A study on estimating the usefulness level of consumer reviews: Comparison of machine learning algorithms

With the development of Web 2.0, information exchange among consumers on online sharing platforms has become widespread, and the limit of information exchange between consumers has almost disappeared. While consumers could only exchange information and ideas with people in their own cities and neighborhoods before Web 2.0 and technologies, this type of communication slowly moved to internet platforms and revealed the concept of Electronic Word of Mouth Communication. In this way, consumers can exchange information about a product or service not only with consumers in their own neighborhood but also all over the world. With the developing technologies and the development of consumers' internet usage habits, it becomes difficult to reach information that will benefit the consumer. While online shopping platforms offer a variety of methods for identifying and highlighting useful information, determining whether a consumer review and evaluation is useful depends on the approval of other consumers. Developments in the field of machine learning enable the transfer of another job to machines day by day. In this way, repetitive work or control processes can proceed automatically with little or no human intervention. Supervised learning, which is a branch of machine learning, can be used to label or classify entities with unknown labels or classes by learning the attributes of data groups with certain labels or classes. The first part of the study includes general information about word of mouth, Web 1.0 and Web 2.0, useful consumer evaluation. The second part includes information about what the algorithm is, what the data is, data preprocessing processes, machine learning, supervised learning and its sub-titles, unsupervised learning, reinforcement learning and testing machine learning algorithms. The third part of the study includes the research model, data acquisition processes, data preprocessing, feature extraction, training and testing of models. In the study, 6 different supervised machine learning algorithms were trained with the attributes obtained from consumer evaluations written on a product in the online shopping platform, and it was tried to predict whether a new user evaluation would be marked as useful by consumers. Algorithms were tested for accuracy, area under the ROC curve and F1 scores, and were also externally tested with consumer reviews from different product categories. It is expected that this study will encourage consumers to prepare a new consumer assessment with the knowledge that a consumer assessment that will be beneficial to other consumers can be determined without the approval of other consumers.