understanding and discovering deliberate self ... - yilin wang · yilin wang1 2jiliang tang 1...
TRANSCRIPT
![Page 1: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/1.jpg)
Yilin Wang1 Jiliang Tang2 Jundong Li1 Baoxin Li1 Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5
1Arizona State University 2Michigan State University 3University of Washington 4 Yahoo Research5 Huawei Research
WWW 2017
Understanding and Discovering Deliberate Self-harm Content in
Social Media
![Page 2: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/2.jpg)
Self harm: What is it about ?
#svv
#blithe
#secret society 123Self harm
Self mutilation
#olive
![Page 3: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/3.jpg)
Self harm: How common?
• 2 Million cases reported annually (US) • 2nd leading cause of teenage deaths
(world wide)
• Existing efforts only relied on self and friends/families reports, but most of self harm symptom is very difficult to discover.
• The relatively rare occurrence of completed self-harm treatment and the rare population made the studies expensive to conduct.
![Page 4: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/4.jpg)
Why Social Media?
• Monitoring human behavior• A better place for young adolescentsSocial stigma exists for people who engaged in self harm“I swear to god, I got worse panic attack ever when adults talk about cutting and force you to show the wrist ”
![Page 5: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/5.jpg)
Motivation A few works on social science and psychology researcher [Meleno 2016, Daine] have studied some language usage, and social influences for self harm people.
Limitations• Heavily rely on survey and self reports• relative small scale (hundreds or thousands)
![Page 6: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/6.jpg)
Research Questions
• Is there any distinct characteristics of self-harm content from normal content? (insights ?)
• If so, how can we leverage these characteristic to build models to discover self harm content?
![Page 7: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/7.jpg)
Data Collection Flickr: 10 Billion posts with 50 Million users • Self-harm Content: “#selfharm” “#selfinjury” 1B-> 15,792 posts
Refine 383,614 posts and 63,949 users • Normal User drawn from YFCC 19720
users and 93286 posts for each group
![Page 8: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/8.jpg)
Data Analysis
• Textual Analysis
• User Analysis
• Temporal Analysis
• Visual Content Analysis
![Page 9: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/9.jpg)
Textual Analysis
• self-harm content tends to include more verbs and adjectives/adverbs than nouns which is very consistent with suicidal word usage.
• The poor linguistic structure usage and language suggest the decreased cognitive functioning and coherence.
• A large portion of negative sentiment words are used in self-harm content.
![Page 10: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/10.jpg)
Beyond Text
![Page 11: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/11.jpg)
User Analysis
More Active: average post from create account to last login in.
High proportion of reply and number of favorites indicate that self harm content receives more social response.
Self harm uses has less friends shows that it could be berried by the large portion of normal users.
![Page 12: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/12.jpg)
Temporal Analysis
Normal users:• fewer number is published later in the night and early morning.• the number generally increases through the day (peaks in 3pm )
Such reason could be the mental issues related insomnia.
![Page 13: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/13.jpg)
Visual Content Analysis
saliency
Image Style
Image emotion
patterns
![Page 14: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/14.jpg)
The importance of our finding
• Let self harm post to be heard.
Common feature: visual feature and textural feature (CNN+WE) Our findings: language usage, sentiment and lexicon, temporal, user information and visual patterns
![Page 15: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/15.jpg)
How to utilize the findings?
• supervised Features Labels
Training the classifier
![Page 16: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/16.jpg)
How to utilize the findings? • unsupervised
visual information
Model Learning
textual information Our findings
![Page 17: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/17.jpg)
Experiments Dataset: • Balanced dataset: equal size of self harm content
and normal content. (150k) • Imbalanced dataset : 1:10 with self harm to normal
content. (150k and 850k)
Metric: • Supervised: F1 and precision • Unsupervised: accuracy and NMI Parameter analysis : alpha from 0.0001 to 10
![Page 18: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/18.jpg)
Results of Supervised Method
Textual
Visual
![Page 19: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/19.jpg)
Results of Unsupervised Method
Textual
Visual
![Page 20: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/20.jpg)
Parameter
![Page 21: Understanding and Discovering Deliberate Self ... - Yilin Wang · Yilin Wang1 2Jiliang Tang 1 Jundong Li1 Baoxin Li Yali Wan3 Clayton Mellina4 Neil O’Hare4 Yi Chang5 1Arizona State](https://reader031.vdocuments.us/reader031/viewer/2022041021/5ed0568bfd7bcf77cf7c8199/html5/thumbnails/21.jpg)
Conclusion • Our analysis suggest that the characteristic of
self harm content is very different with normal content.
• Features inspired by our findings improve detection of identify self harm content.
• We can extend our work to a semi supervised learning problem for real-world data.
• We will explore the network influences to self harm users.