k-skip-n-gram Graph Classification for Temporal Event Sequences
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing online help tools face challenges in real-time, accurate classification of highly variant natural language phrases due to their high volume and temporal bursts, requiring computationally efficient and precise methods to provide relevant information to users.
Innovation Solution
The implementation of a system that uses k-skip-n-grams, graph generation, and supervised machine learning to classify sequences of objects into event or issue categories, employing a trained classifier and metadata analysis to identify and respond to user requests in real-time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional natural language classification methods are used, then the system can provide basic help functionality, but the classification accuracy deteriorates when handling high-volume, temporally-bursty, and highly-variant user requests
Solution Approach 1:
The patent segments the classification task into multiple stages: (1) detecting temporal bursts in request sequences, (2) generating event representations from bursty sequences using k-skip-n-grams, (3) classifying individual requests based on event context. This segmentation allows the system to handle high-volume requests efficiently while maintaining accuracy through specialized processing of temporally-related events.
Solution Approach 2:
The system performs preliminary action by detecting and analyzing temporally-related events and bursts before classifying individual requests. By pre-processing sequences of requests to identify patterns and generate event representations in advance, the system prepares classification models that can quickly and accurately categorize incoming requests during high-volume periods without sacrificing precision.
2Speed
If real-time classification is implemented to meet user needs, then the responsiveness improves, but the computational complexity increases due to the need to process highly-variant natural language phrases
Solution Approach 1:
The patent implements dynamics by making the classification system adaptive to temporal patterns in request data. The system dynamically detects bursts, adjusts its analysis focus based on temporal relationships, and modifies classification behavior according to the detected event context. This dynamic approach allows real-time processing without requiring overly complex static models, as the system adapts its complexity only when temporally-related patterns are detected.
Solution Approach 2:
The system changes parameters by adjusting its processing intensity based on detected temporal patterns. During burst periods, the system activates specialized event detection and k-skip-n-gram analysis with specific parameter settings (k=2, n=3). During normal periods, it uses simpler classification. This parameter adjustment allows real-time responsiveness while managing computational complexity through conditional processing intensity.
3Measurement precision
If the system processes all user requests with high analytical depth, then the classification precision improves, but the processing time increases unacceptableably during high-volume periods
Solution Approach 1:
The patent applies partial action by selectively applying deep analytical processing only to requests that belong to detected temporally-related events or bursts. Not all requests receive the full k-skip-n-gram analysis and event context processing. This selective approach maintains high classification precision for critical temporally-related requests while reducing average processing time across the entire request volume, as normal requests are handled with lighter processing.
Data Source
AI summary
Systems and methods for enhanced classification of sequences of objects based on clique similarity and metadata associated with the sequences are presented. Sequences are received. Events are detected based on analyzing k-skip-n-grams included in the sequences. For each event of the detected plurality of events, a graph is generated. The graph for a particular event includes z-cliques that correspond to portions of the k-skip-n-grams that are included in the sequences that are associated with the particular event. A first sequence, which is separate from the other sequences, is received. The first sequence includes a first plurality of k-skip-n-grams. A trained classifier is employed to classify the first sequence as being associated with a first event of the detected events. Classifying the first sequence is based on a comparison between the first plurality of k-skip-n-grams and the z-cliques of the graph that is generated for the first event.