Social network analysis for customer churn prediction

Incorporating textual information in customer churn prediction models based on a convolutional neural network

In the mobile telecommunication industry, call networks have been used with great success to predict customer churn. These social networks are complex and rich in features, because the telecommunications operators have a lot of information about their customers. In this paper we leverage a novel framework called GraphSAGE for inductive representation learning on networks with the goal of predicting customer churn. The technique has an advantage over previously proposed representation learning techniques because it leverages node features in the learning process. It also features a supervised learning process, which can be used to predict churn directly, as well as an unsupervised variant which produces an embedding. We study how the number of node features impacts the predictive performance of churn models as well as the benefit of a complete learning process, compared to an embedding with supervised machine learning techniques. Finally, we compare the performance of GraphSAGE to that of standard local models.

The recommender system is of great significance to alleviate information overload. The rise of online social networks leads to a promising direction—social recommendation. By injecting the interaction influence among social users, recommendation performance has been further improved. Successful as they are, we argue that most social recommendation methods are still not sufficient to make full use of social network information. Existing solutions typically either considered only the local neighbors or treat neighbors’ information equally, even or both. However, few studies have attempted to solve these social recommendation problems jointly from both the perspective of social depth and social strength. Recently, graph convolutional neural networks have shown great potential in learning graph data by modeling the information propagation and aggregation process. Thus, we propose an attention-based social aggregation neural networks (abbreviated as SAN) model to build a recommendation system. Different from previous work, our proposed SAN model simulates the recursive social aggregation process to spread the global social influence, and simultaneously introduces social attention mechanism to incorporate the heterogeneous influences for better model user embedding. Instead of a shallow linear interaction function, we adopt multi-layer perception to model the complex user–item interaction. Extensive experiments on two real-world datasets show the effectiveness of our proposed model SAN, and further analysis verifies the generalization and flexibility of the model.

조직유효성(직무만족, 조직몰입)은 기업 경쟁력 가치사슬의 근간을 이루기 때문에, 조직성과모델의 중요한 선행변수이다. 그래서 많은 연구들에서 조직유효성의 원인들을 규명하고자 하였는데, 주로 인적자원관리 변수들인 직무특성(자율성, 책임감, 도전심), 임금, 승진, 복지정책이나 관계 변수들로 상사의 리더십이나 동료 관계 등을 논의하였다. 한편 조직구성원들의 구직 경로에 관한 기존 연구들은 취업 희망자들이 구직 정보에 대한 불확실성과 비대칭성을 줄이기 위해서 다양한 사회적 네트워크-개인적 친분, 부모, 친구, 이웃 등을 활용하는 것이 효과적이라는 결과를 제시하고 있다. 이 연구는 앞서 언급한 두 가지 개념을 사회적 자원동원 관점에서 통합적 시각에서 접근하였다. 즉, 사회적 네트워크를 통한 구직과 조직유효성의 관계를 논의하였다. 이를 위해 문헌들을 고찰하고, 사회적 네트워크를 통한 구직과 조직유효성의 관계에 관한 이론을 구축하여 명제들을 제안하였다. 주요 연구내용 및 결과는 다음과 같다. 첫째, 구직의 사회적 네트워크 이론으로 균형이론, 약한 유대의 강함, 사회적 자원동원이론 등을 고찰하였다. 그리고 사회적 네트워크를 통한 구직 유형을 사회 유대관계 차원(약함과 강함)과 사회 위치적 차원(부모와 자신의 사회적 지위)을 기준으로 구분하여, 네 가지 구직 유형을 제시하였다. 둘째, 직무만족에 관한 기존 연구들을 고찰하고, 앞서 제시한 구직 유형들과의 관련성을 제시하였다. 네 가지 유형 가운데 구직자 부모의 사회적 지위가 높고, 부모가 다양한 네트워크를 보유할수록, 구직자는 더 좋은 직장에 입사할 가능성과 직무에 대한 만족도가 클 수 있다는 점을 제시하였다. 그 이유로 부모의 사회적 네트워크를 통해 구직에 관한 광범위하고 다량의 정보수집 효과와 구직자의 임금, 조직 현실 등에 관한 정보 확실성, 면역효과, 그리고 인지부조화 등을 논의하였다. 셋째, 조직몰입에 관한 기존 연구들을 고찰하고, 사회적 네트워크를 통한 구직과의 관련성을 제시하였다. 결론적으로 사회적 네트워크를 통해 구직한 구성원은 조직에서 지속적으로 근무하고자 하는 지속적 몰입과 도덕적 의무감에 의한 규범적 몰입이 높다는 점이다. 그 이유로 사회적 네트워크 구직 비용, 심리적 계약이론, 규범적 몰입의 도덕적 측면 등을 논의하였다. 마지막으로 연구의 시사점 및 한계점 그리고 향후 실증분석을 위한 측정도구들과 방법들을 논의하였다.This study discussed the relationship between job seeking via social networks and organizational effectiveness. The antecedents of job satisfaction and organizational commitment, which are variables of organizational effectiveness, have been discussed mainly on HR variables such as job characteristics, wages, and the company s promotion policy and welfare policy. In this study, it was hypothesized that the organizational effectiveness of employees may be determined by whether they mobilize networks based social resources or not, when they land a job. In terms of social resource mobilization perspective, this study built a theory on the relationship between job seeking via social networks and organizational effectiveness. To verify the relationship, the literatures on the concept of social networks and the types of job seeking via social networks were reviewed, after which the propositions between job seeking via social networks and organizational effectiveness was theoretically suggested. The results of this study are as follows. (i) Based on two dimensions (relational dimension (strong tie or weak tie) and positional dimension (social position of parents or social position of me)), four distinctive types of job seeking via social networks were suggested. (ii) Among the four types of job seeking identified, the higher the social status of the job seeker s parents was and the wider their social network was, the higher the chance the job seeker had of landing a better job and the higher his job satisfaction was. (iii) Job seekers who landed a job via their social network usually demonstrated a higher level of continuance commitment and normative commitment. (iv) Finally, the implications and limitations of this study, as well as suggestions for further study, were presented.

Introduction The residential neighbourhood is thought to affect residents because of presumed neighbourhood effects; the independent effects of a neighbourhood’s characteristics on the life chances of its residents. An enormous body of research has tried to measure neighbourhood effects, however, there are no clear conclusions on how much, if any, effect the neighbourhood has on its residents. There is non-random selection of people into neighbourhoods which causes a bias in the modelling of neighbourhood effects. Any correlation found between neighbourhood characteristics and individual outcomes might be explained by selection bias and can therefore not prove the existence of a causal neighbourhood effect. The question is; do poor neighbourhoods make people poor, or do poor people live in unattractive neighbourhoods because they cannot afford to live elsewhere (Cheshire, 2007). Therefore, insight in selection is important to gain more insight in neighbourhood effects (Van Ham and Manley, 2012). For neighbourhood effects research it is important to study selective mobility and neighbourhood choice and to combine neighbourhood effects research with neighbourhood selection research (Doff, 2010a; Van Ham and Manley, 2012; Van Ham et al., 2012; Galster, 2003; Hedman, 2011). The aim of this thesis therefore is to gain more insight in both the causes and the consequences of segregation and thus to study both individual residential mobility and neighbourhood selection and neighbourhood effects. Besides the neighbourhood effects literature, also the segregation literature will benefit from better insights in selective residential mobility because selective residential mobility is one of the main driving forces of segregation. There are two main research questions for this thesis. Firstly, I try to give insight in selective mobility and neighbourhood choice and thus to study where, when and why which people move. What is the effect of personal characteristics, neighbourhood characteristics and macro level housing market developments on individual neighbourhood satisfaction, moving wishes, moving behaviour and neighbourhood selection and on macro level selective mobility patterns and segregation? Secondly, I will test presumed neighbourhood effect mechanisms. Concentration areas of ethnic minorities are seen as undesirable, because their residents are thought to have less contact with the native majority which might hamper their integration and their life chances. It is, however, unclear to what extent social contact is affected by the residential neighbourhood. The second research question therefore asks whether ethnic minorities have less contact with the native majority if they live in minority concentration neighbourhoods. Segregation Segregation is defined as the population composition of neighbourhoods in relation to each other; that is, the concentration or underrepresentation of population groups in neighbourhoods compared to a city or national level average. Selective residential mobility is one of the main driving forces of segregation. Households move to a certain neighbourhood, either because they choose to live there, or because they are constrained in their choice options. Therefore segregation can be both voluntary and involuntary. This thesis focuses on selective residential mobility as cause of segregation, therefore it tries to understand why and where people move. According to residential mobility theory, personal characteristics determine residential preferences and if the residential situation is not in line with these preferences this will lead to dissatisfaction and a desire to move (Brown and Moore, 1970). Whether a dissatisfied household succeeds in moving to a dwelling and neighbourhood more in line with their preferences, depends on their personal resources and restrictions and macro level opportunities and constraints (Mulder and Hooimeijer, 1999). Successful households will move to a neighbourhood more in line with their preferences. There are, however, differences between households in which neighbourhoods are open to choice. Low income households will only be able to select neighbourhoods in which inexpensive dwellings are available. Similarly, households who depend on the social housing sector, or on the owner-occupied sector will only be able to select neighbourhoods where dwellings of this tenure are available. In addition, there are differences between households in which neighbourhood is (deemed) most attractive. People prefer to live among others who are similar to themselves and also facilities directed towards specific groups will make especially concentration neighbourhoods of the own ethnic or income group attractive. Besides residential preferences, also other factors will affect neighbourhood selection; population groups will differ in access to information on neighbourhood attractiveness or housing opportunities and discrimination, or fear of discrimination, can limit the opportunities of minority groups on the housing market. Neighbourhood effects It is typically assumed in European and American urban policy and academic research that spatial concentrations of low income households or ethnic minorities have negative effects on their inhabitants (Friedrichs et al., 2003). An enormous body of research has tried to measure neighbourhood effects; the independent effect of a neighbourhood on its residents when controlling for individual characteristics (see for a review Dietz, 2002; Ellen and Turner, 1997; Sharkey and Faber, 2014; Van Ham et al., 2012). The research attention for neighbourhood effects started with the seminal work of Wilson (1987). He argued that living in concentration areas of the jobless lowest class, isolated from role models, mainstream values and norms, and informal job networks and social contacts with employed, has a negative effect on your life chances. Neighbourhoods can affect their residents via a number of mechanisms (Ellen and Turner, 1997; Erbring and Young, 1979; Galster, 2012). Firstly, the geographical location determines job access and thereby labour market opportunities. Secondly, pollution, noise and disturbance affect health and (thereby) life chances via environmental mechanisms . In addition, neighbourhood stigmatisation can reduce life chances because others have prejudiced ideas and low expectations of the residents of stigmatised neighbourhoods. Also, the quality of institutions such as schools, museums, libraries and sport facilities will generally be lower in disadvantaged neighbourhoods, thereby reducing the life chances of residents. Finally, neighbourhood effects can transpire via social interactions with neighbours. Through collective socialisation and peer pressure people are thought to conform to local social norms (Jencks and Mayer, 1990). Positive role models and social network contacts with individuals with more social capital can help people advance in their work or educational career (Ellen and Turner, 1997). For ethnic minorities it can be important to have social interactions with the native majority to have the opportunity to learn the majority language, standards and values (Lazear, 1999) and to have bridging network ties that can provide access to valuable information not present within the own ethnic network (Buck, 2001). Many neighbourhood effect researchers believe that living in concentrated poverty has negative effects on individuals, and policymakers try to create mixed neighbourhoods to prevent these negative neighbourhood effects. However, although “it is perfectly plausible that poor people are made poorer by the characteristics of the neighbourhoods in which they live” (…) “a close examination of the best research available does not reveal any clear evidence to support it” (Cheshire, 2007: p. ix). Almost all neighbourhood effect studies struggle with selection bias (Cheshire, 2007). Any relation found between neighbourhood characteristics and individual outcomes might be a selection effect and therefore cannot prove the existence of a causal neighbourhood effect. Neighbourhood effects researchers have tried to reduce or eliminate selection bias. Firstly this is done by using quasi-experimental study designs, using households whose residential neighbourhood is determined by external factors (Sampson et al., 2002). Secondly, advanced statistical methods such as sibling studies, fixed effects studies, instrumental variables and propensity score matching are used to reduce selection bias or to control for selection (Harding, 2003). These advancements in methodology have improved our insight in selection bias and in neighbourhood effects, however, since there are no methods that can completely eliminate selection bias, there is still no clear evidence of causal neighbourhood effects. Generally, methods that apply more controls for selection bias find smaller neighbourhood effects. Instead of trying to eliminate selection bias, this thesis tries to provide insight in selection, in why and where which people move. In addition, this thesis tests presumed neighbourhood effect mechanisms. Social interactive mechanisms assume that neighbourhood effects transpire because the population composition of the residential neighbourhood affects with whom you interact (Ellen and Turner, 1997; Galster, 2012). Therefore I test whether the ethnic composition of the neighbourhood affects interethnic contact. Individual differences in determinants of residential satisfaction<

We formalize the derivation of a generalized coarse-graining $n$-resolved master equation by introducing a virtual detector counting the number of transferred charges in single-electron transport. Our approach enables the convenient inclusion of coherences and Lamb shift in counting statistics. As a Markovian example with Lindblad-type density matrices, we consider the Born-Markov-Secular (BMS) approximation which is a special case of the non-Markovian dynamical coarse graining (DCG) approach. For illustration we consider transport through two interacting levels that are either serially or parallelly coupled to two leads held at different chemical potentials. It is shown that the coherences can strongly influence the (frequency-dependent) transport cumulants: In the serial case the neglect of coherences would lead to unphysical currents through disconnected conductors. Interference effects in the parallel setup can cause strong current suppression with giant Fano factors and telegraph-like distribution functions of transferred electrons, which is not found without coherences. We demonstrate that with finite coarse graining times coherences are automatically included and, consequently, the shortcomings of the BMS approximation are resolved.

Social networks have interesting properties. They influence our lives enormously without us being aware of the implications they raise. The authors investigate the following areas concerning social networks: how to exploit our unprecedented wealth of data and how we can mine social networks for purposes such as marketing campaigns; social networks as a particular form of influence, i.e.., the way that people agree on terminology and this phenomenon's implications for the way we build ontologies and the Semantic Web; social networks as something we can discover from data; the use of social network information to offer a wealth of new applications such as better recommendations for restaurants, trustworthy email senders, or (maybe) blind dates; investigation of the richness and difficulty of harvesting FOAF (friend-of-a-friend) information; and by looking at how information processing is bound to social context, the resulting ways that network topology's definition determines its outcomes.

A linear threshold-hurdle model for product adoption prediction incorporating social network effects

Introduction The presentation of health institutions in the internet is highly variable concerning marketing features and medical information. We aimed to investigate the structure and the kind of information provided on the Web sites of all departments of pediatric surgery in Germany. Furthermore, we aimed to identify the degree to which these Web sites comply with internet marketing recommendations for generating business. Method The Web sites of all pediatric surgery units referred to as departments on the official Web site of the German Society of Pediatric Surgery (GSPS) were assessed. The search engine Google was used by entering the terms "pediatric surgery" and the name of the city. Besides general data eight content characteristics focusing on ranking, accessibility, use of social media, multilingual sites, navigation options, selected images, contact details, and medical information were evaluated according to published recommendations. Results A total of 85 departments of pediatric surgery were included. On Google search results 44 (52%) ranked number one and 34 (40%) of the department's homepages were accessible directly through the homepage link of the GSPS. A link to own digital and/or social media was offered on 11 (13%) homepages. Nine sites were multilingual. The most common navigation bar item was clinical services on 74 (87%) homepages. Overall, 76 (89%) departments presented their doctors and 17 (20%) presented other staff members with images of doctors on 53 (62%) and contact data access from the homepage on 68 (80%) Web sites. On 25 (29%) Web sites information on the medical conditions treated were presented, on 17 (20%) details of treating concepts, and on 4 (5%) numbers of patients with specific conditions treated in the own department per year. Conclusion We conclude that numerous of the investigated online presentations do not comply with recommended criteria for offering professional information for patients and for promoting services. Only less than one-third of the departments of pediatric surgery in Germany offer information about the medical conditions they treat. Features, which may influence the decision of patients and parents such as ranking, accessibility, use of social media, multilingual sites, navigation options, selected images, and contact information were differently lacking on many Web sites.

With the proliferation of smart embedded devices, cyber physical social computing and networking systems (CPSCN) are emerging as a next generation of social networks. Unlike traditional social networks that run on cloud-based infrastructure, CPSCN systems usually depend on a large number of distributed, heterogeneous devices, such as mobile phones, smart vehicles, or network access points. These computing resources, which are often referred to as fog computing systems, provide a gateway to the physical world, and thus offer new possibilities for social applications. Unfortunately, building CPSCN systems that leverage fog computing infrastructure is not straightforward. Significant challenges arise from the large scale distribution of computing resources over a wide area, and the dynamic nature of multiple, possibly mobile, hosts. In this paper, we extend our previous work on a distributed dataflow programming model and propose an application platform for realizing CPSCN systems. A key aspect of our work is the development of an exogenous coordination model, which exhibits a separation of concern between computation and communication activities, and helps resolve some of the challenges brought about by the dynamic and large scale nature of CPSCN systems.

This thesis explores the application of complex systems research to understand the dynamics of the art world, considering art as a complex system and investigating its various components through data-driven methodologies. By studying art as a complex system, we contribute to a systematic understanding of human development and behavior in creative domains. The social network plays a crucial role in the cultural evolution of art, shaping our cultural identity and collective memory. Analyzing network characteristics provides insights into how individual decisions influence collective dynamics and sustain social phenomena. Previous studies have used network models and data analysis to examine the role of network position and connectedness in artistic collaborations, individual success, and the transmission of artistic knowledge. In this thesis, we focus on ballet as a unique art form with a rich historical and social structure. Ballet provides an opportunity to investigate the role of the social network in shaping collective dynamics in performing arts. We present two research articles that examine gender inequalities and the role of social connections on the career success of ballet dancers. The first article investigates the social network structure of ballet creations and its potential impact on gender disparities in leading positions. The second article explores the influence of social connections and prestige on the career trajectories of ballet dancers, using network analysis and centrality metrics to uncover hierarchical stratification within ballet academies. Our research highlights the significance of social dynamics and network effects in understanding complex social phenomena in the art world. It offers insights into gender inequalities and career success in ballet and demonstrates the value of data-centric methodologies in arts research. By generating a unique dataset and applying interdisciplinary approaches, we contribute to the scientific examination of the arts and enhance our understanding of human creativity and cultural heritage. This thesis contributes to the broader goal of fostering diversity, equity, and inclusion within the arts by shedding light on social structures and suggesting potential avenues for change. Lastly, this work underscores the importance of interdisciplinary research in enriching our understanding of human development and behavior in creative domains.

Butterfly Effect in Social Network: Analysis of Churn Under Network Influence

Traditional mobile crowdsensing has shown a great potential to address large-scale data sensing problems by allocating sensing tasks to pervasive mobile users with smart devices. As crowdsensing service provider involves data collection from mobile users, the issue of providing rewards to incentivize mobile users is fundamentally important to ensure the sustainability of crowdsensing system. Recently, socially-aware crowdsensing services have been introduced as the integration of social networks and crowdsensing platforms. For example, in health-related crowdsensing applications, a mobile user benefits from information regarding food, exercise, medicine and medical treatment collected and shared by her socially-connected friends and family members. In this thesis, incentive mechanism design is revisited in the context of socially-aware crowdsensing, where the game theory tool is used to study the interactions between crowdsensing service provider and mobile users. In the first part of the thesis, incentive mechanism is designed by considering the underlying social network effects amid mobile social networks, for motivating the participants. Namely, one MU will obtain additional benefits from information contributed or shared by local neighbors in social networks. The process of rewarding and participating is modeled as a two-stage game, and backward induction is used to analyze the mobile users' participation level and the crowdsensing service provider's optimal reward mechanism. The analytical expressions are derived for the discriminatory incentive as well as the uniform incentive mechanisms. In the second part of the thesis, information asymmetry challenges are investigated. A Bayesian Stackelberg game with incomplete information is formulated to analyze the interaction between the crowdsensing service provider and mobile users, where the social structure information, i.e., the social network effects, among mobile users is uncertain. The existence and uniqueness of the Bayesian Stackelberg equilibrium is analytically validated by identifying the best response strategies of the mobile users. To be more general and practical, different with the former two parts, incentive mechanism in presence of multiple crowdsensing service providers is studied in the third part of the thesis. Understanding the behaviors of mobile users and service providers in socially-aware crowdsensing is of paramount importance for incentive mechanisms. With this focus, a multi-leader and multi-follower Stackelberg game approach is proposed to model the strategic interactions among service providers and mobile users, where the social influence of mobile users and the strategic interconnections of service providers are jointly and formally integrated into the game modeling. Through backward induction methods, the existence and uniqueness of the Stackelberg equilibrium are theoretically proved. In summary, this thesis addresses a few urgent challenging problems with the tools of game theory in the real implementation of incentive mechanism design in the context of socially-aware crowdsensing. Extensive numerical simulations are also conducted to illustrate some important properties of the equilibrium. Finally, several promising research directions are outlined for the future work.

In an attempt to replicate Berkman and Syme's study of social networks and mortality in Alameda County, California, the authors investigated the relationship between a social network index and survivorship from 1967 to 1980 in the Evans County, Georgia, cohort. They constructed an index modeled after the Berkman Social Network Index and tested it in race- and sex-specific proportional hazards models for 2,059 subjects who were examined in 1967-1969 during the Evans County Cardiovascular Epidemiologic Study. The present study emphasized a priori specification of the social network index and statistical hypothesis test. Descriptive analyses were consistent with a modest social networks effect (e.g., hazard ratio (95 per cent confidence interval) of 1.6 (1.2-2.2) ). Among white males, the age-adjusted hazard ratio comparing the lowest to the highest value of our six-level index was 2.0 (1.2-3.4), but control for potential confounders (principally cardiovascular disease risk factors) reduced this value to 1.5 (0.8-2.6). The social networks effect among white females, black males, and black females was weaker and clearly nonsignificant. Exploratory analyses suggested that marital status, church activities, and an alternate social network index predicted survivorship, but not in a dose-response fashion. Reduced survivorship among older subjects with few social ties was the most important feature of the data.

The DR-SCIR network public opinion propagation model was employed to study the characters of S-state users stopping transmitting information for the first time and secondary transmission of immune users. The model takes into account symmetry and complexity such as direct immunization and social reinforcement effect, proposes the probability of direct immunity Psr and the probability of transform from the immune state to the hesitant state Prc, and divides public opinion information into positive public opinion and negative public opinion based on whether the public opinion information is confirmed. Simulation results show that, when direct immunity Psr = 0.5, the density of I-state nodes in the model decreased by 54.12% at the peak index; when the positive social reinforcement effect factor b = 10, the density of I-state nodes in the model increased by 16.67% at the peak index; and when the negative social reinforcement effect factor b = -10, the density of I-state nodes in the model decreased by 55.36% at the peak index. It shows that increasing the positive social reinforcement effect factor b can promote the spread of positive public opinion, reducing the negative social reinforcement effect factor b can control the spread of negative public opinion, and direct immunization can effectively suppress the spread of public opinion. This model can help us better analyze the rules of public opinion on social networks, so as to maintain a healthy and harmonious network and social environment.