Academic literature on the topic 'Crowd risk prediction'

Can machines outperform crowds in financial lending decisions? Using data from a crowd-lending platform, we show that, compared with portfolios created by crowds, a reasonably sophisticated machine can construct financial portfolios that provide better returns while controlling for risk. Further, we find that the machine-created portfolios benefit not only the lenders, but also the borrowers. Borrowers receive loans at a much lower interest rate as the machine can weed out the riskiest loans better than the crowds. We also find suggestive evidence of algorithmic bias in machine decisions. We find that, compared with women, men are more likely to receive loans by machine. We propose a general and effective “debiasing” method that can be applied to any prediction-focused machine learning (ML) applications. We show that the debiased ML algorithm, which suffers from lower prediction accuracy, still improves the crowd’s investment decisions in our context. Our results indicate that ML can help crowd-lending platforms better fulfill the promise of providing access to financial resources to otherwise underserved individuals and ensure fairness in the allocation of these resources.

Background. Acute kidney injury (AKI) is an important comorbidity of ST-Segment Elevation Myocardial Infarction (STEMI) and worsens the prognosis. The purpose of this study was to investigate the relationship between clinical data, test results, surgical findings, and AKI in STEMI patients and to develop a simple, practical model for predicting the risk of AKI. Method. Prognostic prediction research with clinical risk score development was conducted. The data used for model development was derived from the database of the Henan Province Cardiovascular Disease Clinical Data and Sample Resource Bank Engineering Research Center. The data used for external validation was derived from the China Chest Pain Center database. The study endpoint was defined as the occurrence of AKI. Logistic regression analysis was used to identify independent predictors of AKI. Logistic coefficients of each predictor were used for score weighting and transformation. The predictive performance of the newly derived risk scores was validated, respectively, by receiver operating characteristic (ROC) regression in the development population and the external validation population. Result. A total of 364 patients, 57 (15.6%) with AKI and 307 (84.4%) without AKI, were included for score derivation. The validation crowd includes 88 STEMI patients in different institutions. A total of 11 potential predictors were explored under the multivariable logistic regression model. The new risk score was based on five final predictors which were age > 72 years , ejection fraction of no more than 40%, baseline serum creatinine > 102.7 mmol / L , red blood cell distribution width > 13.15 , and culprit lesion located in the middistal segment. With only five predictor variables, the score predicted the risk of AKI with the effective discriminative ability (area under the receiver operating characteristic curve (AuROC): 0.721, 95% confidence interval (CI): 0.652-0.790). In the external validation, the newly developed score confirmed a similar discrimination as the crowd used for derivation (AuROC: 0.731, 95% CI: 0.624-0.838). Conclusion. The newly developed score was proved to have good predictive performance and could be practically applied for risk stratification of AKI in STEMI patients.

With the development of COVID-19, the epidemic prevention requirements of city subway system have become stricter. This study studies the transmission path of epidemic disease in city subway system. Using FLUENT software and AnyLogic software, the simulation models of subway platform ventilation structure and crowd behavior mode in subway system are constructed, respectively, and SEIR (vulnerable exposed affected recovered) is used as the general infection model of epidemic disease. According to the actual situation, the parameters such as shoulder width, flow, and moving speed of crowd are determined, and the simulation analysis of epidemic disease transmission in subway system is carried out. The analysis results show that the transmission speed of the disease in the subway will increase with the enhancement of the transmission capacity of the disease and the increase of the contact rate. When the disease transmission capacity is 0.14, the number of latent persons reaches the peak at 14.115 time units, which is 1374, and the number of patients reaches the peak at 28.541 time units, which is 1925. According to the simulation results, the simulation analysis results show that with the enhancement of disease transmission ability and the increase of exposure rate, the maximum number of latent and sick people in the subway environment will increase. The corresponding suggestions on risk management and control of infectious disease transmission in subway are put forward. The research results can provide a useful reference for the epidemic prevention management of urban subway transportation system in China.

In today's information age, the necessary means exist for clinical risk prediction to capitalize on a multitude of data sources, increasing the potential for greater accuracy and improved patient care. Towards this objective, the Prostate Cancer DREAM Challenge posted comprehensive information from three clinical trials recording survival for patients with metastatic castration-resistant prostate cancer treated with first-line docetaxel. A subset of an independent clinical trial was used for interim evaluation of model submissions, providing critical feedback to participating teams for tailoring their models to the desired target. Final submitted models were evaluated and ranked on the independent clinical trial. Our team, called "A Bavarian Dream", utilized many of the common statistical methods for data dimension reduction and summarization during the trial. Three general modeling principles emerged that were deemed helpful for building accurate risk prediction tools and ending up among the winning teams of both sub-challenges. These principles included: first, good data, encompassing the collection of important variables and imputation of missing data; second, wisdom of the crowd, extending beyond the usual model ensemble notion to the inclusion of experts on specific risk ranges; and third, recalibration, entailing transfer learning to the target source. In this study, we illustrate the application and impact of these principles applied to data from the Prostate Cancer DREAM Challenge.

Human movements in complex traffic environments have been successfully simulated by various models. It is crucial to improve crowd safety and urban resilience. However, few studies focus on reproducing human behavior and predicting escape reaction time in the initial judgement stage in complex traffic environments. In this paper, a pedestrian pre-evacuation decision-making model considering pedestrian heterogeneity is proposed for complex environments. Firstly, the model takes different obvious factors into account, including cognition, information, experience, habits, stress, and decision-making ability. Then, according to the preference of the escapees, the personnel decision-making in each stage is divided into two types: stay and escape. Finally, multiple influencing factors are selected to construct the regression equation for prediction of the escape opportunity. The results show that: (1) Choices of escape opportunity are divided into several stages, which are affected by the pedestrian individual risk tolerance, risk categories strength, distance from danger, and reaction of the neighborhood crowd. (2) There are many important factors indicating the pedestrian individual risk tolerance, in which Gen, Group, Time and Mode are a positive correlation, while Age and Zone are a negative correlation. (3) The analysis of the natural response rate of different evacuation strategies shows that 19.81% of people evacuate immediately. The research in this paper can better protect public safety and promote the normal activities of the population.

Purpose Docetaxel has a demonstrated survival benefit for patients with metastatic castration-resistant prostate cancer (mCRPC); however, 10% to 20% of patients discontinue docetaxel prematurely because of toxicity-induced adverse events, and the management of risk factors for toxicity remains a challenge. Patients and Methods The comparator arms of four phase III clinical trials in first-line mCRPC were collected, annotated, and compiled, with a total of 2,070 patients. Early discontinuation was defined as treatment stoppage within 3 months as a result of adverse treatment effects; 10% of patients discontinued treatment. We designed an open-data, crowd-sourced DREAM Challenge for developing models with which to predict early discontinuation of docetaxel treatment. Clinical features for all four trials and outcomes for three of the four trials were made publicly available, with the outcomes of the fourth trial held back for unbiased model evaluation. Challenge participants from around the world trained models and submitted their predictions. Area under the precision-recall curve was the primary metric used for performance assessment. Results In total, 34 separate teams submitted predictions. Seven models with statistically similar area under precision-recall curves (Bayes factor ≤ 3) outperformed all other models. A postchallenge analysis of risk prediction using these seven models revealed three patient subgroups: high risk, low risk, or discordant risk. Early discontinuation events were two times higher in the high-risk subgroup compared with the low-risk subgroup. Simulation studies demonstrated that use of patient discontinuation prediction models could reduce patient enrollment in clinical trials without the loss of statistical power. Conclusion This work represents a successful collaboration between 34 international teams that leveraged open clinical trial data. Our results demonstrate that routinely collected clinical features can be used to identify patients with mCRPC who are likely to discontinue treatment because of adverse events and establishes a robust benchmark with implications for clinical trial design.

Metastatic castrate resistant prostate cancer (mCRPC) is the major cause of death in prostate cancer patients. Even though some options for treatment of mCRPC have been developed, the most effective therapies remain unclear. Thus finding key patient clinical variables related with mCRPC is an important issue for understanding the disease progression mechanism of mCRPC and clinical decision making for these patients. The Prostate Cancer DREAM Challenge is a crowd-based competition to tackle this essential challenge using new large clinical datasets. This paper proposes an effective procedure for predicting global risks and survival times of these patients, aimed at sub-challenge 1a and 1b of the Prostate Cancer DREAM challenge. The procedure implements a two-step feature selection procedure, which first implements sparse feature selection for numerical clinical variables and statistical hypothesis testing of differences between survival curves caused by categorical clinical variables, and then implements a forward feature selection to narrow the list of informative features. Using Cox’s proportional hazards model with these selected features, this method predicted global risk and survival time of patients using a linear model whose input is a median time computed from the hazard model. The challenge results demonstrated that the proposed procedure outperforms the state of the art model by correctly selecting more informative features on both the global risk prediction and the survival time prediction.

College students often affiliate with similar peers, forming identity-based peer crowds. Research has shown that affiliations with certain peer crowds is associated with risky behaviors, thus derailing college success. This study examined whether college peer crowd affiliations predicted risky and prosocial behaviors. Participants were 527 students at a public university in the Midwest (aged 18 - 26). Hierarchical multiple regression analyses showed that Counterculture and Athletic/Social affiliations positively predicted risky behaviors. Arts/Ethnic and Scholastic affiliations positively predicted prosocial behaviors and negatively predicted risky behaviors. In addition, hierarchical multiple regression analyses showed that gender moderated the relation between peer crowd affiliation and prosociality. The results highlight the importance of college peer crowds and their implications for academic success. The discussion focuses on ways to promote positive behavior among college peer crowds using research.

Business Administration/Accounting
D.B.A.
Regulation Crowdfunding, enacted in May 2016, is intended to facilitate capital formation in startups and small businesses funded primarily by small investors (Securities and Exchange Commission (SEC), 2016b). This dissertation investigates (1) the risk of fraud in equity crowdfunding offerings and (2) whether investors respond to fraud signals by selecting (rejecting) offers with low (high) fraud risk. Because equity crowdfunding is quite new, no frauds have yet been identified. Therefore, I employ a predictive analytics tool, Benford’s Law, to assess the fraud risk of the offering. I select observable indicators to represent the Fraud Triangle dimensions—incentives, opportunities and rationalization—and test if they predict fraud risk. I also compare offer funding outcomes to my fraud risk assessments to identify if investors’ selections consider fraud risk appropriately. The relaxed auditor assurance and disclosure requirements attracts both honest and dishonest founders, but I find that the risk of fraud is higher in equity crowdfunding offers than in public offerings as reported by other studies. I find that there are several individual fraud indicators and models that explain fraud risk, but these do not predict whether the offer is funded or not (funding outcomes) or the amount that is raised if funded. This dissertation is the first to apply Benford’s Law to equity crowdfunding offers and map fraud attributes to fraud risk and funding outcomes. My dissertation can inform investors, issuers, regulators, intermediaries and practitioners of the high risk of fraud in equity crowdfunding offerings and of several noteworthy fraud indicators.
Temple University--Theses

Understanding the principles and applications of crowd dynamics in mass gatherings is very important, specifically with respect to crowd risk analysis and crowd safety. Historical trends from India and other countries suggest that the crowd crushes in mass gatherings, especially in religious events, frequently occur, highlighting the importance of studying crowd behaviour more scientifically. This is required to support appropriate and timely crowd management principles in planning crowd control measures and providing early warning systems at mass gatherings. Hitherto, the researchers have studied the previous incidents of crowd crushes from the viewpoint of high density and the resulting physical forces and poor geometric facilities, but the factors such as psychological triggers and weather are overlooked. Further, although the average number of victims per panic event seems to decrease, their total number increases with the frequency of mass religious gatherings. Unless proper measures are in place, this trend will continue. Therefore, a comprehensive risk assessment is required to assess the potentially risky situations associated with an event that can lead to crowd crushes. To manage large crowds, an understanding of crowd dynamics is required to reasonably predict the level of risk and implement appropriate crowd management measures. However, there is a lack of empirical studies with real-world data on crowd behaviour and dynamics. Therefore, deriving motivation from the given background, the objectives of this research are: (1) to conduct a detailed empirical data collection in a mass religious gathering in an uncontrolled setup, (2) to understand the fundamental relationships between speed, flow, and density across different sections of case study, (3) to analyse the potentially risky situations observed in the site, and (4) to develop a comprehensive crowd risk model concerning crowd movement in mass religious gatherings and arrive at a Crowd Risk Index (CRI) which can give a range of values on scale defining the possibilities of crowd risks in a given area of mass religious gathering. The case study considered was Kumbh Mela 2016, held in Ujjain, India, between 22 April and 21 May. It attracted an estimated population of 75 million with an interesting mix of domestic and international pilgrims, spiritual leaders, and holy men, who journeyed to Ujjain from short duration (one day) to long-term stay (throughout the event). The key attractions of Kumbh were (1) taking a dip in the river Kshipra and (2) visiting temples. Data was collected throughout the event, covering the important days on which the crowd was expected to be more. Data in video form was recorded using Go-Pro, head-mount cameras, mobile phones and CCTV cameras. Additionally, data was also collected using GPS trackers and survey forms. Further, quantitative data was collected through visual observations. The Crowd Risk Index was developed from three pillars of indices: Crowd Dynamic Index (CDI), Crowd Anxiety Index (CAI), and Temperature-Humidity Index (THI). CDI include (i) macroscopic fundamental flow diagrams of a spiritually motivated crowd (ii) characteristics of stop and go waves in one-dimensional interrupted pedestrian flow through narrow channels (iii) understanding social group behaviour in the crowd and the effect of the presence of groups on the crowd movement, and (iv) understanding serpentine group behaviour and its impact on crowd dynamics. Using the above-mentioned study observations, the CDI was developed for ghat and temple locations as they were the two key attractions of Kumbh Mela. All the variables were used both for ghat and temple model. About 53 expert opinions were gathered separately for the temple and ghat videos. The experts rated the risk levels from the video clippings as low, medium, or high. Low was taken as class 1, medium as class 2, and high as class 3, which was given as an input to the CDI. The dataset was imbalanced, and so the SMOTE-Tomek Link method was used to balance out the dataset. Cross Validation technique using the Random Forest algorithm was used to predict the level of risk for CDI. CAI included the patience and aggression scores obtained from the study conducted on understanding the crowd’s emotions. A Structural Equation Modelling (SEM) was performed, and hypotheses testing were done to verify the relationship between the first order (cue-dependence (CD), tolerance (TO) and goal-oriented (GO); norm violation (NV), obstruction to movement (DO) and social display of power (SP)) and second-order factors (patience and aggression). All the first-order factors under patience and aggression were found to have a direct and significant impact on the second-order factors, i.e., patience and aggression, respectively. The patience and aggressions scores were obtained from the path loadings. Moreover, the effect of high temperature can have an indirect impact on the CRI through increasing aggression. This was also included in the index. The dataset here was also imbalanced, and so the SMOTE-Tomek Link method was used to balance out the dataset. The same Cross Validation technique using the Random Forest algorithm was used to predict the level of risk for CAI. A value between 0 and 1is class 1 (low), a value between 1 and 2 is class 2 (medium), and a value between 2 and 3 is class 3 (high). THI from literature was used to gauge the effect of temperature on the crowd risk. Kumbh Mela 2016 was held during peak summer under the scorching heat. The average temperature across the event duration was above 91-degree Fahrenheit, which implies that the event happened under severe stress conditions. This indicates the importance of including temperature effects into the model, especially for events that happen under high-temperature conditions. The comfort zone values were considered as class 1 (low), mild and severe stress conditions are combined as class 2 (medium), and severe stress conditions as class 3 (high). The CAI, CDI, and THI together form the CRI. The relative importance of these indices was also gathered from the same 53 experts. The weights were then calculated using the AHP process. Then the final CRI prediction equation was formulated. A CRI value between 0 and 1 indicates low risk, a value between 1 and 2 indicates medium risk, and a value between 2 and 3 indicates high risk. This can help in predicting the level of risk in a given area for every one-minute interval. Therefore, the CRI developed includes factors such as crowd anxiety and temperature, other than the crowd dynamics and behaviours, as it is important to include a comprehensive set of factors for a better prediction. With an overarching understanding of the factors leading to critical crowd conditions, the CRI developed in this work can help reasonably predict the level of risk and implement appropriate crowd management measures. However, the approach used in the study has its own set of limitations. There are other important factors that could endanger crowd safety, including bottleneck movement and crowd turbulence, among others, which are not considered. Studying and incorporating these into the CRI can result in a more accurate model. Adding health-related aspects and studying other psychological aspects supplemented with video data can also improve the model's precision. In addition, a comparison of different machine learning techniques to assess their performance could be a follow-up to this research. Despite these limitations, the study proposes a novel methodology for predicting crowd risk in mass religious gatherings. This is a one-of-a-kind study in crowd disaster and crowd safety that has never been attempted before in the literature.

Presented reasearch aimed to develop and analyse the suitability of the CART models for prediction of the extent and probability of occurrence of damage to outer soil layers caused by timber harvesting performed under varied conditions. Having employed these models, the author identified certain methods of logging works and conditions, under which they should be performed to minimise the risk of damaging forest soils. The analyses presented in this work covered the condition of soils upon completion of logging works, which was investigated in 48 stands located in central and south-eastern Poland. In the stands selected for these studies a few felling treatments were carried out, including early thinning, late thinning and final felling. Logging works were performed with use of the most popular technologies in Poland. Trees were cut down with chainsaws and timber was extracted by means of various skidding methods: with horses, semi-suspended skidding with the use of cable yarding systems, farm tractors equipped with cable winches or tractors of a skidder type, and forwarding employing farm tractors with trailers loaded mechanically by cranes or manually. The analyses also included mechanised forest operation with the use of a harvester and a forwarder. The information about the extent of damage to soil, in a form of wheel-ruts and furrows, gathered in the course of soil condition inventory served for construction of regression tree models using the CART method (Classification and Regression Trees), based on which the area, depth and the volume of soil damage under analysis, wheel-ruts and furrows, were determined, and the total degree of all soil disturbances was assessed. The CART classification trees were used for modelling the probability of occurrence of wheel-ruts and furrows, or any other type of soil damage. Qualitative independent variables assumed by the author for developing the models included several characteristics describing the conditions under which the logging works were performed, mensuration data of the stands and the treatments conducted there. These characteristics covered in particular: the season of the year when logging works were performed, the system of timber harvesting employed, the manner of timber skidding, the means engaged in the process of timber harvesting and skidding, habitat type, crown closure, and cutting category. Moreover, the author took into consideration an impact of the quantitative independent variables on the extent and probability of occurrence of soil disturbance. These variables included the following: the measuring row number specifying a distance between the particular soil damage and communication tracks, the age of a stand, the soil moisture content, the intensity of a particular cutting treatment expressed by units of harvested timber volume per one hectare of the stand, and the mean angle of terrain inclination. The CART models developed in these studies not only allowed the author to identify the conditions, under which the soil damage of a given degree is most likely to emerge, or determine the probability of its occurrence, but also, thanks to a graphical presentation of the nature and strength of relationships between the variables employed in the model construction, they facilitated a recognition of rules and relationships between these variables and the area, depth, volume and probability of occurrence of forest soil damage of a particular type. Moreover, the CART trees served for developing the so-called decision-making rules, which are especially useful in organising logging works. These rules allow the organisers of timber harvest to plan the management-related actions and operations with the use of available technical means and under conditions enabling their execution in such manner as to minimise the harm to forest soils. Furthermore, employing the CART trees for modelling soil disturbance made it possible to evaluate particular independent variables in terms of their impact on the values of dependent variables describing the recorded disturbance to outer soil layers. Thanks to this the author was able to identify, amongst the variables used in modelling the properties of soil damage, these particular ones that had the greatest impact on values of these properties, and determine the strength of this impact. Detailed results depended on the form of soil disturbance and the particular characteristics subject to analysis, however the variables with the strongest influence on the extent and probability of occurrence of soil damage, under the conditions encountered in the investigated stands, enclosed the following: the season of the year when logging works were performed, the volume-based cutting intensity of the felling treatments conducted, technical means used for completion of logging works, the soil moisture content during timber harvest, the manner of timber skidding, dragged, semi-suspended or forwarding, and finally a distance between the soil damage and transportation ducts. The CART models proved to be very useful in designing timber harvesting technologies that could minimise the risk of forest soil damage in terms of both, the extent of factual disturbance and the probability of its occurrence. Another valuable advantage of this kind of modelling is an opportunity to evaluate an impact of particular variables on the extent and probability of occurrence of damage to outer soil layers. This allows the investigator to identify, amongst all of the variables describing timber harvesting processes, those crucial ones, from which any optimisation process should start, in order to minimise the negative impact of forest management practices on soil condition.

The annual task of forecasting forest fire danger is becoming increasingly relevant, especially in the context of global warming. The forecast of surface fires is most important, as more than 80% of all vegetation fires are surface fires. Practically all crown fires develop from surface fires. This chapter discusses the deterministic-probabilistic method for predicting the number of forest fires in a controlled forest area. This methodology is based on the assumption that the number of registered and projected forest fires is related to the probability of their occurrence. The influence of forest fire retrospective data on the predicted number of forest fires for some sites of the Timiryazevskiy forestry of the Tomsk region was studied. This chapter presents the results of a comparative analysis of forecast data and statistics.

This chapter covers modelling of business-credit risk, whether retail or wholesale. (1) Risk 101—i) data sources—variations by firm or loan size (financial statements, traded securities prices, environmental assessments); ii) assessment tools—rating agency grades, business-report scores, public and private firm, hazard, portfolio, and exposure models; iii) rating grades—internal and external (Moody’s, Standard and Poor (S&P), Fitch; S&P provided further insights); iv) small and medium enterprises (SME) lending—including reviewing principals in the personal capacities. (2) Financial-ratio scoring—i) pioneers—including Altman’s Z score and Moody’s commercially successful RiskCalc; ii) predictive ratios—that have appeared; iii) agency usage—for the development of public- and private-firm models; iv) Moody’s RiskCalc—basics and results when first launched; v) non-financial factors—those typically considered, and how objectivity can be improved. (3) Forward-looking data—most provided by human judgment, even the ‘wisdom of the crowd’ inherent in market prices. Rating transitions and functional versus reduced-form models are also used.

The meteoric rise of smart devices in dominating worldwide consumer electronics market complemented with data-hungry mobile applications and widely accessible heterogeneous networks e.g. 3G, 4G LTE and Wi-Fi, have elevated Mobile Internet from a ‘nice-to-have' to a mandatory feature on every mobile computing device. This has spurred serious data traffic congestion on mobile networks as a consequence. The nature of mobile network traffic today is more like little Data Tsunami, unpredictable in terms of time and location while pounding the access networks with waves of data streams. This chapter explains how Big Data analytics can be applied to understand the Device-Network-Application (DNA) dimensions in annotating mobile connectivity routine and how Simplify, a seamless network discovery solution developed at Nextwave Technology, can be extended to leverage crowd intelligence in predicting and collaboratively shaping mobile data traffic towards achieving real-time network congestion control. The chapter also presents the Big Data architecture hosted on Google Cloud Platform powering the backbone behind Simplify in realizing its intelligent traffic steering solution.

The meteoric rise of smart devices in dominating worldwide consumer electronics market complemented with data-hungry mobile applications and widely accessible heterogeneous networks e.g. 3G, 4G LTE and Wi-Fi, have elevated Mobile Internet from a ‘nice-to-have' to a mandatory feature on every mobile computing device. This has spurred serious data traffic congestion on mobile networks as a consequence. The nature of mobile network traffic today is more like little Data Tsunami, unpredictable in terms of time and location while pounding the access networks with waves of data streams. This chapter explains how Big Data analytics can be applied to understand the Device-Network-Application (DNA) dimensions in annotating mobile connectivity routine and how Simplify, a seamless network discovery solution developed at Nextwave Technology, can be extended to leverage crowd intelligence in predicting and collaboratively shaping mobile data traffic towards achieving real-time network congestion control. The chapter also presents the Big Data architecture hosted on Google Cloud Platform powering the backbone behind Simplify in realizing its intelligent traffic steering solution.

Influenza is an important public health concern. Influenza leads to the death or hospitalization of thousands of people around the globe every year. However, the flu-season varies every year viz. when it starts, when it peaks, and the severity of the outbreak. Knowing the trajectory of the epidemic outbreak is important for taking appropriate mitigation strategies. Starting with the 2013–2014 flu season, the Influenza Division of the Centers for Disease Control and Prevention (CDC) has held a “Predict the Influenza Season Challenge” to encourage the scientific community to make advances in the field of influenza forecasting. A key observation from these challenges is that a simple average of the submitted forecasts outperformed nearly all of the individual models. Further, ongoing efforts seek ways to assign weights to individual models to create high-performing ensemble models. Given the sheer number of models, as well as variation in methodology followed among teams contributing influenza-risk forecasts, multiple forecasting models can be combined, by capturing human judgment, to outperform a simple average of these same models. This project exploits such a “wisdom of crowds” approach, using public votes acquired with the help of an R/Shiny based web-application platform in order to assign weights to individual forecasting models on a week-over-week basis, in an effort to improve overall ILI risk prediction accuracy. We describe a strategy for improving the accuracy of influenza risk forecast modeling based on a crowd-sourced set of team-specific forecast votes and the results of the 2017–2018 season. Our approach to assigning weights based on crowd-sourced votes on individual models outperformed an average forecasts of the individual models. The crowd was statistically significantly more accurate than the average model and all but one of the individual models.

The COVID-19 pandemic has disrupted the lives of millions across the globe. In Singapore, promoting safe distancing by managing crowds in public areas have been the cornerstone of containing the community spread of the virus. One of the most important solutions to maintain social distancing is to monitor the crowdedness of indoor and outdoor points of interest. Using Nanyang Technological University (NTU) as a testbed, we develop and deploy a platform that provides live and predicted crowd counts for key locations on campus to help users plan their trips in an informed manner, so as to mitigate the risk of community transmission.

With the rise in popularity of short-term Human Trajectory Prediction (HTP), Long-Term Crowd Flow Prediction (LTCFP) has been proposed to forecast crowd movement in large and complex environments. However, the input representations, models, and datasets for LTCFP are currently limited. To this end, we propose Fourier Isovists, a novel input representation based on egocentric visibility, which consistently improves all existing models. We also propose GeoInteractNet (GINet), which couples the layers between a multi-scale attention network (M-SCAN) and a convolutional encoder-decoder network (CED). M-SCAN approximates a super-resolution map of where humans are likely to interact on the way to their goals and produces multi-scale attention maps. The CED then uses these maps in either its encoder's inputs or its decoder's attention gates, which allows GINet to produce super-resolution predictions with substantially higher accuracy than existing models even with Fourier Isovists. In order to evaluate the scalability of models to large and complex environments, which the only existing LTCFP dataset is unsuitable for, a new synthetic crowd dataset with both real and synthetic environments has been generated. In its nascent state, LTCFP has much to gain from our key contributions. The Supplementary Materials, dataset, and code are available at sssohn.github.io/GeoInteractNet.

Attaining the proper balance between underfitting and overfitting is one of the central challenges in machine learning. It has been approached mostly by deriving bounds on generalization risks of learning algorithms. Such bounds are, however, rarely controllable. In this study, a novel bias-variance balancing objective function is introduced in order to improve generalization performance. By utilizing distance correlation, this objective function is able to indirectly control a stability-based upper bound on a model's expected true risk. In addition, the Generalization-Aware Collaborative Ensemble Regressor (GLACER) is developed, a model that bags a crowd of structured regression models, while allowing them to collaborate in a fashion that minimizes the proposed objective function. The experimental results on both synthetic and real-world data indicate that such an objective enhances the overall model's predictive performance. When compared against a broad range of both traditional and structured regression models GLACER was ~10-56% and ~49-99% more accurate for the task of predicting housing prices and hospital readmissions, respectively.