Research On The Application Of An Integrated Evaluation Method Based On Siamese Neural Network

Evaluations are frequent,significant cognitive activities in human society.People make comprehensive judgments and rank the evaluated objects based on multiple factors,including relevant information from them and alternative data.The obtained integrated evaluation results are not only convenient for people to recognize the rankings of evaluation objects and compare the advantages and disadvantages of different objects,but also can provide great reference significance for subsequent selection and decision-making,whcih is a classic research problem in statistics.Existing integrated evaluation methods assign weights based on expert opinions or data dispersion and provide evaluated rankings of objects through simple mapping.These methods are concise and easy to implement,and allow intuitive comparison of the importance of different indicators by weights.However,the subjective methods are too labor-intensive,and the objective methods may produce biased results when the differences of all features in the evaluated objects are not significant enough.Therefore,in this study,in order to reduce such errors,we borrow the idea of the ranking method and train the model by combining the attributes of the indicators and the apparent ordering relations between samples pairs.We re-formulate the evaluation problem in mathematics,construct ranking model,then design a new method named Ranking by Siamese Neural Network(RankSNN)adaptive to the model.RankSNN method uses sample pairs as learning instances,and the ordering relations of sample pairs can greatly reduce the labor cost,and then the Siamese neural network structure is trained to measure the difference between the samples,and the greater the difference between the unknown and known samples,the greater the difference in evaluation scores,and the rankings can be obtained by this way.The deep neural network is then used to fit the scoring function,and its adaptive weight learning mechanism can make the trained scoring function more objective and accurate,and it can accurately correspond to the ordering relations amongst the sample pairs by calculating evaluation values and obtaining the best evaluation ranking results.We first conducted a simulation experiment on on the public dataset of learning to rank.After comparing with the commonly used integrated evaluation methods and baseline ranking methods,the results of RankSNN showed obvious superiority.Then,we conducted empirical analysis of the application of the integrated evaluation problem examples in two practical scenarios of the business environment of Chinese cities and the risk of urban transmission of COVID-2019.For each instance,we first constructed the evaluation indicator system for the corresponding problem,collected and preprocessed the data,and then conducted model training experiments using the RankSNN method,divided the training and test sets to verify and compare the model effects.And then we compared the results with the commonly used comprehensive evaluation methods.The RankSNN method can reach 100%in the prediction of the accuracy of the sample order relationship,,the ranking quality also reached the best,which shows the excellent performance of our method and the reliability of the ranking results.