Machine Learning based Dataset for Finding Suicidal Ideation on Twitter

摘要

Suicidal ideation is a major health issue nowadays. This may lead to death of various people. Suicide is also one of the major reason of death in many of the countries [9] [14]. Automatically finding people having suicidal ideation on social media is a major concern and a lot of people are working in this direction [10] [11]. There are many risk factor associated with suicidal ideation such as anxiety, depression, mental disorder etc. [13] [15]. A number of methods have been made to prevent deaths because of suicide. With the advent of social networking site, people have started expressing their feelings more on social media rather than someone in personal [6] [12]. Text classification has proven to be a successful method to prevent suicides [8]. This article describes a dataset of people having suicidal ideation on twitter. The data was extracted from an Application Programming Interface provided by Twitter. Various features/keywords related to suicidal ideation shown in table 2 were used to identify persons having such ideation. These keywords have been gathered from various web forums and previous year papers [7]. Initially the dataset have been taken from Twitter public application programming interface using its access key and access token. The raw data comprises of various fields such as: user__id, user__name, created_at, text, user__screen_name, user__friends_count, user__listed_count, user__favourites_count, user__followers_count, user__statuses_count, user__created_at, user_location with around 14202 tweets a part of which is shown in table 3. After that a sample of 1897 tweets were extracted depending upon the keywords selected and merely the text and class fields are set aside as needed to be given as input to any of the algorithms as shown in table 4. The class consists of binary values having either value 0 (non-suicidal) or 1 (suicidal) based on whether the tweet is related to suicidal ideation or not. This is done by a manual annotation by a human annotator and a psychiatric expert as shown in table 4. In the final step the preprocessing of the tweets are done based on the semantics of the keywords recognized and then based on the text fileld new colums are added to the table which contains all the keywords and the table is altered into the probabilistic values i.e. either 0 or 1. Based on the occurrence/non-occurrence of the keyword, a value 0 or 1 is assigned to each keyword and tweet in the particular record. We have given a value 1 if the specific keyword exists in that particular tweet and we have given a value 0 if a keyword doesn’t exist in the particular tweet and hence the resultant dataset consists of only binary (0 or 1) values as given in table 5 [1]. The resultant dataset consists of 1897 tweets and 34 features. A number of machine learning algorithms like Multinomial Naïve Bayes, Bernoulli Naïve Bayes, Logistic Regression, Decision Tree, Support Vector Machine, Random Forest, Voting Ensemble and AdaBoost are then used on this dataset for testing the dataset and finding the accuracy, recall and precision.