In this paper, we present an ontology-based approach to generate work ows which are core in adaptive data processing by taking advantage of ontologies in characterizing implicit relations among tasks of data processing. Moreover, compared with manual con gurations of current approaches, ontology reasoning can automatically infer preference orders of tasks and overcome the limitation of manual con gurations. Experimental results show that our proposal is e ective and e cient.
The rst step of our proposal is constructing ontologies in Protege [
Cla and { { { ss De nition There are three classes, namely, ML Process, ML Process Parameters. Each of them has some sub classes.
The ML Process class represents the whole pro cedure of data pro cessing and it contains four main sub classes, namely, Data Preprocess, Feature Preprocess, Modeling, and Evaluation.
The ML Process Service class represents a set of all services and it contains four sub classes, namely, Data Preprocess Service, Feature Preprocess Service, Modeling Service, and Evaluation Service.
The Parameters class represents a set of all parameters and it contains many sub classes of parameters w.r.t. variable mo dels.
Service, Property De nition There are two kinds of prop erties, namely, object (relations among classes) and data (relations b etween entities and datatyp e). There are 14 ob ject prop erties as well as 7 data prop erties.
Note that getModelRequirement and getParameters contain many subprop erties which are used to select mo dels and parameters. Besides, three subprop erties hasC, hasGamma and hasClass Weight of getParameters are used to select parameters which are imp ortant to SVM.
Entity De nition According to a sp eci c application scenario, we
entities as normal [
Citi-bike pre dictio n.
Rule De nition We adopte SWRL [
New York Citi-bike parking quanlity predication (www.citibikenyc.com/systemdata): As the rents/returns of bikes at di erent stations in di erent periods are unbalanced, it is interesting to predict the number of bike parking spots to be rent from/returened to each station cluster. There are 3 steps to generate an optimal work ow as follows: Step 1 Instantiating a Citi-bike parking quanlity predication, named citibike, which have three properties, namely, isCase (to a string \citibike "), formatOfData (to a string \SQL"), and accuracy in general (to a boolean value \ture ") shown in Fig. 2 as well as SWRL rules (e.g., rules in Table 1). Step 2 Generating a work ow by ontology reasoning via Protege as follows: data preprocess ! feature preprocess ! modeling ! evaluation shown Fig. 2. Step 3 Executing the work ow through properties in red from data preprocess to feature preprocess, modeling, and evaluation.
As we can see, the results of Citi-bike parking quanlity predication via the generated work ow are the same as the results via optimally manual con gurations shown in Fig. 3.
In this paper, we presente an ontology-based approach to generate work ows so that optimal models and parameters can be automatically selected in data processing. Our proposal provides a novel way for adaptive data processing via ontologies and is helpful to apply ontology techniques for data processing.
This work is supported by the Key Technology Research and Development Program of Tianjin (16YFZCGX00210), the National Natural Science Foundation of China (61502336), the National Key R&D Program of China (2016YFB1000603), and the Seed Foundation of Tianjin University (2018XZC-0016).