Dual-Model Chain Scoring for Call Center Satisfaction
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current call center analytics systems evaluate interactions independently, failing to account for the evolution of customer experience over a chain of interactions, leading to inaccurate results due to insufficient data from follow-up surveys.
Innovation Solution
A system and method that train two models: a first model predicts an initial customer satisfaction score for each interaction, and a second model refines this score by considering the previous interaction in the chain, using metadata and features of the chain to compute a combined customer satisfaction score for the entire chain.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If each interaction is evaluated independently using current analytics systems, then the evaluation process is simple and fast, but the accuracy of customer experience scoring deteriorates because the evolution of customer experience over the chain is not captured
Solution Approach 1:
The evaluation process is segmented into two distinct phases: a first training phase that evaluates each interaction independently to capture baseline patterns, and a second training phase that refines scores by incorporating chain-level context. This segmentation allows the system to progressively build accuracy without overwhelming complexity in a single model.
Solution Approach 2:
The first model is trained in advance to establish baseline interaction evaluations before the second model is applied. This preliminary action creates a foundation of independent interaction scores that the second model can then refine by incorporating chain-level dependencies, improving overall accuracy while maintaining a structured approach.
2Quantity of substance
If follow-up surveys are used to train models, then customer feedback is captured, but the number of interaction chains with complete survey data is very small, leading to insufficient training data
Solution Approach 1:
The training data is segmented into two layers: individual interaction-level data from surveys (which may be sparse) and chain-level synthesized data generated by the first model. This segmentation allows the system to utilize all available survey responses while supplementing them with model-generated chain contexts, effectively increasing the training dataset volume.
Solution Approach 2:
The first model acts as an intermediary that generates predicted scores for interactions lacking survey data. These predicted scores serve as synthetic training labels, allowing the second model to be trained on complete interaction chains even when original survey data is missing, thereby expanding the effective training dataset.
3Adaptability or versatility
If a single model is used to evaluate interactions, then the system is simpler to implement, but it cannot capture the evolution of customer experience across multiple interactions in a chain
Solution Approach 1:
The system is segmented into two specialized models: the first model handles independent interaction evaluation, and the second model handles chain-level context integration. This segmentation allows each model to be optimized for its specific function, with the second model specifically designed to capture customer experience evolution across interactions.
Solution Approach 2:
The first model serves as an intermediary that processes individual interactions and generates preliminary scores, which are then fed into the second model. This intermediary role allows the system to progressively build context awareness, with the second model receiving both raw interaction data and preliminary evaluations to produce refined scores that capture chain-level patterns.
Data Source
AI summary
Calculating a score for a chain of interactions in a call center may include: during a first training phase, train a first model which, given an interaction and interaction metadata, predict an initial estimated customer satisfaction score; during a second training phase, train a second model which, given an interaction and interaction metadata, text and metadata of an immediately preceding interaction in a chain of interactions, and features of the chain, predict a refined estimated customer satisfaction score; and during an inference phase: given a chain of interactions and metadata of each interaction, compute an initial estimated customer satisfaction score for each interaction using the first model; beginning with a second interaction in the chain and metadata of each interaction, compute a refined estimated customer satisfaction score for each interaction using the second model; combine the interaction scores into a combined customer satisfaction score; and output the score.