The Effect of Channel Size on Performance of 1D CNN Architecture for Automatic Detection of Self-Reported COVID-19 Symptoms on Twitter

Social media has a crucial role as the most generally accessed source of information by the public for obtaining information on various topics, such as COVID-19 and natural disasters. Using Twitter messages can be beneficial in tasks like detecting symptoms of COVID-19. The results of COVID-19 symptom detection can be used as a benchmark by relevant parties to create policies that can be useful in the future. However, social media data contains various topics, requiring various classification approaches to obtain the most accurate results. The classification was done using variations of word embedding (Word2vec, fastText, and a combination of Word2vec and fastText) and variations of CNN architecture (single-channel and multi-channel). The best accuracy result was obtained by combining fastText and fourth channels, which was 88.5%. The comparison results showed both the single-channel and fourth-channel architectures achieved the highest average accuracy of 86.33%.