A framework for tailoring Knowledge Graph-based similarity for supervised learning

Rita T. Sousa, Sara Silva, Catia Pesquita

LASIGE, Faculdade de Ciências, Universidade de Lisboa

When predicting external relations between KG entities...

A classification task that takes as input a KG and a set of KG entity pairs for which we want to predict a relation that is not encoded in the KG.

... not everything in the KG will be useful.

A large portion of a general-purpose KG may be irrelevant for a specific prediction task.

In complex domains...

KGs can represent multiple views (or semantic aspects) over the underlying data and some may be less relevant to train the model towards a specific target.

... we need to tailor semantic representations

of the KG entities to an independent and specific classification task when the classification target is not encoded in the KG.

Taxonomic Similarity Measures

KG Embeddings Methods

Benchmark datasets


KGs have been explored as providers of features and background knowledge in several machine learning application scenarios, given their ability to provide meaningful context to the data through semantic representations. Despite the recent advances in semantic representations, existing methods are unsuited to tailoring semantic representations to a specific learning target that is not encoded in the KG.

This problem is exacerbated in complex domains, such as the biomedical, where KGs represent multiple views (or semantic aspects) over the underlying data, some of which may be less relevant to train the model towards a specific target. For instance, the prediction of protein-protein interactions using the Gene Ontology is more accurate if only a portion of the ontology is used (in this case, the one concerning biological processes).

evoKGsim+ is a framework that learns suitable semantic similarity-based semantic representations of data objects extracted from KGs optimized for supervised learning. This tailoring is achieved by evolving a suitable combination of semantic aspects using genetic programming (GP). The evolution of the semantic representations is guided by a fitness function based on the success of a given semantic representation in a specific task.

Previously, we presented evoKGsim, but it is limited to taxonomy-based semantic similarity measures, which do not consider the full spectrum of semantic relations that can be established within a KG. evoKGsim+ is an extension of evoKGsim into a full framework that combines several taxonomic and embedding similarity measures and provides several baseline evaluation approaches that emulate domain expert feature selection and optimal parameter setting.


evoKGsim+, available on GitHub , targets classification tasks that take as input a KG and a set of KG individual pairs for which we wish to learn a relation that is outside the scope of the KG.

Step I

Represent each instance (i.e., a pair of KG entities) according to KG-based similarities computed for each semantic aspect.

evoKGsim+ currently supports 10 different KG-based similarity measures based on a selection of representative state-of-the-art approaches:

6 taxonomic semantic similarity measures, derived by combining one of two information content approaches ( ICSeco and ICResnik ) with one of three set similarity measures ( ResnikMax , ResnikBMA , and SimGIC ).

4 measures based on cosine similarity over embeddings generated from TransE , distMult , RDF2Vec and Owl2Vec .

Step II

Employ GP to learn a suitable combination of the different aspect-based similarities, using a set of predefined operators, to address a given task.

GP is a population-based search procedure inspired in Darwinian evolution and Mendelian genetics.

GP starts by creating an initial population of random individuals representing potential combination of semantic aspects. Each individual solution is evaluated and assigned a fitness value that quantifies how accurate are the classification predictions made by using this combination. New generations of potential combinations are iteratively created by selecting parents based on their fitness and breeding them using genetic operators like crossover and mutation. The fitter individuals are selected more often to be parents and thus pass their characteristics to their offspring, which causes the population to improve in quality along successive generations.

Learn more here .

Step III

Evaluate the predictions made on the test set and comparing them against optimized static representations that represent expert feature selection and parameter tuning.

GP is used in a regression-like fashion, treating the expected class labels (0 and 1) as numeric expected outputs, and calculating fitness as the root mean squared error between the expected and predicted values, a normal procedure when using GP for binary classification.

The predicted numeric outputs are transformed in class labels, by applying the natural cutoff of 0.5, only for reporting performance.

Evaluation Strategy

We evaluate the framework in its full extension in protein-protein interaction (PPI) prediction using the Gene Ontology (GO) as the KG and a set of benchmark data sets with a wide range in size and semantic annotation characteristics.

Gene Ontology Knowledge Graph

GO describes protein function with respect to three semantic aspects: biological process (BP), cellular component (CC), and molecular function (MF). GO and the annotations that link proteins to GO classes make up a KG.

Protein-Protein Interaction Datasets

PPI prediction was chosen as evaluation domain since it is well known that BP and CC aspects describe properties that are stronger indicators for PPI than the MF aspect for protein interaction.

The PPI Benchmark datasets are available on GitHub.

Static Semantic Representations

Five static representations as baselines: the BP, CC and MF single aspectsand the average and maximum of the single aspects.

The static representations are employed as a simple similarity threshold-based classifier, where a semantic similarity score for a protein pair exceeding a certain threshold predicts a positive interaction.

Evolved Semantic Representations

The models returned by GP are the combinations of the semantic similarity scores of the three GO aspects, evolved to support PPI prediction.


For evaluating the quality of a predicted classification, we use the weighted average of F-measures (WAF). Stratified 10-fold cross-validation was used for each experiment performed. The results we report are the median of the 10 WAF values calculated on the 10 folds.

One-Minute Video


Rita T. Sousa

LASIGE, Faculdade de Ciências

Sara Silva

LASIGE, Faculdade de Ciências

Catia Pesquita

LASIGE, Faculdade de Ciências


Catia Pesquita, Sara Silva, Rita T. Sousa are funded by the FCT through LASIGE Research Unit, ref. UIDB/00408/2020 and ref. UIDP/00408/2020. Catia Pesquita and Rita T. Sousa are funded by project SMILAX (ref. PTDC/EEI-ESS/4633/2014), Sara Silva by projects BINDER (ref. PTDC/CCI-INF/29168/2017) and PREDICT (ref. PTDC/CCI-CIF/29877/2017), and Rita T. Sousa by FCT PhD grant (ref. SFRH/BD/145377/2019).It was also partially supported by the KATY project which has received funding from the European Union’s Horizon 2020 research and innovation programme under grant agreement No 101017453.