Updated: October 2, 2012
The Data Quality Services (DQS) data matching process enables you to reduce data duplication and improve data accuracy in a data source. Matching analyzes the degree of duplication in all records of a single data source, returning weighted probabilities of a match between each set of records compared. You can then decide which records are matches and take the appropriate action on the source data.
The DQS matching process has the following benefits:
Matching enables you to eliminate differences between data values that should be equal, determining the correct value and reducing the errors that data differences can cause. For example, names and addresses are often the identifying data for a data source, particularly customer data, but the data can become dirty and deteriorate over time. Performing matching to identify and correct these errors can make data use and maintenance much easier.
Matching enables you to ensure that values that are equivalent, but were entered in a different format or style, are rendered uniform.
Matching identifies exact and approximate matches, enabling you to remove duplicate data as you define it. You define the point at which an approximate match is in fact a match. You define which fields are assessed for matching, and which are not.
DQS enables you to create a matching policy using a computer-assisted process, modify it interactively based upon matching results, and add it to a knowledge base that is reusable.
You can re-index data copied from the source to the staging table, or not re-index, depending on the state of the matching policy and the source data. Not re-indexing can improve performance.
You can perform the matching process in conjunction with other data cleansing processes to improve overall data quality. You can also perform data de-duplication using DQS functionality built into Master Data Services. For more information, see Master Data Services Overview.
The following illustration displays how data matching is done in DQS:
As with other data quality processes in DQS, you perform matching by building a knowledge base and executing a matching activity in a data quality project in the following steps:
Create a matching policy in the knowledge base
Perform a de-duplication process in a matching activity that is part of a data quality project.
You prepare the knowledge base for performing matching by creating a matching policy in the knowledge base to define how DQS assigns matching probability. A matching policy consists of one or more matching rules that identify which domains will be used when DQS assesses how well one record matches to another, and specify the weight that each domain value carries in the matching assessment. You specify in the rule whether domain values have to be an exact match or can just be similar, and to what degree of similarity. You also specify whether a domain match is a prerequisite.
The matching policy activity in the Knowledge Base Management wizard analyzes sample data by applying each matching rule to compare two records at a time throughout the range of records. Records whose matching scores are greater than a specified minimum are grouped in clusters in the matching results. These matching results are not added to the knowledge base; you use them to tune the matching rules. Creating a matching policy can be an iterative process in which you modify matching rules based on the matching results or profiling statistics.
You can specify for a domain that data strings will be normalized when you load data from the data source into the domain. This process consists of replacing special characters with a null or a space, which often removes the difference between two strings. This can increase matching accuracy, and can often enable a matching result to surpass the minimum matching threshold, when without normalization it would not pass.
Null values in the corresponding fields of two records will be considered a match.
The matching policy is run on domains mapped to the sample data. You can specify whether data is copied from the data source into the staging table and re-indexed when you run the matching policy, or not. You can do so both when building the knowledge base and when running the matching project. Not re-indexing could result in improved performance. Re-indexing is not necessary if the following is true: the matching policy has not changed, and you have not updated the data source, remapped the policy, selected a new data source, or mapped one or more new domains.
Each matching rule is saved in the knowledge base when it is created. However, a knowledge base is available for use in a data quality project only when it is published. In addition, until the knowledge base is published, the matching rules in it cannot be changed by a user other than the person who created it.
DQS performs data de-duplication by comparing each row in the source data to every other row, using the matching policy defined in the knowledge base, and producing a probability that the rows are a match. This is done in a data quality project with a type of Matching. Matching is one of the major steps in a data quality project. It is best performed after data cleansing, so that the data to be matched is free from error. Before running a matching process, you can export the results of the cleansing project into a data table or .csv file, and then create a matching project in which you map the cleansing results to domains in the matching project.
A data matching project consists of a computer-assisted process and an interactive process. The matching project applies the matching rules in the matching policy to the data source to be assessed. This process assesses the likelihood that any two rows are matches in a matching score. Only those records with a probability of a match greater than a value set by the data steward in the matching policy will be considered a match.
When DQS performs the matching analysis, it creates clusters of records that DQS considers matches. DQS randomly identifies one of the records in each cluster as the pivot, or leading, record. The data steward verifies the matching results, and rejects any record that is not an appropriate match for a cluster. The data steward then selects a survivorship rule that DQS will use to determine the record that will survive the matching process and replace the matching records. The survivorship rule can be “Pivot record” (the default), “most complete and longest record”, “most complete record”, or “longest record”. DQS determines the survivor (leading) record in each cluster based upon which record most closely matches the criteria or criterion in the survivorship rule. If multiple records in a given cluster comply with the survivorship rule, DQS selects one of those records randomly. DQS gives you the choice of displaying clusters that have records in common as a single cluster by selecting “show non-overlapping clusters”. You must execute the matching process in order to display the results according to this setting.
You can export the results of the matching process either to a SQL Server table or a .csv file. You can export matching results in two forms: first, the matched records and the unmatched records, or second, survivorship records that include only the survivor record for a cluster and the unmatched results. In the survivorship records, if the same record is identified as the survivor for multiple clusters, that record will only be exported once.