Table 1.
Sample of depressive-indicative phrases collected from tweets.
Fig 1.
The age distribution for depressed and control users in ground-truth dataset.
Fig 2.
Gender and depressive behavior association (Chi-square test: Color-code: (blue:Association), (red: Repulsion), size: Amount of each cell’s contribution).
Table 2.
Facial presence comparison in profile/posted images for depressed and control users—*** alpha = 0.05.
Table 3.
Statistics of processed shared/profile images.
Fig 3.
The Pearson correlation between the average emotions derived from facial expressions through the shared images and emotions from textual content for depressed-(a) and control users-(b).
Pairs without statistically significant correlation are crossed (p-value <0.05).
Table 4.
Statistical significance (t-statistic) of the mean of salient features for both depressed and control classes—** alpha = 0.05, *** alpha = 0.05/223.
Table 5.
Statistical significance test of linguistic patterns/visual attributes for different age groups with one-way ANOVA, *** alpha = 0.001, ** alpha = 0.01.
Fig 4.
Characterizing linguistic patterns in two aspects: Depressive-behavior and age distribution.
Table 6.
Age Prediction performance from visual and textual content for different age group(years old).
Table 7.
Facial presentation distribution for different age group(in years old) in profile and media.
Table 8.
Gender prediction performance through visual and textual content.
Fig 5.
Ranking features obtained from different modalities with an ensemble algorithm.
Fig 6.
The explanation of the log-odds prediction of outcome (0.31) for a sample user (y-axis shows the outcome probability (depressed or control), the bar labels indicate the log-odds impact of each feature).
Table 9.
Model’s performance for depressed user identification in Twitter using different data modalities.
Fig 7.
Word usage difference of likely vulnerable individuals versus random profiles.