Advanced Analytics for Call Center Operations Ali Cabukel, Senior Data Mining Specialist Global Bilgi Kubra Fenerci Canel, Big Data Solutions Lead Oracle
Speaker Bio Ali Çabukel Graduated from Hacettepe University Statistics department Experienced in development, DWH, Advanced Analytics projects in many sectors including service, telecommunication and call center Kübra Fenerci Canel Msc. in Bogazici University industrial engineering Currently, Big Data Solutions Lead at Oracle 6+ years experience in data mining, forecasting, optimization and big data
Agenda Turkcell Global Bilgi Company Profile Global Bilgi Data Mining Process Project Timeline Oracle Advanced Analytics Advantages for Global Bilgi Project I: Collection (accessibility) Scoring Project II: IVR Analytics Project III: Agent Analytics (ongoing) Next Steps Q&A
Turkcell Global Bilgi Company Profile Was established in 1999 and provides services with a total of 25 locations, including 20 in Turkey, 4 in Ukraine and 1 in Russia, with over 12.000 Expertise in the telecommunication, public sector, finance, energy and retail industries The company creates value in the fields of customer services, customer acquisition, telesales, technical support, customers retention and loyalty, collections, customer information management and analysis Won 1st place in the category of Best Customer Experience and Application at the Contact Center World's 2014 Top Ranking Performance Awards, the largest worldwide organization of the Contact Center Industry.
Global Bilgi Data Mining Process Oracle DWH, feeded by different source system ü Data Integration Layer Data Pool ü Data Summarization Layer Presentation Layer OAA Option ODM + ORE June 2015+
Global Bilgi Data Mining Process
Project Timeline
1 OAA Advantages for Global Bilgi In- DB analytics Contact center data is big enough! Using in- DB R functionalities without memory limitations, in a scalable architecture 2 Analysing data in original environment Ensuring data consistency Facilitating all cycle management
3 OAA Advantages for Global Bilgi Workflow Logic Providing data preparation, modelling and deployment in the same flow 4 Building Dynamic Structures Contribution of PL/ SQL, ORE and DBMS_DATA_MINING packages for dynamism ODM GUI based workflows with SQL flexibility
Collection Scoring Business Requirement : We reach our customers, who are late in payment to remind their debt and convince them about on- time bill payment. In this project, we aim at reaching customers faster and more efficiient. Solution Suggestion : Analysing past behaviour of the customers, predictive models for call period. and building Action : Prioritization of calling customers regarding their accessibility scores. the medium run, reaching from different channels depending on their score segments. In Expected Benefit : Increasing profitability and operational efficiency by calling prioritization as a result of «increasing customer access ratio». Realized Benefit : Increased customer access ratio 12 % after our first calls Project Time : 5 months
Collection Scoring Call detail and other traces of our customers in our system are used Investigated 270 different variable o Analytics functions o Derived attributes o Trend variables onormalizations o Processed 37 Mio call details Scores 2 mio customers every month o Collection calls are done using our accessibility scores
Collection Scoring Data Preparation o Defining target variable o Data Quality o Designing Integrated data envrionment o Data Decomposition o Analytics Datamart Design osampling Data Analysis o Investigating Variables o Classification and Weighting o Variable Selection Modelling o Predictive Modelling / GLM Deployment
Collection Scoring Understanding Business Needs Understanding business need: Understanding faster reachable customers in collection list so that business will be able to reach more debtors and save time Scope: 2014 Feb 2015 May Individual customers and breakdowns
Collection Scoring Data Understanding
Collection Scoring Data Preparation OBSERVATION PERIOD ( L12M) PERFORMANCE PERIOD ( M13) Feb. 14 Mar. 14 Apr. 14 May. 14 June. 14 July. 14 Aug. 14 Sept. 14 Oct. 14 Nov. 14 Dec. 14 Jan. 15 Feb. 15 TRAINING May. 14 June. 14 July. 14 Aug. 14 Sept. 14 Oct. 14 Nov. 14 Dec. 14 Jan. 15 Feb. 15 Mar. 15 Apr. 15 May. 15 VALIDATIO N Data is splitted for each campaign For training and validation, data is grouped by GSM Target variable definition is done Observation period is divided into quarters, and activeness is investigated Recent data ( Will be modelled) Non- recent data ( Will be modelled) New comers= Who are in collection list for performance period, but does not exist in Observation period ( Will not be modelled)
Collection Scoring Data Preparation Example: Total call time/ number of collections Example: Last 6 months/ last 12 months call time 203 variables derived ( gsm basis) from 23 variables ( from call data) 90 out of 203 new variables are derived using business aspect
Collection Scoring Data Preparation ORE is used for binning, WOE and IV calculations. WOE values of high IV variables are integrated to modelling dataset. Higher the variable IV, more explanatory on target Using binning, extreme values and missing values are handled without data loss ODM Attribute importance is also used to understand explanatory variables. Correlation matrices are calculated using ORE, and using highly correlated inputs is avoided
Collection Scoring Modelling
IVR Analytics Business Requirement : Understanding underlying reasons of «not finding a solution on IVR channel» and directing to agent. Solution Suggestion : What are the patterns of customers, who could not solve their issues on 1st IVR interaction? What are the differences between customers who find a solution on 1st IVR int./ customers who cannot find a solution on 1st IVR int. Action : Optimizing IVR process, revising menu design. In the medium run, real time actions about our customers on IVR. Expected Benefit : Increasing operational efficiency, customer experience and service quality. Project Time : 4 Months
IVR Analytics Call detail and other traces of our customers in our system are used o IVR master/ IVR details tables o Training data: 15. 04. 2015-30. 09. 2015 Investigated 280 different variable o Analytics functions o Derived attributes o Trend variables onormalizations o Pattern recognition function in DB 12c Processed 134 Mio call details
IVR Analytics Data Preparation o Defining target variable o Data Quality o Designing Integrated data envrionment o Data Decomposition o Analytics Datamart Design osampling Data Analysis o Investigating Variables o Classification and Weighting o Variable Selection Modelling o Classification models/ decision trees/ SVM/ Naive Bayes o Asscociation Rules Presentation
IVR Analytics Project Steps
IVR Analytics Data Decomposition
IVR Analytics Sampling Cross-validation is used due to data size Data is divided into 10 random equal parts attribute importance is ran (targets: otomation nonautomation / Bubble non-bubble/ Agent Direct Non-Agent direct Which inputs are valuable for which target
IVR Analytics Data Enrichment Techniques used Min, Max, Avg, Sum, Standart deviation methods and normalization Relative variables Flag variables Moving Average, Cumulative Sum, Rank, Percentile, Distinct Count fuctions Variables generated by pattern recognition
IVR Analytics Descriptive Analytics General Results If customer spends longer time in an IVR module, they leave without continuing If they have error at later times, they leave without continuing Underlying patterns for quitting, MT direct, otomation are understood.
IVR Analytics Descriptive Analytics
IVR Analytics Variable Selection Single/multiple interaction 4 different samples Ending type: Quitting, automation, directing agent
IVR Analytics Target Definition What is a recurring transaction? Understanding calling patterns in a slipping time window Analysing patterns with Oracle DB pattern matching function
IVR Analytics Modelling SVM, Decision Tree, ARM models are built SVM stats are as follows:
Agent Analytics (ongoing) Ongoing projects and prototype project ideas Customer Agent Profiling o Clustering Models Customer Agent Anomaly Detection o Descriptive Statistics o Anomaly Detection Target Budget Prediction o Classification Models