Data Quality Class 2 David Loshin
Goals Cost of low data quality Mapping the information chain Data Quality impacts Economic measures Impact domains Building the Data Quality ROI Model
Goals 2 Data Cleansing Project –Goal of the application –Components of the application
Cost of Low Data Quality Data quality is measured using anecdotes “Hazy” feeling of wrongness Desire to gauge the true cost of poor data quality
5 Steps Map the Information Chain Categorize costs associated with low data quality Identify and estimate actual effect Determine cost of fixing problem Calculate Return on Investment (ROI)
Evidence of Economic Impact Frequent service interruptions and system failures Drop in productivity vs. volume High employee turnover High new business/continued business ratio Increased customer service requirements Customer Attrition
The Information Chain Data flow model Processing stages Communication/data transfer
The Information Chain 2 Data Supply Data Acquisition Data Creation Data Processing Data Packaging Decision Making Decision Implementation Data Delivery Data Consumption
Information Chain 3 Information chain = data flow graph Processing stages are vertices in graph Directed message-passing channels = directed edges Examples
Impacts of Low Data Quality Hard impacts: can be estimated and/or measured Soft impacts: hard to measure, but definitely are evident
Hard Impacts Customer attrition Costs attributed to error detection Costs attributed to error rework Costs attributed to prevention of errors Costs associated with customer service Costs associated with fixing customer problems Costs associated with enterprisewide data inconsistency Costs attributable to delays in processing
Soft Impacts Difficulty in decision making Time delays in operation Organizational mistrust Lowered ability to effectively compete Data ownership conflicts Lowered employee satisfaction
Economic Measures Cost Increase Revenue Decrease Cost Decrease Revenue Increase Delay Speedup Increase Satisfaction Decrease Satisfaction
Impact Domains Operational Tactical/Strategic
Operational Impacts Detection Correction Rollback Rework Prevention Warranty Reduction Attrition Blockading.
Tactical/Strategic Impacts Delays Preemption Idling Increased Difficulty Lost opportunities Organizational mistrust Alignment Acquisition overhead Decay Infrastructure
Putting it Together Map the information chain Conduct interviews to locate data quality problems Annotate information chain with location of data qualty problems Identify impact domains for each problem Characterize economic impact (=cost!) Aggregate totals
ROI Model Create a spreadsheet with assigned costs Add in costs of improvements Determine best return on investment
Data Cleansing Project Write an application to cleanse data –Record Parsing –Metadata cleansing –Data standardization –Data correction –Data enhancement
Record Parsing Data element types –first names –last names –honorifics –titles –street names –directions –business words –etc.
Data Domains Data types Subclassed data types = domains Mappings between domains
Data Domains 2 Data type = char(2) –676 possible non-punctuation members Data Domain: US State abbreviations –62 possible members Subclassed data domain: “New England” –{“ME”, “NH”, “VT”, “MA”, “CT”, “RI”}
Data Domains 3 Enumerated domains –All values are explicit Rule-based domains –Domain definition is generative
Record Parsing Tokenizing data elements within an attribute Assign meaning to tokens –Domain membership –Patterns –Context
Tokenizing Straightforward –white-space separated –punctuation – important or not? –Result: stream of tokens
Domain Membership Can each token be assigned to a domain? –Based strictly on token value –Based on patterns –Based on context
Domain Membership 2 Domains can be maintained in memory using hash tables Search for domain membership is the same as hash table lookups What if a token belongs to more than one domain?
Patterns Certain kinds of data attributes are organized around token patterns Example: names can appear using these kinds of patterns: (title) (first) (middle) (last) (title) (first) (initial) (last) (first) (middle) (last) (last) (comma) {first) (middle) etc.
Context What happens when a token belongs to more than one domain? We can use context to infer decision Build weights based on frequency = training
Next Week Dimensions of Data Quality Project specification