单类分类方法结合光谱分析在食品真实性鉴别中的应用

近年来，假冒伪劣食品已日益成为广大消费者密切关注的问题，食品真实性评估是缓解这一问题、保护公众健康的有力手段。在仪器设备和样品处理的高要求下，现代检测技术通常需要大量时间和金钱的成本消耗，而如今食品掺假手段不断变换，花样日益翻新，使得这类检测技术存在一定的局限性。为促进食品安全质量监管的效率和水平提高，为监管工作提供有力的科学技术支撑和保障，需要寻求新型检测技术。光谱分析技术，以操作简单、快速无损的优势近年来被广泛应用，作为一种间接分析技术，结合数据统计学中的分类方法建立模型后更能有效进行真假鉴别。在分类方法中，由于现实生活中五花八门的掺假类型以及在真假样本数量差异大的情况下，常用的分类方法效果可能出现偏差。但单类分类方法（one-class classification）是一种只针对一类实例建模分析，以特定的置信水平固定目标样本类的边界，对新样本的类别进行判定的方法，利用这一特点能有效区分不同于真实样本的数据，大大减少了检测的工作量，在食品掺假检测应用领域有一定的发展潜力。对近年来模式识别中的分类方法——单类分类方法进行了综述。通过阐述光谱分析结合分类方法用于食品掺假检测的必要性，比较在同一情形下多类分类方法和单类分类方法的判别率，简介单类分类方法的特点，并重点介绍几种常见的单类分类方法如数据驱动的簇类独立软模式（DD-SIMCA）、单类偏最小二乘（OCPLS）、单类支持向量机（OCSVM）以及单类随机森林（OCRF），论述单类分类方法在食品真实性鉴别中的应用，具体在食用油，乳制品，饮料，保健品，香辛料及谷物方面进行了阐述，还分析了当前单类分类方法存在的问题，最后对该技术的应用前景进行展望，为食品认证分析提供了一定的理论依据。

Application of One-Class Classification Combined With Spectral Analysis in Food Authenticity Identification

In recent years, counterfeit and substandard food products have become an increasing concern to consumers, and food authenticity assessment is a powerful tool to address this problem and protect public health. Under the high requirements of equipment and sample processing, modern detection technologies usually require a lot of time and money cost consumption. However, as food adulteration methods change and become more sophisticated, traditional modern food quality detection technologies have certain limitations. Therefore, it is necessary to seek new detection technology to effectively promote the efficiency and improvement of food safety quality control and provide strong scientific and technological support and protection for regulatory work-spectroscopic analysis technology, which has been used extensively in recent years for its simplicity and rapidity. As an indirect analysis technique, it needs to be combined with classification methods in data statistics to establish models and achieve rapid analysis requirements. Commonly used classification methods are ineffective in the face of the enormous variety of adulteration types in real life and the considerable variation in the number of true and false samples. One-class classification is a method that models and analyses only one class of instances, fixing the boundaries of the target sample class at a specific confidence level for classification and then using the edges of the target sample to predict the class of the new sample, distinguishing it from all other possible objects. Using this feature to effectively differentiate between samples that are different from the actual data, significantly reducing the detection effort, and has some potential for development in food adulteration detection applications. This paper reviewed the one-class classification method, which has been used in pattern recognition in recent years and described the need for spectral analysis combined with classification methods for food adulteration. The classification results of traditional -and one-class classification methods were compared in the same scenario, and the latter’s characteristics were briefly introduced. Then, several common one-class classification methods were highlighted, such as data-driven class comparison soft independent modelling (DD-SIMCA), one-class partial least squares (OCPLS), and one-class support vector machines (OCSVM), and one-class random forests (OCRF). The applications of one-class classification methods in the food authenticity identification were also discussed, specifically in edible oils, dairy products, beverages, herbs, spices, and agricultural products. At last, the problems of the current one-class classification were analyzed, and the prospects for applying the technique were outlined. This paper is expected to provide some theoretical basis for food certification analysis.