João Saúde

Xiaoxiao Li, Joao Saude

Real data collected from different applications that have additional topological structures and connection information are amenable to be represented as a weighted graph. Considering the node labeling problem, Graph Neural Networks (GNNs) is a powerful tool, which can mimic experts' decision on node labeling. GNNs combine node features, connection patterns, and graph structure by using a neural network to embed node information and pass it through edges in the graph. We want to identify the patterns in the input data used by the GNN model to make a decision and examine if the model works as we desire. However, due to the complex data representation and non-linear transformations, explaining decisions made by GNNs is challenging. In this work, we propose new graph features' explanation methods to identify the informative components and important node features. Besides, we propose a pipeline to identify the key factors used for node classification. We use four datasets (two synthetic and two real) to validate our methods. Our results demonstrate that our explanation approach can mimic data patterns used for node classification by human interpretation and disentangle different features in the graphs. Furthermore, our explanation methods can be used for understanding data, debugging GNN models, and examine model decisions.

Guilherme Ramos, Joao Saude, Carlos Caleiro et al.

A usual way to model a recommendation system is as a matrix completion problem. There are several matrix completion methods, typically using optimization approaches or collaborative filtering. Most approaches assume that the matrix is either low rank, or that there are a small number of latent variables that encode the full problem. Here, we propose a novel matrix completion algorithm for recommendation systems, without any assumptions on the rank and that is model free, i.e., the entries are not assumed to be a function of some latent variables. Instead, we use a technique akin to information theory. Our method performs hybrid neighborhood-based collaborative filtering using Kolmogorov complexity. It decouples the matrix completion into a vector completion problem for each user. The recommendation for one user is thus independent of the recommendation for other users. This makes the algorithm scalable because the computations are highly parallelizable. Our results are competitive with state-of-the-art approaches on both synthetic and real-world dataset benchmarks.

João Saúde, Guilherme Ramos, Carlos Caleiro et al.

The spread of online reviews, ratings and opinions and its growing influence on people's behavior and decisions boosted the interest to extract meaningful information from this data deluge. Hence, crowdsourced ratings of products and services gained a critical role in business, governments, and others. We propose a new reputation-based ranking system utilizing multipartite rating subnetworks, that clusters users by their similarities, using Kolmogorov complexity. Our system is novel in that it reflects a diversity of opinions/preferences by assigning possibly distinct rankings, for the same item, for different groups of users. We prove the convergence and efficiency of the system and show that it copes better with spamming/spurious users, and it is more robust to attacks than state-of-the-art approaches.