Real-Time Social Media Content Filtering via ML Classifiers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Social media users face harassment through unwanted content, and existing solutions rely on manual human evaluations that are inefficient and occur after the fact, failing to create a safe environment due to the lack of automated filtering systems on the receiver side.
Innovation Solution
A system utilizing Machine Learning classifiers to filter out unwanted content, such as harassment, threats, and fake news from incoming social media data in real-time, allowing only neutral content to be displayed while enabling users to access filtered-out material and automate reporting processes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If manual human evaluation is used to filter content, then accuracy in identifying harmful content is maintained, but efficiency and timeliness deteriorate due to manual processing requirements
Solution Approach 1:
The patent replaces manual human evaluation with automated machine learning classifiers that process social media content. The system uses trained ML models to automatically identify and filter harmful content such as harassment, threats, and abuse, eliminating the need for manual intervention while maintaining filtering accuracy.
Solution Approach 2:
The system performs self-service by automatically detecting, classifying, and filtering harmful content without requiring user intervention. The machine learning models independently evaluate incoming social media data and make filtering decisions, allowing the system to operate autonomously at scale.
2Reliability
If content filtering is performed after the fact through removal, then post-hoc correction is achieved, but prevention of harmful content exposure deteriorates
Solution Approach 1:
The patent implements preliminary action by filtering harmful content before it reaches the user's feed. The machine learning classifiers process incoming social media data in real-time and block harmful content prior to display, preventing exposure rather than correcting it after the fact. This includes filtering harassment, threats, abuse, and other harmful content before it can be viewed.
3Productivity
If automated ML filtering is implemented, then filtering speed and coverage improve, but false positives and accuracy deteriorate due to algorithm limitations
Solution Approach 1:
The system incorporates feedback mechanisms where users can report filtering errors or false positives. This feedback is used to retrain and refine the machine learning models, continuously improving accuracy. The feedback loop allows the system to learn from real-world performance and adjust its filtering decisions over time.
Solution Approach 2:
The patent employs dynamic filtering approaches where the machine learning models continuously adapt to new patterns and contexts. The system adjusts its filtering criteria based on evolving social media trends, emerging forms of harassment, and user feedback, maintaining high accuracy despite changing conditions.
4Object-affected harmful factors
If strict content filtering is applied, then safety and harm reduction improve, but user freedom of expression and access to content deteriorate
Solution Approach 1:
The patent segments content into different categories based on harm level and user preferences. The system creates separate feeds or sections for filtered content and accessible content, allowing users to navigate between them. Users can access filtered content through specific interfaces while maintaining a safe default feed, preserving both safety and access.
Solution Approach 2:
The system applies local quality by allowing users to customize their filtering preferences and control what content is blocked in different contexts. Users can adjust filter sensitivity, select specific types of harmful content to block, and control the level of filtering applied to different parts of their feed, balancing safety with expression freedom.
Data Source
AI summary
A method, system, and product for filtering out unwanted social media content in real-time. The system comprises multiple sets of machine learning classifiers to filter out the unwanted content on any media including but not limited to text, images, audio, and video. Classifiers are trained with labeled data. After being trained, the models screen the incoming real-time data either on a server or a mobile device. A user application is run that results in only approved content to be displayed on the main screen of the user application device. The unwanted data are still available if the user desires to access them. The classifiers are trained with labeled data; and with input parameters in addition to the labeled data. On the device, customized models are trained with the individual user data and Transfers Learning models. When unwanted content is detected, a report is sent to an entity that might help support the receiver.