The use of machine learning has become a common approach for solving complex problems across multiple application domains. As its usage often requires training and validation of models with large and heterogeneous datasets, the engineering of these datasets becomes a critical task, although in many cases it does not follow any well-de ned process. In this demonstration paper, we present a novel approach to dataset engineering, which comprises the construction, structuring, understanding, and reuse of datasets from a semantic perspective. Our approach uses a hybrid conceptual model called Hyperknowledge, which can semantically describe both symbolic and non-symbolic nodes, including representing the datasets' structure and enabling dataset retrieval/creation queries.
As the popularity of Machine Learning (ML) tasks increases, so does the amount
of data used to train and test them. The e ectiveness of such algorithms is related
to the quality and variety of the data applied during the training stage [
HKW is a conceptual model that can represent, in the same description
framework, high-level semantic concepts and unstructured data. By semantic
concepts, we mean high-level description regarding linked data, such as facts
about subjects, or specialized knowledge from a given domain, formally described
by ontologies [
To support the understanding of the examples illustrated in the remainder of this paper, we present a simpli ed formalization of HyQL grammar 2 and its processing engine. First, it is important to understand the basic behavior of HyQL when selecting HKW entities using the SELECT clause. A basic selection returns the concept passed in the clause and any other concept that is linked to this concept by a link that contains a connector of type hierarchical 3. Figure 1(a) illustrates the instanceOf (i.e., the connector is declared as hierarchical) relationship between Dog and Animal. Figure 1(b) depicts the result of a query to select animals and returns the Animal node and also the Dog node.
(a) (b) Fig. 1. Representation of the relationship between concepts Animal and Dog (a) and a query to select all animals (including Animal) from the base (b).
Dataset Representation: The dataset representation task stands for structuring data and its metadata in a common knowledge framework (e.g., ontology). It promotes the enhancement of dataset querying and reuses capabilities, which are important properties for supporting the other DE tasks. As an illustrative example, consider the dataset ontology depicted in Figure 2. It shows the concept 2 Accessible at https://ibm.box.com/v/iswc2021-hyql-grammar. For clarity, we have suppressed the handling of spaces, comments, and lower case keywords. 3 HKW connectors have types. The hierarchical type can be used to represent taxonomical relationships, such as instantiations and specializations of concepts. dataset modelled within its data and metadata concepts, i.e., datatype and class (red arrows) and possible instances of each represented concept (blue arrows).
HKW can be used for describing this ontology. Nodes can represent the ontology's classes and instances (round rectangles in the gure) and links can represent the relationships among these entities (arrows in the gure). The description of each dataset could be organized in contexts, promoting a better organization of the data. Likewise, Figure 3 illustrates an example of how HKW can be used for structuring the information of datasets. In this example, the Cleansed Dataset A is described as a context having other three nested contexts: Training, Validation, and Testing. In each of these contexts, there are descriptions of their content using nodes and links. For instance, it describes that the training dataset contains an image (Image1 ) that has a cat and a dog.
Dataset Retrieval: Suppose one is given the task of classifying wild animals in images. For performing such a task, one could query the KB for retrieving all datasets that contain images using lines 1 to 2 of the query depicted in Figure 4. However, such a query is too broad, as it returns datasets having any type of image.
An HKW entity can have anchors that represent part of its content. For instance, a Fig. 4. Query to return all contexts de ned as node representing an image dataset, that contain images with animals. FROM may have several anchors and clause reduces the search scope to entities contained one can de ne relationships in HKW contexts that are instances of Dataset. using these anchors, that is, de ning a meaning to fragments of the image. If we assume that our KB has images of animals and that information is expressed in terms of HKW relationships, one could de ne an additional constraint to this query, which is depicted in line 3 of Figure 4. If we have a more detailed ontology, one can be even more speci c in the construction of the query.
Dataset Creation: As aforementioned, dataset creation is one of the DE tasks that has been drawing the most attention from the research community, mostly because of its importance and di culty. The creation of new datasets with data from a variety of sources is di cult and can lead to errors. We argue that a full- edged knowledge-oriented approach for DE should support the creation of datasets through queries, which can avoid representation errors by applying a common ontology to describe di erent instances of datasets. For example, in the query depicted in Figure 5(a), every image from CopiedDataset is inserted into NewDataset. Practical use of this query is when one wants to reuse images from a preexisting dataset into a new one. Another example is depicted in Figure 5(b), where a new dataset is created from a selection of data with speci c features. (a) (b) Fig. 5. Query that copies the content of CopiedDataset to the a new context (a) and query that copies the images of dogs from DatasetOfImages (b).
The main goal of this demo is to show HKW features that support dataset
engineering tasks through KES and HyQL. In this sense, we use the Pascal
VOC2012 [
In this paper, we presented a knowledge-based approach for DE in the ML domain. The main contributions of this work are the proposal and demonstration of this approach, which comprises core DE tasks such as dataset representation, retrieval, creation. Data scientists could use the approach here described to nd new datasets, balance, clean, and resample them, and select speci c features, which saves time and promotes better data exploitation.
As the main drawback of our approach is the need to annotate datasets for achieving richer descriptions, the (semi-)automatic annotation of datasets is a future work that will allow saving more of data scientists' time.