DeVisa Use Cases

Devisa Use Cases

Summary

Name	Documentation
DM Consumer	The DM Consumer is a WS client application that uses DeVisa DM models.
download	A DM Consumer can download a DeVisa model in order to use it internally, e.g import into the DM application.
authenticate	The WS client that invokes a certain method might need to be authenticated in order to execute the required function.
admin	The Admin use case involves uploading/downloading the DeVisa models without any other processing.
replace	Replace occurs when a model is replaced completely by a newer one. The new model will get the same model ID in the repository.
upload	A DM Producer can upload models in the DeVisa repository via WS methods (SOAP, XMLRPC).
update	Update occurs when a model needs only a certain type of adjustment. This use case should use a specific XML update technique. The update procedure is triggered by the system itself in most of the cases.
search	The searching functions allow inspecting the properties of the PMML models in the repository. The search functions conform and therefore are limited to the information that a PMML model can incorporate according the PMML 3.2 specification. DeVisa provides searching functions such as: Selecting the models with desired properties function type: classification, regression, clustering etc type of model: tree, SVM, cluster etc producer e.g all the models belonging to a certain producer application degree of freshness: e.g models newer than a certain date performance measures fields statistics Selecting the models that conform to a certain schema Exact Schema. the document refers a data dictionary but does not refer a model; It can specify desired properties though. DeVisa engine will select the appropriate model. Match Schema. the document describes a data dictionary (and possibly a mining schema according to the data dictionary). In this case the DeVisa engine is responsible for identifying the matching schema and the appropriate model (should these exist). This type of search is used especially as an extension to the scoring use care Full text search in the model repository This type of search is useful when the client is looking for keywords in the description of the model, in the field names or field description, in the function type etc Future work: create a search mechanism that allows more complex predicates on the search criterions.
scoring	The scoring use case means applying the models on the new instances. The scoring occurs via web service methods. Depending on the models, there are several types of DeVisa scoring procedures: Classification Scoring. Cluster Scoring. Association rules scoring.
classification scoring	The scoring method receives as input a set of instances and one or more classification models and classifies the instances with respect to the models.
clustering scoring	The scoring method receives as input a set of instances and one or more clustering models and assigns the instances to the most appropriate cluster in each of the models.
association rules scoring	The scoring method receives as input a set of items (instances) and one or more association rule models. It determines all rules of each of the input models whose antecedent itemset is a subset of a the input itemset and returns the consequents of these rules as the inferred itemsets. An extension of this procedure computes all rules whose antecedent and consequent itemsets are included in the input itemset. This version is useful to determine which itemsets support which rules.
compose	Model Composition allows the combination of simple models into a single composite PMML model. PMML version 3.2 supports the combination of decision trees and simple regression models. More general variants would be possible and may be defined in future versions of PMML. In PMML Model composition uses three syntactical concepts The essential elements of a predictive model are captured in elements that can be included in other models. Embedded models can define new fields, similar to derived fields. The leaf nodes in a decision tree can contain another predictive model. In DeVisa simple models can be combined into more complex ones forming new valid PMML documents. A client application can specify the models subject to composition and the combination method. DeVisa identifies the specified models. If the models do not exist in the repository then the process stops. The found models are checked for compatibility. If they are not compatible the process stops. The new valid model is returned to the user/stored in the repository.
sequencing	Model sequencing is the process through which two or more models are combined into a sequence where the results of one model are used as input in another model. Model sequencing is supported partially by the PMML specification. Examples of sequencing: The missing values in a regression model can be replaced by a set of rules (or decision tree) Several classification models with the same target value can be merged via a voting scheme, i.e the final classification result can be computed as a an average of the results of the initial classifiers.The average can be computed by a regression model. Prediction results may have to be combined with a cost or profit matrix before a decision can be derived. A mailing campaign model may use tree classification to determine response probabilities per customer and channel. The cost matrix can be appended as a regression model that applies cost weighting factors to different channels, e.g., high cost for phone and low cost for email. The final decision is then based on the outcome of the regression model.
selection	Model selection in PMML allows for combining multiple 'embedded models', aka model expressions, into the decision logic that selects one of the models depending on the current input values. Examples of selection A common method for optimizing prediction models is the combination of segmentation and regression. Data are grouped into segments and for each segment there may be different regression equations. If the segmentation can be expressed by decision rules then this kind of segment based regression can be implemented by a decision tree where any leaf node in the tree can contain an embedded regression model.
compare	A client application (DM Consumer) wants to compare two models. The client needs to specify: The two models to be compared, through an exact reference (model id) The comparison type, which can be syntactic or semantic. Syntactic Comparison. Two PMML models are compared through a XML differencing approach. To be researched if the eXist's XML diff extension module can be used here. Semantic Comparison. Two models can be compared from the following points of view: Schema compatibility. This involves a DataDictionary compatibility check and a MiningSchema compatibility check. Function comparison. It is checked if the models fulfil the same prediction task, the algorithms used to fulfil the task etc. Performance measures (to be researched) The Semantic comparison is useful for the cases in which the client envisions a model composition and wants to pre-check the compatibility of the models
statistics	An application can invoke this service to obtain statistics on the models. Example of statistics: frequencies per domain, producer, schema, or function type etc. ?? based on performance measures (to be researched): sensitivity, specificity, accuracy (from what I've seen they are not supported by PMML) can we connect performance/assesment with model verification  in PMML? At the beginning DeVisa should be able to provide only a full report (PMQL) for the point (1).
DM Producer	The DM Producer is a WS client application -typically a DM application, like Weka- who uploads models in the DeVisa Repository.
DeVisa

Details

DM Consumer

Name	Value
Visibility	public
Abstract	false
Leaf	true
Root	false
Documentation	The DM Consumer is a WS client application that uses DeVisa DM models.
Business Model	false

Relationships

Unamed Association
To	Name Value End Model Element download Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Unamed Association
To	Name Value End Model Element search Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Unamed Association
To	Name Value End Model Element classification scoring Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Unamed Association
To	Name Value End Model Element clustering scoring Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Unamed Association
To	Name Value End Model Element sequencing Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Unamed Association
To	Name Value End Model Element selection Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Unamed Association
To	Name Value End Model Element compare Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Unamed Association
From	Name Value End Model Element statistics Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Unamed Association
From	Name Value End Model Element association rules scoring Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

download

Extension Points

laxly specified model

Relationships

Unamed Include
To	authenticate
Visibility	Unspecified
Stereotypes	Include

Unamed Extend
To	search
Visibility	Unspecified
Stereotypes	Extend

Unamed Generalization
From	admin
Substitutable	false
Visibility	Unspecified

Unamed Association
From	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

authenticate

Relationships

Unamed Include
From	scoring
Visibility	Unspecified
Stereotypes	Include

Unamed Include
From	compare
Visibility	Unspecified
Stereotypes	Include

Unamed Include
From	compose
Visibility	Unspecified
Stereotypes	Include

Unamed Include
From	upload
Visibility	Unspecified
Stereotypes	Include

Unamed Include
From	download
Visibility	Unspecified
Stereotypes	Include

Unamed Include
From	search
Visibility	Unspecified
Stereotypes	Include

Unamed Include
From	statistics
Visibility	Unspecified
Stereotypes	Include

admin

Relationships

Unamed Generalization
To	replace
Substitutable	false
Visibility	Unspecified

Unamed Generalization
To	upload
Substitutable	false
Visibility	Unspecified

Unamed Generalization
To	download
Substitutable	false
Visibility	Unspecified

Unamed Generalization
To	update
Substitutable	false
Visibility	Unspecified

Use Case Descriptions

Admin
Super Use Case
Author	dianagorea
Date	Jan 17, 2008 2:19:00 PM
Brief Description	The DM client uploads/downloads the DeVisa models without any other processing.
Preconditions
Post-conditions
Flow of Events	Actor Input System Response 1

replace

Relationships

Unamed Extend
From	upload
Visibility	Unspecified
Stereotypes	Extend

Unamed Generalization
From	admin
Substitutable	false
Visibility	Unspecified

upload

Extension Points

existing model
Documentation	If the model that the DM producer has uploaded already exists in the DeVisa repository then and the model is newer than the existing one then it is replaced.

Relationships

Unamed Extend
To	replace
Visibility	Unspecified
Stereotypes	Extend

Unamed Include
To	authenticate
Visibility	Unspecified
Stereotypes	Include

Unamed Generalization
From	admin
Substitutable	false
Visibility	Unspecified

Unamed Association
From	Name Value End Model Element DM Producer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

Use Case Descriptions

Upload
Super Use Case	admin
Author	dianagorea
Date	Jan 17, 2008 3:04:44 PM
Brief Description
Preconditions	The DM producer has a model expressed in PMML
Post-conditions	The model is stored in DeVisa PMML repository.
Flow of Events	Actor Input System Response 1 upload request 2 authentication process 3 checks PMML valid

update

Relationships

Unamed Generalization
From	admin
Substitutable	false
Visibility	Unspecified

search

Relationships

Unamed Include
To	authenticate
Visibility	Unspecified
Stereotypes	Include

Unamed Extend
From	scoring
Visibility	Unspecified
Stereotypes	Extend
Condition	The client has specified the model through "Match" or "" case.

Unamed Extend
From	compose
Visibility	Unspecified
Stereotypes	Extend

Unamed Extend
From	download
Visibility	Unspecified
Stereotypes	Extend

Unamed Association
From	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

References

Sequence Diagram/search
Description	A complete search sequence diagram (also interaction with the client)
Type	Diagram

scoring

Extension Points

laxly specified model

Documentation

There are three ways in which a client application -a DM consumer that invokes a scoring method- requests a model in DeVisa. Exact Model. the document refers a model in the DeVisa catalog; The DeVisa engine will use the specified model to execute the scoring task Exact Schema. the document refers a data dictionary but does not refer a model; It can specify desired properties though. DeVisa engine will select the appropriate model. Match Schema. the document describes a data dictionary (and possibly a mining schema according to the data dictionary). In this case the DeVisa engine is responsible for identifying the matching schema and the appropriate model (should these exist). Except for the first case, the model is laxly specified.

Relationships

Unamed Include
To	authenticate
Visibility	Unspecified
Stereotypes	Include

Unamed Extend
To	search
Visibility	Unspecified
Stereotypes	Extend
Condition	The client has specified the model through "Match" or "" case.

Unamed Generalization
To	classification scoring
Substitutable	false
Visibility	Unspecified

Unamed Generalization
To	clustering scoring
Substitutable	false
Visibility	Unspecified

Unamed Generalization
To	association rules scoring
Substitutable	false
Visibility	Unspecified

References

Sequence Diagram/scoring
Type	Diagram

Communication Diagram/scoring - Communications
Type	Diagram

classification scoring

Relationships

Unamed Generalization
From	scoring
Substitutable	false
Visibility	Unspecified

Unamed Association
From	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

References

Communication Diagram/scoring - Communications
Type	Diagram

clustering scoring

Relationships

Unamed Generalization
From	scoring
Substitutable	false
Visibility	Unspecified

Unamed Association
From	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

References

Sequence Diagram/scoring
Type	Diagram

association rules scoring

Relationships

Unamed Generalization
From	scoring
Substitutable	false
Visibility	Unspecified

Unamed Association
To	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

compose

Extension Points

laxly specified models

Relationships

Unamed Include
To	authenticate
Visibility	Unspecified
Stereotypes	Include

Unamed Extend
To	search
Visibility	Unspecified
Stereotypes	Extend

Unamed Generalization
To	sequencing
Substitutable	false
Visibility	Unspecified

Unamed Generalization
To	selection
Substitutable	false
Visibility	Unspecified

sequencing

Relationships

Unamed Generalization
From	compose
Substitutable	false
Visibility	Unspecified

Unamed Association
From	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

selection

Relationships

Unamed Generalization
From	compose
Substitutable	false
Visibility	Unspecified

Unamed Association
From	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

compare

Relationships

Unamed Include
To	authenticate
Visibility	Unspecified
Stereotypes	Include

Unamed Association
From	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

statistics

Relationships

Unamed Include
To	authenticate
Visibility	Unspecified
Stereotypes	Include

Unamed Association
To	Name Value End Model Element DM Consumer Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

DM Producer

Relationships

Unamed Association
To	Name Value End Model Element upload Provide Property Getter Method false Provide Property Setter Method false Multiplicity Unspecified Visibility private Aggregation Kind None Navigable true
Abstract	false
Leaf	false
Visibility	Unspecified
Derived	false

DeVisa

Children

Name	Documentation
scoring	The scoring use case means applying the models on the new instances. The scoring occurs via web service methods. Depending on the models, there are several types of DeVisa scoring procedures: Classification Scoring. Cluster Scoring. Association rules scoring.
compare	A client application (DM Consumer) wants to compare two models. The client needs to specify: The two models to be compared, through an exact reference (model id) The comparison type, which can be syntactic or semantic. Syntactic Comparison. Two PMML models are compared through a XML differencing approach. To be researched if the eXist's XML diff extension module can be used here. Semantic Comparison. Two models can be compared from the following points of view: Schema compatibility. This involves a DataDictionary compatibility check and a MiningSchema compatibility check. Function comparison. It is checked if the models fulfil the same prediction task, the algorithms used to fulfil the task etc. Performance measures (to be researched) The Semantic comparison is useful for the cases in which the client envisions a model composition and wants to pre-check the compatibility of the models
compose	Model Composition allows the combination of simple models into a single composite PMML model. PMML version 3.2 supports the combination of decision trees and simple regression models. More general variants would be possible and may be defined in future versions of PMML. In PMML Model composition uses three syntactical concepts The essential elements of a predictive model are captured in elements that can be included in other models. Embedded models can define new fields, similar to derived fields. The leaf nodes in a decision tree can contain another predictive model. In DeVisa simple models can be combined into more complex ones forming new valid PMML documents. A client application can specify the models subject to composition and the combination method. DeVisa identifies the specified models. If the models do not exist in the repository then the process stops. The found models are checked for compatibility. If they are not compatible the process stops. The new valid model is returned to the user/stored in the repository.
authenticate	The WS client that invokes a certain method might need to be authenticated in order to execute the required function.
admin	The Admin use case involves uploading/downloading the DeVisa models without any other processing.
search	The searching functions allow inspecting the properties of the PMML models in the repository. The search functions conform and therefore are limited to the information that a PMML model can incorporate according the PMML 3.2 specification. DeVisa provides searching functions such as: Selecting the models with desired properties function type: classification, regression, clustering etc type of model: tree, SVM, cluster etc producer e.g all the models belonging to a certain producer application degree of freshness: e.g models newer than a certain date performance measures fields statistics Selecting the models that conform to a certain schema Exact Schema. the document refers a data dictionary but does not refer a model; It can specify desired properties though. DeVisa engine will select the appropriate model. Match Schema. the document describes a data dictionary (and possibly a mining schema according to the data dictionary). In this case the DeVisa engine is responsible for identifying the matching schema and the appropriate model (should these exist). This type of search is used especially as an extension to the scoring use care Full text search in the model repository This type of search is useful when the client is looking for keywords in the description of the model, in the field names or field description, in the function type etc Future work: create a search mechanism that allows more complex predicates on the search criterions.
classification scoring	The scoring method receives as input a set of instances and one or more classification models and classifies the instances with respect to the models.
clustering scoring	The scoring method receives as input a set of instances and one or more clustering models and assigns the instances to the most appropriate cluster in each of the models.
sequencing	Model sequencing is the process through which two or more models are combined into a sequence where the results of one model are used as input in another model. Model sequencing is supported partially by the PMML specification. Examples of sequencing: The missing values in a regression model can be replaced by a set of rules (or decision tree) Several classification models with the same target value can be merged via a voting scheme, i.e the final classification result can be computed as a an average of the results of the initial classifiers.The average can be computed by a regression model. Prediction results may have to be combined with a cost or profit matrix before a decision can be derived. A mailing campaign model may use tree classification to determine response probabilities per customer and channel. The cost matrix can be appended as a regression model that applies cost weighting factors to different channels, e.g., high cost for phone and low cost for email. The final decision is then based on the outcome of the regression model.
selection	Model selection in PMML allows for combining multiple 'embedded models', aka model expressions, into the decision logic that selects one of the models depending on the current input values. Examples of selection A common method for optimizing prediction models is the combination of segmentation and regression. Data are grouped into segments and for each segment there may be different regression equations. If the segmentation can be expressed by decision rules then this kind of segment based regression can be implemented by a decision tree where any leaf node in the tree can contain an embedded regression model.
download	A DM Consumer can download a DeVisa model in order to use it internally, e.g import into the DM application.
replace	Replace occurs when a model is replaced completely by a newer one. The new model will get the same model ID in the repository.
upload	A DM Producer can upload models in the DeVisa repository via WS methods (SOAP, XMLRPC).
statistics	An application can invoke this service to obtain statistics on the models. Example of statistics: frequencies per domain, producer, schema, or function type etc. ?? based on performance measures (to be researched): sensitivity, specificity, accuracy (from what I've seen they are not supported by PMML) can we connect performance/assesment with model verification  in PMML? At the beginning DeVisa should be able to provide only a full report (PMQL) for the point (1).
update	Update occurs when a model needs only a certain type of adjustment. This use case should use a specific XML update technique. The update procedure is triggered by the system itself in most of the cases.
association rules scoring	The scoring method receives as input a set of items (instances) and one or more association rule models. It determines all rules of each of the input models whose antecedent itemset is a subset of a the input itemset and returns the consequents of these rules as the inferred itemsets. An extension of this procedure computes all rules whose antecedent and consequent itemsets are included in the input itemset. This version is useful to determine which itemsets support which rules.