Select Page

Our work was a problem of machine learning regression. We established 7 regression equations for machine learning to predict in each message the average of the Judgments of the Turks on the amount of emotional or informative support, question-asking or self-revelation (see the dimensions of the judgment listed in Table 5). The predictor variables were the dictionaries and other features listed in text box 1 and the Story function. We used Weka, a machine learning toolkit, to create the Support Vector Machine (SMOreg) regression models [35]. The 1000 thread launchers or their first answers, coded by MTurk workers, were randomly divided into a training game (80%), a development rate (10%) and a test game (10%). The training kit was used to build the models. The SDK was used to evaluate the accuracy of different model configurations and variations in the functions used. After the models performed well in terms of development data, we used the test to evaluate the proper functioning of the final regression equations. We evaluated the predictions from the pearson torque correlation between human-coded assessments and machine measurements for the 100 messages in the sample test. The correspondence between human-coded assessments and machine measurements was .65 average over the 7 dimensions and ranged from .85 for providing information assistance to .44 for positive emotional self-revelation. Table 5 presents the evaluation results for each support-related construction. Most of the results from this large sample of machine-coded data replicated the results of the small sample of human-encoded data. The model showed that the more one of the 4 types of self-opening presented itself in the thread start message, the more emotional support the initial response contained (all P<.001), although the effect of positive emotional self-revelation was not significant in the human-encoded data model (P = .10).

The effect of self-opening negative information (beta = 0.18) was the effect of self-opening of negative information, especially in the context of emotional support, SE 0.01) stronger than the effect of positive emotional self-opening (beta = 0.09, SE 0.01) or negative emotional self-opening (beta = 0.10, SE 0.01), which were in turn stronger than the effect of positive self-opening of information (beta = 0.06, SE 0.01). However, these comparisons of effect effectiveness should be treated with caution, as some constructs in Table 6 have high correlations (e.g.B. the correlation between positive and positive self-disclosure of information was .83). As with the human-coded recording, when the thread launcher asked questions, the answer contained less emotional support (beta =-.17, SE 0.01), but more information support (beta=.29, SE 0.01). Positive self-presentation of information appeared to weaken information assistance (beta=-.06, SE<0.00). Self-presentation of negative information appeared to attract support for information (beta=.07, SE<00.00), although this association was not significant in the human-encoded dataset, likely due to the small sample size. . . .