Maryam Mohammadi
Abstract
Extended Abstract
Background and Objectives: The examination of users’ emotions through social media has developed into an impactful domain across diverse scientific fields, ...
Read More
Extended Abstract
Background and Objectives: The examination of users’ emotions through social media has developed into an impactful domain across diverse scientific fields, appealing not only to business proprietors and politicians but also to general users. In the meantime, this field has infiltrated urban studies and has been used by urban planners and designers due to its methodology; whether in the form of research that aims solely at emotion analysis or as an integrated layer within broader research endeavors. The aim of this article is to explain this field in the analysis of urban emotions as modeling methods in order to identify the position of this field in urban studies by examining the importance of emotion and the methods of its study in the city.Methods: This research used the supervised machine learning approach and analyzed the sentiments of tweets related to eight major cities in Iran. The data collection consists of 930 tweets that were collected in a period of 10 years from 2011 to 2022. Initially, over 5000 tweets were collected, and during the tagging process, 80% of them were excluded due to their limited relevance to the city, emphasizing tweets related to urban space. The name of cities and tourist areas were searched to establish a balance between positive and negative data. The tweets are downloaded through Twitter streaming API and the metadata along with the text, including the number of retweets, number of likes and tweet ID, language and location. The data sets have been used for machine training after standard and normalization steps. In this research, the ratio of training data to testing data is 80 to 20. According to the supervised approach, the data were labeled by the researcher with three negative, neutral, and positive labels, and where the researcher had doubts, the opinions of two other experts were used. In general, both machine learning and deep learning have been used. In order to check the validity of the model and to test it, the confusion matrix has been used.Findings: Firstly, the machine was trained based on 3 algorithms that were used in many research related to text sentiment analysis. Based on the test results presented on the confusion matrix, the accuracy of the trained machine in determining the polarity of the text in three polarities was defined. Among the three used algorithms, support vector machine and random forest have performed better than other algorithms. Given that the model’s highest accuracy was approximately 70%, deep learning was employed to train the machine in order to assess the potential for achieving improved results. In the following, machine learning with a convolutional neural network algorithm and a hybrid algorithm were considered. At first, the machine was trained using a convolutional neural network. The results of the accuracy of the model showed that the model is predictable by up to 75%. Next, an attempt was made to improve the predictive accuracy of the model by writing a hybrid algorithm based on the convolutional neural network. The architecture of this network is such that two types of data are considered as input to the neural network, text data and other features in the data set, including location, number of retweets, number of likes, city codes and searched content (as metadata). Therefore, based on this input and output (classification based on the polarity of the text by the researcher), the machine was trained and finally tested. As depicted in the structure of the hybrid algorithm, the significance of the text is assigned a weight of 90%, while the importance of metadata is weighted 10%. It should be noted that different percentages were given to the importance of each of the inputs and the predictability accuracy of the model was checked. As the model test results show, the designed algorithm has improved the predictability of the machine by 4%.Conclusion: In this article, sentiment analysis based on model-oriented methods - machine learning and deep learning - was scrutinized, and therefore, while comparing it with traditional methods and lexical methods, the process of urban sentiment analysis was developed and the different levels of the process were described in detail. As stated, these methods have many advantages and can be useful for analyzing the current situation or predicting different urban projects. Besides, compared to traditional methods, they are less expensive, faster and have sufficient accuracy. According to the appropriate capability of the trained machine, this machine can predict the polarity of the data. This means that by using the text data published in social networks, it is possible to analyze the feelings of users. Certainly, when these data are geodatabases, there is also the capability to geolocate emotions. This approach allows for a swift, accurate, and cost-effective general assessment of city spaces. By identifying areas where users perceive negative emotions, the reasons can be investigated and addressed accordingly. This research has been innovative in two aspects, 1) preparing a collection of data related to the sentiment of Persian language users related to the city and 2) analyzing urban sentiment in the country using machine learning in the field of urban planning and design. Some limitations of this research include limited access to all data published on Twitter using Twitter streaming API; the small amount of available and relevant data; low use of Twitter by users due to filtering; and the unavailability of financial resources to prepare and use a larger set of data.
