Systems and methods for unsupervised cyberbullying detection via time-informed Gaussian mixture model

Patent metadata

Field	Value
Publication number	US-11916866-B2
Application number	US-202117546648-A
Country	US
Kind code	B2
Filing date	Dec 9, 2021
Priority date	Dec 9, 2020
Publication date	Feb 27, 2024
Grant date	Feb 27, 2024

Abstract

A computer-implemented framework and/or system for cyberbullying detection is disclosed. The system includes two main components: (1) A representation learning network that encodes the social media session by exploiting multi-modal features, e.g., text, network, and time; and (2) a multi-task learning network that simultaneously fits the comment inter-arrival times and estimates the bullying likelihood based on a Gaussian Mixture Model. The system jointly optimizes the parameters of both components to overcome the shortcomings of decoupled training. The system includes an unsupervised cyberbullying detection model that not only experimentally outperforms the state-of-the-art unsupervised models, but also achieves competitive performance compared to supervised models.

First claim

What is claimed is:

1. A framework for identification of cyber-bullying on social media sites, comprising:
a model trained according to a representation learning framework configured to:
construct one or more multi-modal representations of one or more social media sessions; and

a learning network configured to:
estimate a likelihood of bullying associated with each of the one or more social media sessions using the one or more multi-modal representations; and
predicting a time interval between one or more comments of the one or more social media sessions;

wherein a graph reconstruction error determined by the representation learning framework, an energy estimation loss determined by the learning network, and a time interval prediction error determined by the learning network are used to determine a total loss associated with bullying identification; and
wherein the total loss associated with bullying identification is used to optimize the representation learning framework and the learning network,
wherein the model is configured for unsupervised cyberbullying detection and incorporates inter-arrival times of a social media session to leverage temporal dynamics associated with repeated acts of aggression over time.

2. The framework of claim 1 , wherein the representation learning network comprises:
a graph auto-encoder configured to embed user attributes associated with each of the one or more social media sessions as low-dimensional vectors representative of a social network structure of each of the one or more social media sessions.

3. The framework of claim 2 , wherein the graph-auto encoder is implemented using one or more neural networks.

4. The framework of claim 1 , wherein the representation learning network comprises:
a hierarchical attention network configured to generate a textual representation of each of the one or more social media sessions by modeling a sequence of words and a sequence of comments for each of the one or more social media sessions.

5. The framework of claim 4 , wherein the hierarchical attention network captures long-term contextual dependencies between the sequence of words and the sequence of comments for each of the one or more social media sessions.

6. The framework of claim 1 , wherein the learning framework comprises:
a Gaussian mixture model-based density estimator configured to infer a probability density function associated with likelihood estimation of bullying in the one or more social media sessions.

7. The framework of claim 6 , wherein the Gaussian mixture model-based density estimator uses one or more user attributes and one or more textual attributes determined by the representation learning network to estimate the likelihood estimation of bullying in the one or more social media sessions.

8. A processor adapted for cyberbullying detection, the processor configured to:
implement a model configured via a representation learning network that constructs multi-modal representations of social media sessions; and
implement a multi-task learning network that simultaneously with the representation learning network estimates a likelihood of input samples and predicts time intervals between comments associated with the social media sessions,
wherein implementation of the representation learning network and the multi-task learning network outputs,
wherein the model is configured for unsupervised cyberbullying detection and incorporates inter-arrival times of a social media session to leverage temporal dynamics associated with repeated acts of aggression over time.

9. The processor of claim 8 , being further configured to:
combine, by the representation learning network a user representation in a graph auto-encoder and social representation in a hierarchical attention network to form a sessions representation.

10. A tangible, non-transitory, computer-readable media having instructions encoded thereon, such that a processor implementing the instructions, is operable to:
implement a system for unsupervised cyberbullying detection via time-informed Gaussian Mixture Model (UCD) that predicts bullying instances without labeled data, the system incorporating comment inter-arrival times of a social media session which accommodates classification of cyberbullying instances using a full commenting history,
wherein the system includes a representation learning network that learns a compact multi-modal representation of the social media session and a multi-task learning network that predicts whether or not the social media session contains bullying behaviors while modeling temporal dynamics of all social media comments, and
wherein the representation learning network models social media sessions using a hierarchical attention network (HAN) for textual features of a plurality of features and a graph auto-encoder for user and network features of the plurality of features, and the multi-modal task learning network takes the plurality of features as input to estimate a likelihood of bullying using a time-informed Gaussian Mixture Model (GMM).

Assignees

• Univ Arizona State

Inventors

• Cheng Lu
• Shu Kai
• Wu Siqi
• Silva Yasin
• Hall Deborah
• Liu Huan

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• H04L51/52 Primary

  for supporting social networking services · CPC title

• G06F40/30

  Semantic analysis · CPC title