Track: Ph.D. Thesis/Dissertation Presentation

Abstract

Contact centers, as the organization’s touch point, have a considerable effect on customer experience and retention.
It has been shown that 70% of all business interactions are handled in contact centers. A framework is proposed in this conceptual
paper to build cleaned interactive customer recognition framework (CICRF) in CCs. CICRF consists of two integrated modules:
cleansing and ICRF. The first module focuses on the detection and resolution of duplicate records to improve the effectiveness and
efficiency of customer recognition. The second module focuses on interactive customer recognition in a customer database when
there are multiple records with the same name. Cleansing module uses Semi-Automatic deduplication process by incorporating
three main functions in its design, namely: DedupCrowd, DedupNN and DedupCSR. DedupCrowd is a function that provides
training pairs of records for DeduppNN which is a deduplication based neural network. Researchers suggest leveraging human
computing power in managing duplicate data which is scalable top the large size of contact centers data. However completion of
crowdsourcing tasks is an error-prone process that affects the overall performance of the crowd. Thus, controlling the quality of
workers is an essential step for crowdsourcing systems and for that I propose OSQC, an online statistical quality control
framework, to monitor the performance of workers. DeduppNN is a neural network based deduplication method that uses output
of DedupCrowd for the training purposes. DeduppNN has two features: first is that it is an online deduplication method which is
essential for the purposes of customer recognition. Second is that in terms of costs it is much lower in comparison with
DedupCrowd. The last function is designed for providing label to pairs when DedupNN is not sure about their label. The intuition
behind this function is similar with active learning area which selects appropriate data for labeling. ICRF consists of three
integrated sub-modules. The first sub-module (DedupNNSelect) focuses on the detection and resolution of duplicate records to
improve the effectiveness and efficiency of customer recognition. The second sub-module determines the level of ambiguity in the
recognition of an individual customer in a customer database when there are multiple records with the same name. The third submodule,
depending on the level of determined ambiguity from the second module, recommends to the CSR the series of featurerelated
questions that need to be asked of the customer for his/her recognition.

A dynamic-programming-based technique is proposed in the second module to determine the level of ambiguity that the customer database entails. This module calculates the level of ambiguity in the recognition of an individual customer in a normal scalar form and utilizes fuzzy logic to express it in the form of linguistic variable which is more readable and convenient for CSRs. For the feature selection method in our third module, three statistical approaches, namely (i) Levenshtein edit distance in combination with weights based on the Inverse Document Frequency (IDF) of terms, (ii) statistical tests based on the Analysis of Variance (ANOVA) method, and (iii) decision trees based on the C4.5 based algorithm, are proposed and empirically compared. This module aims to select the algorithm which requires a minimum amount of questions to be asked from the customer for his/her recognition. I evaluate our proposed framework on a synthetic dataset and demonstrate how it assists the CSR in the recognition of the correct customer.