A privacy-preserving socio-demographic enrichment framework for big data and its empirical application

近几十年来，立法机构和公众对隐私保护的意识和需求大幅增长。这可以部分归因于信息和通信技术的扩散，这些技术在使用时会产生大量数据（即大数据）。与此同时，越来越多的大数据的可用性为它们的应用创造了机会，利用它们可以洞察人们的日常行为模式。然而，由于社会人口信息作为人类行为的外生决定因素的重要作用，匿名大数据在基于行为的分析中的适用性受到限制。因此，出现了大量的研究来预测各种大数据源中受访者缺乏的社会人口统计学属性，称为大数据的社会人口统计学丰富性。现有的社会人口统计学丰富方法要么使用基于性能的数据挖掘和机器学习方法，要么使用面向统计拟合的模型，这些模型通常缺乏理论基础，无法解释或证明受访者的行为模式（输入特征）与其社会人口统计学属性（丰富的输出）之间的假设关系。然而，理论基础是至关重要的，因为微观经济消费者理论表明，人们的行为是由他们的社会人口属性驱动的。忽视基础的微观经济学和/或社会学行为理论的一个直接后果是，现有方法既无法预测富集的质量，也无法解释由于数据分布的变化而导致的性能变化。这激发了我的博士研究，我提出并形式化了一个新的丰富框架，称为逆离散选择建模（IDCM）框架。IDCM社会人口浓缩框架允许定量地理解浓缩准确性和隐私保护之间的权衡。具体来说，IDCM方法对离散选择模型（DCM）进行统计反转，这是一种成熟的建模技术，依赖于基于社会科学的明确行为假设，包括微观经济学、社会学和心理学。建立了IDCM性能理论，根据已知的富集样品中数据分布信息来估计IDCM富集性能。这是通过在信息论中对人类行为进行类比来实现的，即观察个体作为通过信息通信通道传输的“消息”，这允许使用几个强大的信息论概念来数学地联系我们如何预测这个人是谁和他/她的隐私。到目前为止，IDCM绩效理论的能力是为了丰富社会人口对单一二元行为特征的观察而发展起来的。为了提高经验丰富的性能，本研究项目的目标是通过将多种行为模式作为输入特征来扩展当前的IDCM方法。这可以通过使用几个dcm来实现，这些dcm分别捕获每个行为特征与丰富属性之间的关系，然后找到最有可能导致联合行为模式的社会人口统计属性的值。IDCM方法的拟议扩展涉及结合机器学习或深度学习算法，用于从原始大数据中提取有意义的行为模式，这些行为模式可以进一步用作后续IDCM丰富的输入特征。相应的，随附的IDCM性能理论将得到相应的扩展，以适应基于使用多种行为特征的浓缩性能估计，以保持拟议的IDCM方法扩展的可转移性。

Recent decades have seen a substantial growth in the awareness and demand for privacy preservation set in legislation and by the public. This can be partly attributed to the proliferation of information and communications technologies, which generates large amounts of data (i.e. big data) when used. In parallel, the growing availability of big data has created opportunities for their applications, drawing upon the insights they can provide insights into people's daily behaviour patterns. However, the applicability of anonymous big data has been limited in behaviour-based analysis because of the essential role of socio-demographic information as exogenous determinants of human behaviour. Therefore, a plethora of studies have emerged to predict the absent socio-demographic attributes of respondents in various big data sources, termed the socio-demographic enrichment of big data.Existing socio-demographic enrichment approaches use either performance-based data mining and machine learning methods or statistical-fit-oriented models, which typically lack theoretical underpinnings that can explain or justify the postulated relationship between respondents' behaviour patterns (input features) and their socio-demographic attributes (output of the enrichment). A theoretical underpinning is, however, crucial because microeconomic consumer theory suggests that people's behaviour is driven by their socio-demographic attributes. One immediate consequence to neglect the underpinning microeconomic and/or sociological behavioural theoriesconcerns the incapability of existing methods to either predict the quality of enrichment or interpret the change in their performance due to the variation in data distributions. This motivates my PhD research in which I propose and formalise a new enrichment framework, called the Inverse Discrete Choice Modelling (IDCM) framework. The IDCM socio-demographic enrichment framework allows to quantitatively understand the trade-offs between enrichment accuracy and privacy preservation. Specifically, the IDCM approach performs statistical inversion to a discrete choice model (DCM), which is a well-established modelling technique relying on explicit behavioural assumptions grounded in social science, including microeconomics, sociology and psychology. The IDCM performance theory is established to estimate the IDCM enrichment performance based on known information about the data distribution in the enriched sample. This is enabled by drawing an analogy of human behaviour in information theory, i.e. observed individual as a 'message' transmitted over an information communication channel, which allows to use several powerful information-theoretic concepts to mathematically link how well we can predict who the person is and his/her privacy.So far, the ability of the IDCM performance theory is developed for socio-demographic enrichment of observation of a single, binary behaviour feature. To improve the empirical enrichment performance, the aim of the proposed research project is to extend the current IDCM approach by including multiple behaviour patterns as the input features. This can be achieved by using several DCMs that respectively captures the relationship between each behaviour feature and the enriched attribute and then to find the value of the socio-demographic attribute that is most likely to result in the joint behaviour patterns. The proposed extension of the IDCM approach involves the incorporation of machine learning or deep learning algorithms, applied to extract meaningful behaviour patterns, from raw big data, that can be further employed as the input feature for the subsequent IDCM enrichment. Correspondingly, the accompanying IDCM performance theory will be extended accordingly to accommodate the estimation of the enrichment performance based on the use of multiple behaviour features to retain transferability of the proposed extension of the IDCM methodology.