Call Volume Prediction Using Embedding Neural Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Unified Communications as a Service (UCaaS) platforms face challenges in predicting call volumes and network traffic with limited data sets, making it difficult to efficiently manage resources and deploy servers effectively.
Innovation Solution
Implementing a machine learning model with embedding functions that apply to time series data with associated metadata to predict call volumes, using deep learning neural networks and transferring learned embeddings from similar time series data to reduce the need for vast training data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional machine learning models are used for call volume prediction, then prediction accuracy may be achieved with sufficient training data, but the system requires vast amounts of training data which is often unavailable in UCaaS environments
Solution Approach 1:
The patent applies preliminary action by pre-training embedding functions on external time series data sources (such as weather data, traffic data, or other temporal patterns) before deploying them for call volume prediction. This pre-training establishes foundational temporal pattern recognition capabilities that can be transferred to the UCaaS prediction task, eliminating the need to start from scratch with limited call volume data.
Solution Approach 2:
The patent introduces embedding functions as an intermediary layer between the input time series features and the prediction model. These embedding functions serve as a mediator that transforms raw time series data into meaningful representations that capture temporal dependencies. By using this intermediary, the system can leverage patterns from diverse data sources and transfer them to the prediction task, reducing dependence on large volumes of task-specific training data.
2Reliability
If UCaaS platforms deploy sufficient telephony servers to handle peak call volumes, then service reliability is improved, but resource utilization efficiency decreases due to servers remaining inactive during low-volume periods
Solution Approach 1:
The patent applies dynamics by implementing a dynamic server provisioning system that adjusts the number of active telephony servers based on predicted call volumes. The machine learning model forecasts future call volumes, and the system dynamically scales server resources accordingly - activating additional servers before predicted peak periods and deactivating them during low-volume periods. This dynamic adjustment maintains service reliability during high demand while optimizing resource utilization during low demand.
Solution Approach 2:
The system performs preliminary action by predicting future call volumes and proactively provisioning server resources before peak demand occurs. Rather than reacting to actual call volumes after they occur, the machine learning model forecasts upcoming call patterns and triggers server activation in advance, ensuring service availability is maintained while avoiding the need for permanently over-provisioned infrastructure.
3Adaptability or versatility
If on-premises telephony systems are used, then system control and security are maintained, but remote workers and mobile users cannot access communication services outside the physical location
Solution Approach 1:
The patent applies universality by implementing a cloud-based UCaaS platform that consolidates multiple communication services (voice, video, messaging, conferencing) into a single accessible system. This unified cloud platform replaces the need for separate on-premises telephony infrastructure at each location, providing universal access to all communication services through the network for both remote and on-site users, while maintaining centralized control and security management.
Data Source
AI summary
A machine learning model (e.g., including a deep learning neural network) with learned embeddings is applied to time series data with associated metadata to obtain predictions of the time series value. For example, a call volume in a period of time may be predicted based on call volume data for a sequence of time bins in a window of preceding time. Time bins may be associated with respective metadata, such as day of week, hour of day, day of month, holiday, part of business cycle, weather, and/or tide. These pieces of metadata may be mapped to embedding vectors using trained embedding functions. The resulting embedding vectors may be input to a neural network along with the corresponding time series data (e.g., call volumes) to make a prediction for future time bin. For example, the prediction may be used to provision servers in a network infrastructure.