The 'risk of risk': remodelling artificial intelligence algorithms for predicting child abuse.

英国的儿童保护在很大程度上依赖于风险预测，这一领域自20世纪80年代末以来在英国日益受到关注(Browne&Saqi 1988，Creighton 1992)。人们普遍认为，可以而且应该通过风险预测来发现虐待儿童的情况，以确定其儿童可能受到虐待或被忽视的脆弱和危险家庭。及早识别这类家庭的目的，是为了及早对他们进行干预，以减少虐待的风险。为了满足这一需求，个别地方当局委托盈利提供商提供算法风险预测系统。这个拟议的项目解决的问题是，鉴于相关的纵向数据显示，儿童保护结果的准确性很差，以及在风险预测方面的假阳性和假阴性的数量高得令人无法接受，这种系统是否“适合于目的”。家庭司司长最近强调了这一关切(Munby 2016)。这一拟议的项目通过探讨以更现实的方式预测风险的新方法的可能性来解决这一问题，为儿童保护系统提供更好的手段，使其得到它们的支持，而不是不得不处理可能不准确的数据。它提出了一种新的、变革性的方法，可以从以前的研究中整理、评估和提取一致的信息，并以一致和可靠的方式对其进行测试。存在着将算法风险预测移动到新领域的新系统的范围的可能性；现有系统没有从这些错误中‘学习’，因此该技术在算法预测阶段停滞不前，而不是发展成基于证据的、可靠的和反应灵敏的人工智能(AI)。关键研究问题/目标是：-什么是现有儿童保护风险预测研究中的归一化置信限(S)；-开发一种计算风险的新方法，并为其在儿童保护中的应用进行设计；-评估设计一种新的、符合GDPR的人工智能风险预测模型的可能性，该模型适用于诉讼前和诉讼后的儿童保护工作。这项研究的方法具有变革性，将传统方法和开创性方法结合在一起。对该方法的每个阶段都进行了评估，以确定其方法和/或结果的潜在转变程度。该小组将开始拟议的项目，首先创建以前相关研究的第一个全面和可重复使用的数据库。其余研究人员采用的创造性和新方法风险较高，但如果成功，将产生相应的高回报。在创建了研究数据库后，该小组将分析它们的特点、规模、范围和方法，以应用一致的方法计算它们的功率比，创建包括优势、劣势和可信度在内的比较分析。这些结果将在Eggleston关于在事实发现过程中使用概率的工作的背景下使用贝叶斯统计进行分析(Eggleston，1983)。贝叶斯网络提供了一种新的方法来建立社会和技术问题的证据加权标准，包括推理(使用贝叶斯推理算法)、学习(使用期望最大化算法)、规划(使用决策网络)和感知(使用动态贝叶斯网络)。概率算法还可以用于过滤、预测、找到数据流的解释，以及帮助系统分析随时间推移的过程。在这种情况下，我们将提供对各种风险因素的信心的一致衡量，以及对其证据正直的衡量。我们方法的这一核心变革因素将使风险预测系统的范围能够考虑到优势和劣势，包括确定差距，向法院提供可靠的法律指标，说明作为项目成果的适当权重。

Child protection in the UK relies heavily on risk prediction, an area of growing interest in the UK since the late 1980s (Browne & Saqi 1988, Creighton 1992). It is generally taken as an axiom that child abuse can and should be detected via risk prediction to identify vulnerable and risky families whose children may become abused or neglected. The purpose of identifying such families at an early stage is to target early intervention towards them to reduce the risk of abuse. To service this need, individual local authorities commission algorithmic risk prediction systems from profit making providers. The question this proposed project addresses is whether such systems are 'fit for purpose' given the concerning longitudinal data showing poor accuracy in child protection outcomes and an unacceptably high number of false positives and false negatives in risk prediction. This concern was recently highlighted by the President of the Family Division (Munby 2016).This proposed project addresses the issue by looking at the possibilities for a new method of predicting risk in a more realistic way that provides a better means for child protection systems to be supported by them, rather than have to work potentially inaccurate data. It sets out a new and transformative means of collating, assessing and extracting consistent information from previous studies and testing them in a consistent and reliable way. The potential exists for scoping a new system which moves algorithmic risk prediction into new territory; existing systems do not 'learn' from these errors so the technology stalls at the stage of algorithmic prediction rather than developing into evidenced-based, reliable and responsive artificial intelligence (AI).The key research questions/objectives are:- What is a normalised confidence limit(s) in existing risk prediction studies in child protection;- To develop a new method of calculating risk, and design for its application in child protection;- To assess the possibility of designing a model for a new, GDPR-compliant, AI model of risk prediction suitable for use in pre- and post-proceedings child protection work.This study's methodology is transformative, bringing together a mix of traditional and pioneering methods. Each stage of the methodology has been assessed for the level of potential transformation in either its approach and/or outcome. The team will start the proposed project by creating the first, comprehensive and re-usable database of previous relevant studies. The creative and new methods employed by the rest of the study is higher risk, but if successful will yield a correspondingly high reward. Having created the database of studies, the team will analyse their characteristics, size, scope and methods to apply a consistent means of calculating their power ratio, creating a comparative analysis including strengths, weaknesses and confidence limits. These results will be analysed using Bayesian statistics in the context of Eggleston's work in respect of the use of probability in fact finding processes (Eggleston 1983). Bayesian networks provide a novel means of establishing criteria for weighting of evidence for social and technical problems including reasoning (using the Bayesian inference algorithm), learning (using the expectation-maximization algorithm), planning (using decision networks) and perception (using dynamic Bayesian networks). Probabilistic algorithms can also be used for filtering, prediction, finding explanations for datastreams, and helping systems to analyse processes over time. Used in this context, we will provide a consistent measure of confidence across risk-factors and measure of their evidential probity. This core transformative element of our methods will enable scoping of a risk prediction system to take account of strengths and weaknesses, including identifying gaps, providing a reliable legal indicator to courts as to the appropriate weighting as a project outcome.