive more citations and collaborations overall improve paper quality.Source: §4.2.3 · p. 68
  • Smartphone sensing identified conversational interaction, class attendance, and study hours as behavioral factors correlated with students’ term and cumulative GPA.Source: §4.2.3 · p. 68
  • Behavioral features predict cumulative GPA with r = 0.81.Source: §4.2.3 · p. 69
  • Cao et al. defined orderliness from meal and shower timing records using actual entropy, finding positive GPA correlation and improved prediction accuracy alongside other behavioral indicators.Source: §4.2.3 · p. 69
  • Yao et al. used smart-card records from 6,597 students to derive orderliness, diligence, and sleep-pattern features for a multi-task academic-performance prediction framework.Source: §4.2.3 · p. 69

4.3. Demographics and personal variables

The section reviews how social-media, mobile-phone, and search data can support inference of demographic and personal attributes, including personality, reputation, emotion, mental-health signals, and suicidal risk. It describes computational and predictive approaches, findings, and challenges in risk assessment and ethics.

4.3.1. Demographic inference

  • Including first names improved Twitter gender-inference accuracy by about 20%, and another method identified gender without profile or text analysis.Mobile-phone call-detail records also supported demographic inference: WhoAmI outperformed benchmarks by about 10% in F1, and a later framework predicted gender and age with R2 = 0.47 for age.Source: §4.3.1 · pp. 69–70

4.3.2. Personality analysis

  • Twitter text and image models predicted dark-triad traits at r ≈0.25, while image content predicted neuroticism and conscientiousness at Pearson correlations 0.530 and 0.566.Profile-image features were also linked to personality traits, with Pearson correlation 0.189 for predicting conscientiousness.Source: §4.3.2 · p. 72

4.3.3. Online reputation evaluation

  • The GR method assigns reputation from users’ rating-group sizes, whereas PGR groups users by rating preferences before applying the GR framework.These approaches reflect different assumptions: GR groups similar ratings, while PGR accounts for users’ tendencies to give higher or lower ratings.Source: §4.3.3 · pp. 74–75

4.3.4. Emotion and health analysis

  • SentiStrength predicted positive and negative emotions in MySpace comments with accuracies 0.606 and 0.728, respectively.An analysis of 2,600 human-classified comments found that two thirds expressed positive emotions.Source: §4.3.4 · p. 76
  • Emotion classification and intensity estimation use tweet features including n-grams, sentiment words, POS information, embeddings, and lexicons.With about 2 million training tweets, Wang et al. report highest emotion-identification accuracy of 0.6557; Mohammad and Bravomarquez report r = 0.66 for word-embedding and lexicon features, while Madisetty and Desarkar report Spearman rank correlation 0.725 for an ensemble.Source: §4.3.4 · p. 76
  • Social-media posts reveal mood patterns and associations with well-being and local conditions, including lower Weibo happiness at higher PM2.5 concentrations.Twitter studies report temporal and geographic mood patterns and event-related mood shifts; Facebook emotional expression correlates with anxiety, depression, and stress. The Weibo happiness index was constructed from 210 million geotagged tweets across 144 Chinese cities in 2014.Source: §4.3.4 · p. 76, Figure 41 · p. 77
  • Social-media language and self-disclosures have been used to study mental-health conditions, with Twitter classifiers reporting precision up to 0.85 for anxiety and 0.75 for eating disorders.A Reddit-forum algorithm detected levels of self-disclosure with accuracy 0.78 using content features.Source: §4.3.4 · p. 77
  • Facebook models of postpartum depression found that adding early postnatal information improved prediction, and prenatal plus early-postnatal features explained about 48% of variance.The study used data shared by 165 new mothers and stepwise logistic regression.Source: §4.3.4 · p. 77
  • Depression detection studies use social-media activity, user, and image features, including changes before onset and a two-month optimal Twitter observation window.Reported results include average accuracy about 0.70 for a Twitter SVM, accuracy about 0.80 for a Weibo model, accuracy 0.69 using Twitter activity histories, and F1 measure 0.647 for identifying depressed individuals from Instagram photos.Source: §4.3.4 · p. 77
  • A 1200-tree random forest correctly classified about 88.2% of PTSD predictions, while a hidden Markov model inferred depression onset months before clinical diagnosis.The study used depression histories and Twitter behaviors from 204 individuals; Figure 42 shows ROC curves and top predictors for depression and PTSD.Source: §4.3.4 · p. 78, Figure 42 · p. 78
  • Social-media suicide surveillance uses posts and behavioral features, while search-query estimates vary across settings and may require specific rather than broad terms.Suicide-related tweets correlated with actual age-adjusted suicide data at Spearman’s rank correlation rho = 0.53, and a Twitter ensemble achieved F1 measure 0.69 for suicidal ideation. Google Trends results differed across studies: in England, larger volumes of “suicide” searches indicated more suicides, whereas validation across four countries found weak associations and no association for “suicide” searches in the US.Source: §4.3.4 · p. 78
  • Reviews identify continuing challenges in accurate suicide-risk assessment, while mental-health analyses based on personal sensing and mobile interventions also raise ethical concerns.Mohr et al. proposed translating raw sensor data into behavioral markers related to mental health; the review also notes technological advances alongside ethical challenges.Source: §4.3.4 · pp. 78–79

5.1. Public health and epidemic surveillance

The review surveys digital signals for epidemic surveillance, from search queries and online posts to mobile-phone-derived mobility. These sources support timely monitoring but have reliability and representativeness constraints.

5.1.1. Search queries for epidemic surveillance

  • Search-query studies support timely ILI estimation across platforms, while GFT reliability can erode as search behavior and model conditions change.Source: §5.1.1 · p. 79

5.1.2. Online posts for disease surveillance

  • Twitter surveillance ranges from term-frequency regression to classifiers separating infection reports from concern; combining online and traditional predictors can yield earlier estimates at comparable accuracy.Source: §5.1.2 · pp. 81–82
  • Wikipedia access-log models combine article requests with prior ILI and can forecast official levels weeks ahead; a figure illustrates regression against current ILI.Source: §5.1.2 · p. 82, Figure 44 · p. 83

5.1.3. Mobile phone records for epidemic prediction

  • Mobile-phone mobility maps importation risk: Kenyan analyses identify Nairobi as a major destination for humans and parasites, while rubella risk varies seasonally and spatially.Source: §5.1.3 · pp. 83–84, Figure 45 · p. 84
  • Mobility estimates can diverge in magnitude and fit by resolution; mobile-phone data also face ownership bias, availability limits, and missing demographic or social identifiers.Source: §5.1.3 · pp. 84–85

5.2. Emergency and disaster monitoring

This section surveys computational approaches to emergency and disaster monitoring using remote sensing, mobile-phone records, and social media to support timely situational awareness and response. It covers damage assessment, emergency detection, and population monitoring, alongside social-media applications to flood mapping, collective attention and damage, human mobility and emotions, and monitoring droughts, wildfires, and snowstorms.

5.2.1. Remote sensing for disaster assessment

  • Satellite remote sensing supports timely, wide-area disaster assessment, with one flood-mapping procedure identifying 96.7% of flooded areas against an official reference map.For earthquake assessment, high-resolution imagery can reveal changes to ground surfaces and buildings before and after an event, enabling analysis at settlement and building levels.Source: §5.2.1 · pp. 86, 88

5.2.2. Mobile phones for emergency management

  • Mobile-phone records reveal emergency signals in calling and mobility, supporting real-time event detection and monitoring of evacuation and population displacement.Calling activity can spike rapidly during emergencies, while mobility records can track displacement; after the Haiti earthquake, estimated movements from Port-au-Prince were highly predictable from normal-time mobility patterns.Source: §5.2.2 · pp. 88–90

5.2.3. Social media for situational awareness

  • 93% of flood locations in tweeted-about regions were detected in Jakarta using filtered, geo-statistical analysis of multiple observations.Social media can provide situational awareness, but crisis-scale message volumes call for automated extraction; the Tweedr pipeline classifies, clusters, and extracts damage information from disaster tweets and identified 12 US crisis events.Source: §5.2.3 · pp. 91–92, Figure 48 · p. 92
  • A Bayesian model combining social media and remote sensing predicted spatial flood inundation with AUC over 0.93, while tweets and crowdsourced photos enabled high-resolution flood monitoring.The photo and tweet workflow extracted locations and flood depth using NLP and classified flooding photos with CNNs.Source: §5.2.3 · p. 93
  • Kendall and Spearman correlations between per-capita Twitter activity and Hurricane Sandy damage reached 0.39 and 0.55, respectively, across New Jersey ZCTAs.Twitter sentiment was a weak predictor of damage.Source: §5.2.3 · p. 93, Figure 49 · p. 93
  • Flickr photo activity tracked atmospheric pressure during Sandy (Kendall’s τ = −0.37), while 9.7 million geotagged tweets linked activity to landfall wind power and proximity to the hurricane path.Sadness was the most pronounced emotion in NYC-area tweets, while surprise and fear peaked on landfall day.Source: §5.2.3 · p. 93
  • Sandy significantly perturbed human mobility, but mobility parameters under disruption remained highly correlated with steady-state values [863]; a crisis-mapping platform mapped impacts at building and street level with F1 0.53.Long-distance movements fell on the first day; tweet displacement distributions had power-law exponents of 1.73 that day and about 1.19 over the following 11 days.Source: §5.2.3 · pp. 93–94
  • Social media supported drought-risk information sharing, wildfire monitoring, and snowstorm preparedness and response; wildfire activity correlated with reported fires at r = 0.529 for Twitter and r = 0.716 for Instagram.Wildfire-related Twitter hotspots were likely to coincide with wildfire locations.Source: §5.2.3 · p. 94

6. Discussions

The review presents computational socioeconomics as a way to draw insights from large-scale data across economic development, inequality, socioeconomic status, public health, and disasters. The field remains at an early stage, with data quality, applicability, causal explanation, and practical implementation still demanding attention.

6. Discussions

  • Across economic networks, inequality, socioeconomic status, public health, and disasters, the review illustrates insights drawn from large-scale data and advanced analysis tools.The review argues that such insights are often difficult to obtain through traditional methods.Source: §6 · p. 94
  • Social-media data may include bots or deliberately fabricated content, which can distort sentiment or spread rumors and produce conclusions far from reality.The authors recommend pretreatment and consideration of risks from noisy or unreal data.Source: §6 · p. 94
  • Large datasets may miss less-connected groups, findings may not transfer across regions, and group-level patterns cannot be assumed to hold for individuals.The authors also caution that patterns observed at the group level may differ from those at the individual level, and vice versa.Source: §6 · pp. 94–95
  • Future work should develop useful new socioeconomic indices, rather than only estimating familiar indicators such as GDP with novel data and tools.The review notes that even 0.99 accuracy would not outperform the original GDP statistical method, whose accuracy is 1.Source: §6 · p. 95
  • Large-scale data-driven methods could reassess traditional socioeconomic theories, many of which lack comprehensive validation across diverse societies and contexts.The authors suggest that such evaluation may also help integrate computational and traditional socioeconomics.Source: §6 · p. 95
  • Because most findings are correlational, the review calls for robust causal relationships and theoretical models that withstand strict evaluation.It describes the theoretical value of findings based only on correlations as limited.Source: §6 · p. 95
  • The review does not advocate abandoning controlled experiments; large-scale Internet experiments can test theories, while data analysis can help guide their design.It cites a 61-million-person Facebook experiment as an example.Source: §6 · p. 95
  • Practical impact remains limited, so the authors propose multisource real-time socioeconomic monitoring and data-driven policies tested in regions with government support.They identify implementation support from the relevant government heads as a critical issue.Source: §6 · pp. 95–96

Related papers

Loading 1905.06166v1…