Being depression a mental disorder that a ects 1 in 10 people world-wide, it is important to react in an e ective way before fatal decisions could be taken. Moreover, with the rise of social media, some issues regarding how to identify patterns of depressed users in a timely fashion is an important task that is attracting the attention from the NLP community. Accordingly, the CLEF 2017 launched a challenge to identify depressed users in reddit forums. Our proposal to attend this task was based on a two-step classi cation procedure, where we rst look at the post level to create basic features that were then applied at user level, to build a pro le for each user. To evaluate this methodology, we applied a 10-fold cross validation on the training data, obtaining an F-score of 0.73 in the identi cation of depressed users. For the CLEF challenge, our team reached the fourth place, obtaining an F-score of 0.48.
According to the National Institute of Mental Health (NIMH), 1 in 10 adults will exhibit some kind of depression at some point in life [15]. People a ected with depression su er from feelings of hopelessness, loss of motivation, have sleep and eating disorders, and feel disconnected from other people. If untreated, depression can lead to suicide. Psychological criteria for detection of depression is presented in the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition (DSM-5), and consists of long-term (two or more years) behavioral analysis performed by a professional. This requires individuals to get diagnosed by a psychologist, something that many people are not willing or capable to do.
Online social platforms are quickly arising in popularity, becoming nearly
ubiquitous in countries where internet is readily-available. These platforms allow
people to share and express their thoughts and feelings freely and publicly with
other people. As could be supposed, this information is a rich source for Natural
Language Processing tasks such as: Opinion Mining [
The proposed approach is based on a two-step classi cation procedure; the rst step focuses on the analysis of individual posts by their content, whereas the second step considers the classi cation of users by their behaviors expressed by their kind of posts. The main idea behind our approach consists in modeling each user post, instead of directly modeling each user, under the assumption that each post is a window into the mind of a person in a particular point in time. Then, if we collect enough of these posts, we can observe how that person's mind changes - or do not - through time, in order to improve the model of each user. In our rst set of experiments, we applied a 10-fold cross validation on the training data, obtaining a F1 score of 0.73 for predicting depressed individuals and above 0.90 for non-depressed individuals. For the challenge task, we ended up in the fourth place (out of eight teams) with an F-score of 0.48. Since our participation in the challenge was late, we only submitted predictions for the last part of the competition, so we would like to further analyze our model simulating their application at di erent levels of evidence for each user.
The organization of this paper is as follows. Section 2 presents some related work. In section 3 we describe our approach. Experimental Settings are presented in section 4, while results are presented in section 5. Finally, Conclusion is presented in section 6. 2
Social media platforms have become an attractive place to infer health related
issues for entire communities or individual users. A study presented in [
This section presents an overview of the proposed two-step classi cation approach. The rst step considers the analysis of individual posts by their content. Its aim is to discriminate messages produced by depressed and non-depressed users. Then, a second step analyzes this information with the aim of modeling the respective users' behaviors based on their category. The following explains in detail these two steps. 3.1
The main goal of the rst step is to classify each post as likely produced by a depressed or non-depressed person. For this, we train a classi er using user posts considering that all posts inherit the label from their respective author. More formally, let D = f(D1; y1); : : : ; (Dn; yn)g be a training set of labeled userdocuments, that is, D is a collection of n tuples of user-documents, i.e. posts, (Di) and category-labels (yi), where yi 2 fdepressed; non-depressedg. Also let Di = f(P1; yi); : : : ; (Pn; yi)g be the set of posts from user Di. We represented each post Pj by a feature vector Pj that combines three di erent kinds of features: unigrams, bigrams and trigrams. For each individual feature space we considered the 10k top frequent items and then we selected all the features with information gain greater than zero. Additionally, we included two extra time attributes: the hour of the post and a binary attribute for whether or not it was posted on a weekend. Therefore, the nal post representation Pj is expressed as follows:
Pj = hPj1gram; Pj2gram; Pj3gram; Pjtime; Pjweekendi (1) 3.2
The intuitive idea behind analyzing posts individually is that users can be understood in terms of their behavior, rather than by their vocabulary. Hence, in this second step we build the user representation by considering some statistics obtained from the rst classi cation step.
Given a document Di, we rst classify all its posts using the previous classi er (refer to Section 3.1). Then, we model Di as the sequence of predicted labels for each post Pj 2 Di, that is, Di0 = (y1; : : : ; yh), where yi corresponds to the predicted label of post Pi. From this sequence of labels, we represent each user by a 12-dimensional vector Di including the features described below. For simplicity we will refer to a post as being "depressive" (or "non-depressive") depending if it was generated by a depressed or non-depressed user. 1. Percentage of posts that were classi ed as \depressive". 2. Percentage of posts that were classi ed as \non-depressive". 3. Percentage of times that a \non-depressive" post is followed by another \non-depressive" post. 4. Percentage of times that a \non-depressive" post is followed by a "depressive" post. 5. Percentage of times that a \depressive" post is followed by a \non-depressive" post. 6. Percentage of times that a \depressive" post is followed by another \depressive" post. 7. Percentage of \depressive" posts written in the morning. 8. Percentage of \non-depressive" posts written in the morning. 9. Percentage of \depressive" posts written in the evening. 10. Percentage of \non-depressive" posts written in the evening. 11. Percentage of \depressive" posts written in the night. 12. Percentage of \non-depressive" posts written in the night.
Time periods were considered after analyzing heatmaps of users' activity, See Figure 1. Accordingly, the periods that we consider were morning [6am-2pm], ii) evening [2pm-10pm], and iii) night [10pm-6am]. 4
Dataset. For the evaluation we use the dataset presented in [14]. This dataset is in XML format, where each user had a list of posts, and each post contained the following elds: title, date, info, and text. The info eld simply contained the string \reddit post" and was discarded for not providing any relevant information. Title and text elds both contained useful information; they were merged into a \text" eld.
Preprocessing. Uppercase letters were turned to lowercase, numbers were removed, and negations were joined to the following word (i.e., \don't want" returns a single token \don't want").
Classi cation. For post and user classi cation we used a Nave-Bayes Classi er (NBC). For validation we applied a strati ed 10 fold-cross validation.
To evaluate the appropriateness of the proposed approach to identify depressed users from online forums, we compared its performance against a traditional text classi cation approach using a BOW representation. The rst part of Table 1 shows the results from this comparison; they indicate that the two-step classi cation approach outperformed by more than 40% the baseline results. In order to evaluate the usefulness of some components of the proposed approach we carried out some experiments disabling some of them. In particular, we ran an experiment without considering the time information for the post representation and other using only word unigrams. The last two rows from Table 1 present the results from these two experiments; they clearly show that word sequences and time information are key elements in the classi cation of depressed posts and therefore on the identi cation of depressed users.
An analysis on the relevance of the features showed us that the most discriminative n-grams were those related to the health condition of the users such as depression, diagnosis and therapy; this result was somehow expected due to the nature of the corpus. However, other n-grams related to feelings, news and interpersonal relationships were also highly discriminative for both kind of users. Table 2 shows the 40 most discriminative word n-grams. depression, pope, to talk to, my depression, miserable, conservative depression , global, depresesion and, therapist, gop, obama
depression ., anxiety, suicidal, meds, $ billion, not alone diagnosed, my life, myself, depressed, when i was, conspiracy diagnosed with, because i have, boyfriend, feel like i
hollywood, iraq, like i, report, president, cnn makes me feel, of my life, isis, me feel, vote, helped me
We also analyze the posting behavior of the users. The heatmaps in Figure 1 show that there is a clear di erence in the posting behaviors of both kinds of users, particularly on weekends. We expected depressed users to be more active during night, as people with depression also su ers from insomnia or has bad sleeping habits overall, but evidence showed us that it was the contrary, not depressed users were more active overall, except on weekends. In this paper we presented our approach to identify depression through online publications, which is based on a two step analysis. The analysis of individual posts to characterize users' behavior was very useful in a second step to analyze the users from higher point of view. In this work, the general behavior of each user is captured by means of observing the posting history, especially the posting behavior and the posting period of time.
We presented experimental results of our method that shows to be useful compared with other methodologies. For this, we study the user-documents from di erent perspectives, including: post, and analysis of period of time. We found that both kinds of users can be di erentiated by exploiting the time of publication, but more importantly, the changes in their behavior related to depressed and non-depressed posts. It was also found that both users have di erent activity during the weekend and during the day, which is valuable for future research paths.
Future work includes developing an early risk assessment strategy. We could not participate until late for this contest, and thus we did not develop an strategy for early risk assessment.
Acknowledgements. This work was supported by CONACYT under scholarship 735623/599448 and project grant PDCPN-2014-01-247870. 12. Park S., Lee S. W., Kwak J., Cha M., Jeong B. (2013). Activities on Facebook reveal the depressive state of users. J. Med. Int. Res. 15:e217 10.2196/jmir.2718 13. De Choudhury M, Counts S, Horvitz EJ, Ho A. Characterizing and predicting postpartum depression from shared Facebook data. 2014 Presented at: ACM Computer Supported Collaborative Work 2014; February 15-19, 2014; Baltimore, MD p. 626-638. 14. Losada, David E., and Fabio Crestani. "A Test Collection for Research on Depression and Language Use." CLEF. 2016. 15. National Institute of Mental Health. Depression Research, 1999. 16. Michalek, R., Tarantello, G.: Subharmonic solutions with prescribed minimal period for nonautonomous Hamiltonian systems. J. Di . Eq. 72, 28{55 (1988) 17. Tarantello, G.: Subharmonic solutions for Hamiltonian systems via a ZZp pseudoindex theory. Annali di Matematica Pura (to appear) 18. Rabinowitz, P.: On subharmonic solutions of a Hamiltonian system. Comm. Pure Appl. Math. 33, 609{633 (1980)