Data Mining  

Purpose

You can use data mining to automatically determine significant patterns and hidden associations from large amounts of data. Data mining provides you with insights and correlations that had formerly gone unrecognized or been ignored because it had not been considered possible to analyze them.

Since each company has different data mining requirements, it is not possible to deliver fixed models for producing prediction results. However, the data mining methods available in SAP BW allow you to create models according to your requirements and then use these models to draw information from your SAP BW data to assist your decision-making. For example, you can analyze patterns in customer behavior and predict trends by identifying and exploiting behavioral patterns. Data mining models can be used to provide answers to decision-making questions like the following:

·        Which offer is most appropriate to which customers and when should that offer be made?

·        Which customers are liable to churn?

·        How high is the cross-selling potential for a new product?

Implementation Considerations

You can access data mining methods from the SAP Easy Access menu under Enhanced Analytics  → Data Mining Models.

The data mining methods can also be accessed from the menu for the role Customer Behavior Analysis (SAP_BWC_CUSTOMER_BEHAVIOR).

Integration

Alongside SAP's own data mining methods, you can alternatively use the role Customer Behavior Analysis (SAP_BWC_CUSTOMER_BEHAVIOR) to access an interface to the IBM Intelligent Miner.

Features

SAP delivers the following SAP-owned data mining methods, which can be supplemented by the models that you create:

·        Decision Trees

·        Clustering

·        Association Analysis

·        Scoring

·        Weighted Score Tables

·        ABC  Classification

Decision trees display data using (non-continuous) category quantities. The display rules are determined in training using those sections of historic data where the assignment to categories is already known.

Clustering is used to split data into homogeneous groups. The model looks for a global structure for the data with the aim of partitioning the data into clusters.

Association analysis can be used to establish composite effects and thereby identify cross-selling opportunities, for example. The search for associations considers objects with information content that is remotely comparable. Statements are formulated about partial structures in the data and take the form of rules.

In contrast to decision tree classification, clustering and association analysis determine the models using the data itself.

In scoring, data is displayed using continuous quantities. If required, discretization can then be applied to split the data into classes. The scoring function can either be specified using weighted score tables or be determined by training using historic data as linear or nonlinear regression of a target quantity.

ABC Classification displays data grouped into classes of A, B, C and so  on, using thresholds and classification rules.  The classified results are displayed in the form of ABC chart or list.

You can use historic data to train the models that you create for these data mining methods. This data helps the model to learn by establishing formerly unrecognized patterns. You can either export the result of this learning process into another system (association rules) or you apply the result during prediction to other data that lacks certain information (clustering, decision trees).

You use BW queries to train the model and perform the prediction. You assign these BW queries to the model as sources for the respective business transaction.