Forecasting Virtual Contact Center Queue Metrics Using Hierarchical ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Managing and forecasting resource usage in virtual contact centers is complex due to increasing scale and complexity, leading to inefficiencies in staffing and resource allocation, which can result in either long queues or idle agents, affecting customer satisfaction and labor costs.
Innovation Solution
Implementing a contact center management system that uses machine learning models, specifically temporal convolutional neural networks, to predict demand metrics like contact volume and average handling times, allowing for accurate forecasting and staffing optimization across multiple queues and channels, with automatic retraining and post-processing for noise reduction.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional forecasting methods are used at queue level, then individual queue accuracy is improved, but data insufficiency leads to poor forecasting quality
Solution Approach 1:
The patent combines multiple queues into a parent queue aggregate, merging data from multiple sources to achieve sufficient training data for machine learning models. By aggregating queues at a higher granularity level, the system overcomes individual queue data insufficiency while maintaining the ability to forecast at the individual queue level through learned patterns.
Solution Approach 2:
The patent changes the forecasting dimension from individual queue level to parent queue aggregate level for training purposes. This dimensional shift allows the system to access sufficient data for model training, then apply the learned models back to individual queues by leveraging the hierarchical relationship between queues and their parent aggregates.
2Measurement precision
If machine learning models are trained at queue level, then individual queue forecasting is improved, but computing resources and training time increase
Solution Approach 1:
The patent segments the training process into two distinct phases: first training at the parent queue aggregate level, then fine-tuning or applying to individual queues. This segmentation reduces the overall computational burden by avoiding redundant full-model training at every queue level, while still achieving accurate individual queue forecasts.
Solution Approach 2:
The patent performs preliminary training at the parent queue aggregate level before applying to individual queues. This preliminary action establishes baseline patterns and reduces the computational complexity required for individual queue forecasting, as the model has already learned general contact center patterns at the aggregate level.
3Reliability
If more agents are allocated to queues, then customer satisfaction is improved, but labor costs increase
Solution Approach 1:
The patent implements feedback loops where forecasting results inform staffing decisions, which in turn affect actual queue performance. The system continuously learns from the relationship between agent allocation and queue metrics, optimizing future staffing recommendations to balance customer satisfaction with cost efficiency.
Solution Approach 2:
The patent changes staffing parameters dynamically based on forecasted contact volumes and queue conditions. Rather than static agent allocation, the system adjusts agent assignments in response to changing demand patterns, optimizing the balance between maintaining low wait times for customers and minimizing labor costs.
Data Source
AI summary
Methods, systems, and computer-readable media for accurate usage forecasting for virtual contact centers are disclosed. A contact center management system configures a contact center instance associated with a client. The contact center instance comprises a plurality of queues configured to store contacts. At least a portion of the contacts are routed to a plurality of agents. The contact center management system determines, using one or more machine learning models associated with the contact center instance, a plurality of predictions for a plurality of metrics for the contact center instance for a plurality of time horizons. At least a portion of the predictions are generated for individual queues of the contact center instance.