نوع مقاله : پژوهشی - کاربردی
نویسندگان
1 گروه طراحی شهری، دانشکده معماری و شهرسازی، دانشگاه هنر ایران، تهران، ایران
2 گروه طراحی شهری، دانشکده معماری و شهرسازی، دانشگاه هنر ایران، تهران
چکیده
کلیدواژهها
عنوان مقاله [English]
نویسندگان [English]
ABSTRACT
Among the types of urban spaces, urban green spaces and parks, as city breathing spaces, are lush and relaxing areas that have been selected as the case of this research paper. Therefore, this article aims to analyze the emotions of users of Mellat Park in Tehran in the form of analytical research based on a quantitative method (supervised machine learning approach and lexical-based. After preprocessing and labeling, the data were examined and analyzed using two methods, such as model-oriented and non-model-oriented. Emotions were also examined and analyzed using the Python programming language. The comparison of these two methods revealed that among the machine learning algorithms, XGBoost has the highest accuracy at 87%, while K-nearest neighbors and support vector machines have lower accuracy but are still capable of predicting emotions in green spaces. The lexical method (using the VADER dictionary) has a lower predictive ability compared to machine learning. Finally, the stacking ensemble learning model, which was used to increase the accuracy of the model, has the ability to predict emotions based on the results of the confusion matrix (96%). Therefore, using the method based on virtual space data, it is possible to predict the emotions of users of other urban green spaces with high speed and accuracy in Tehran.
Extended Abstract
Introduction
Increasing the quality of urban parks is effective in providing the psychological and emotional comfort of users. Sentiment analysis is an approach to measure the level of psychological comfort and emotional response of users to spaces.
Methodology
The research method in the case study is based on the quantitative method. The following presents the selected social virtual space, the time period of data extraction, and the methods used in both methods for data preparation and analysis. It was found that the Google Map social network is suitable for this study. After the pre-processing, in order to train and test the data, the data is divided into 80% training data and 20% testing data (for evaluating the models). Then, the data is labeled to train machine learning models using the points given by the users to the location. In order to select and allocate the appropriate model or algorithm, according to the subject under investigation and the need to classify emotions and achieve a predictive model, the models under the supervision of machine learning were used. The VADER method was used in the non-model method (classification of feelings by the dictionary method). At the same time, the Blending Ensemble model from the stacking family was also used in this research.
Results and discussion
According to the first method, among the implemented models, the XG Boost algorithm correctly recognized 87% of all messages and was the best algorithm. After that, the k nearest neighbor algorithm correctly recognized about 80% of all messages, support vector machine, 70%, and simple Bayes and linear discriminate analysis about 66% and 57% of all messages. Other models performed less than 50%, and the random forest model performed worse than all algorithms. The results of the second method showed that 22% of the data in the category of negative data were correctly identified, 33% of the neutral data were false, and 44% were identified as false positives. In the category of neutral data, 7% of false negative data, 20% of true neutral data, and 73% of false positive data have been detected. In positive data, 3% of false negative data, 15% of false neutral data, and 82% of positive data are correctly identified. As a result, in this method, the highest correct percentage associated with positive data is 82% correct detection. As a result, it had the best performance in detecting positive data, but it did not perform well in detecting negative and neutral data. In order to provide a predictive model with higher accuracy, the Blending Ensemble models and the stacking family, which is included in the Ensemble Learning model category, were used. The result of applying the proposed blending model to the test data in the confusion matrix shows that in the category of negative data, 80% of the data are correct, 10% of the data are neutral and false, and 10% have identified false positives from the data. In the category of neutral data, 0% of false negative data, 84% of true neutral data, and 16% of neutral data were detected as false positives. In the positive data section, 0% of false negative data, 3% of false neutral data, and 97% of true positive data have been identified. As a result, in general, in the proposed blending model, the best performance has been in positive data, which has been correctly recognized in 97%, and the most error in negative data has been in general with 20% false neutrals and false positives, which compared to other algorithms have a small percentage of errors and have performed well in all 3 categories of positive, neutral and negative. In relation to the evaluation criteria of this model, a percentage above 95% can be seen in all the evaluation criteria, which indicates the good performance of the model.
Conclusion
The results of the research showed that the model-oriented methods worked better than the word-based method, and the blending method was better than the machine learning algorithms. Therefore, the algorithm trained with the blending method has the ability to predict urban feelings in the park with a high probability. The chosen method has many applications in the field of urban planning and urban design because it provides the possibility of identifying citizens' feelings about the environment due to its low cost compared to field research and the speed of data analysis.
Funding
There is no funding support.
Authors’Contribution
Authors contributed equally to the conceptualization and writing of the article. All of the authors approved thecontent of the manuscript and agreed on all aspects of the work declaration of competing interest none.
Conflict of Interest
Authors declared no conflict of interest.
Acknowledgments
We are grateful to all the scientific consultants of this paper.
کلیدواژهها [English]