Combining Correlation-Based Feature and Machine Learning for Sensory Evaluation of Saigon Beer

Introduction

Food sensory evaluation (FSE) has an important effect on the food and consumer goods industry. The role of sensory evaluation (SE) is widely recognized through the important tasks it performs in the development of new products, basic research, materials and technology improvements, and reduced production costs, manufacturing, quality assurance, and product optimization (Sidel & Stone, 1993). The result of SE is considered as a way to reduce the risk and uncertainty in final decision making.

A key step in the brewing factory and quality control during beer production is the assessment of beer quality with the support of SE (Dong et al., 2014). SE uses a set of techniques to measure human responses. Its target is to minimize the bias caused by potential confounding sources, which may affect consumer perception (Mukhopadhyay, Majumdar, Goswami, & Mishra, 2013).

SE is an essential step in scientific experimentation that can generate useful and accurate data for analysis and allow for an efficient experimental process. It can be implemented by using an experimental design. Following this, mathematical models are used to interpret the relationship between input and output variables. The input variables may contain food ingredients, physical and sensory properties, while the output variables may contain consumer preference and sensory characteristics. These models highlight any interactions and possible correlations existing in variables or factors. Thus, they provide more objective performance than human panelists, who are likely to have subjective responses (Seisonen, Vene, & Koppel, 2016).

However, FSE still remains some challenges. Particularly, it requires a lot of experimental design if data contains many independent variables, which can lead to time consuming, sensory fatigue among panel members and high financial resources. Moreover, sensory qualities and physiochemical attributes in foods may not be linear in nature. Therefore, factorial designs or response surface methodologies in symmetrical designs may not produce the most ideal experimental design. In addition, the constraints related to the nature of a sensory experiment may be different with the levels of independent variables in the experiment. Although mixed levels factorial designs can be used for this purpose, they may result in non-orthogonal and unbalanced designs (Yu, Low, & Zhou, 2018).

Data mining techniques has been used to extract useful information and relationships between factors from raw datasets. In the literature, several one-dimensional statistical methods such as correlation tests and analysis of variance (ANOVA) are applied to find the trends and patterns inside the data. However, univariate regression methods such as one-factor ANOVA may not sufficient to extract patterns in food physicochemical attributes, hedonic properties and sensory profiles, that contain a large number of compounds and physical attributes present in food products (Zielinski et al., 2014).

In FSE, masking and synergistic effects may generate a nonlinear relationship between taste and odor properties of foods (Noble & Ebeler, 2002). In addition, a single product property such as flavor or texture is related to multiple sensory attributes, as perceived by the human brain. This combination provides a highly complex relationship, which cannot be easily analyzed by univariate methods. Therefore, such analyses need to be solved by other methods such as multivariate statistical techniques.

Computer science and mathematical methods have been widely applied in a variety of domains ranging from natural and social sciences to engineering and health care. Machine learning techniques include tasks for pattern recognition, classification, clustering and regression. These methods provide many promising alternatives and their effectiveness has been proven in a wide range of studies such as ester production during fermentation in beer (Dong et al., 2014) and correlating sensory profiles, correlating sensory properties to chemical components in white wine (Liu et al., 2015), red and white wine discrimination (Lozano, Santos, & Carmen Horrillo, 2008), signal calibration in chemometrics, pattern mining (Dinh, Le, Fournier-Viger, & Huynh, 2018; Le, Dinh, Huynh, Nguyen, & Fournier-Viger, 2018; Quang, Dinh, Huynh, & Le, 2016), and Sake wine data clustering (Dinh, Fujinami, & Huynh, 2019; Dinh & Huynh, 2020; Nguyen, Dinh, Sriboonchitta, & Huynh, 2019).