Adapting k-means for graph clustering

Predicting onset of disease progression using temporal disease occurrence networks

OBJECTIVE

Early recognition and prevention are crucial for reducing the risk of disease progression. This study aimed to develop a novel technique based on a temporal disease occurrence network to analyze and predict disease progression.

METHODS

This study used a total of 3.9 million patient records. Patient health records were transformed into temporal disease occurrence networks, and a supervised depth first search was used to find frequent disease sequences to predict the onset of disease progression. The diseases represented nodes in the network and paths between nodes represented edges that co-occurred in a patient cohort with temporal order. The node and edge level attributes contained meta-information about patients' gender, age group, and identity as labels where the disease occurred. The node and edge level attributes guided the depth first search to identify frequent disease occurrences in specific genders and age groups. The patient history was used to match the most frequent disease occurrences and then the obtained sequences were merged together to generate a ranked list of diseases with their conditional probability and relative risk.

RESULTS

The study found that the proposed method had improved performance compared to other methods. Specifically, when predicting a single disease, the method achieved an area under the receiver operating characteristic curve (AUC) of 0.65 and an F1-score of 0.11. When predicting a set of diseases relative to ground truth, the method achieved an AUC of 0.68 and an F1-score of 0.13.

CONCLUSION

The ranked list generated by the proposed method, which includes the probability of occurrence and relative risk score, can provide physicians with valuable information about the sequential development of diseases in patients. This information can help physicians to take preventive measures in a timely manner, based on the best available information.

Qluster: An easy-to-implement generic workflow for robust clustering of health data

The exploration of heath data by clustering algorithms allows to better describe the populations of interest by seeking the sub-profiles that compose it. This therefore reinforces medical knowledge, whether it is about a disease or a targeted population in real life. Nevertheless, contrary to the so-called conventional biostatistical methods where numerous guidelines exist, the standardization of data science approaches in clinical research remains a little discussed subject. This results in a significant variability in the execution of data science projects, whether in terms of algorithms used, reliability and credibility of the designed approach. Taking the path of parsimonious and judicious choice of both algorithms and implementations at each stage, this article proposes Qluster, a practical workflow for performing clustering tasks. Indeed, this workflow makes a compromise between (1) genericity of applications (e.g. usable on small or big data, on continuous, categorical or mixed variables, on database of high-dimensionality or not), (2) ease of implementation (need for few packages, few algorithms, few parameters, ...), and (3) robustness (e.g. use of proven algorithms and robust packages, evaluation of the stability of clusters, management of noise and multicollinearity). This workflow can be easily automated and/or routinely applied on a wide range of clustering projects. It can be useful both for data scientists with little experience in the field to make data clustering easier and more robust, and for more experienced data scientists who are looking for a straightforward and reliable solution to routinely perform preliminary data mining. A synthesis of the literature on data clustering as well as the scientific rationale supporting the proposed workflow is also provided. Finally, a detailed application of the workflow on a concrete use case is provided, along with a practical discussion for data scientists. An implementation on the Dataiku platform is available upon request to the authors.

Social media analysis of car parking behavior using similarity based clustering

This paper investigates car parking users' behaviors from social media perspective using social network based analysis of online communities revealed by mining the associated hashtags in Twitter. We propose a new interpretable community detection approach for mapping user's car parking behavior by combining Clique, K-core and Girvan-Newman community detection algorithms together with a content-based analysis that exploits polarity, relative frequency and dominant topics. Twitter API was used to collect relevant data by tracking popular car-parking hashtags. A social network graph is constructed using a similarity-based analysis. Finally, interpretable communities are inferred by monitoring the outcomes of clique, K-core and Girvan-Newman community detection algorithms. This interpretability is linked to the aggregation of keywords, hashtags and/or location attributes of the tweet messages as well as a visualization module that enables interaction with users. In parallel, a global trend analysis investigates parking types and Twitter influence with respect to both sentiment polarity and dominant trends (extracted using KeyBERT based approach) is performed. The implementation of this social media analytics has uncovered several aspects associated to car-parking behaviors. A comparison with some state-of-the-art community detection methods has also been carried out and revealed some similarities with our developed approach.

A Scalable Deep Network for Graph Clustering via Personalized PageRank

Recently, many models based on the combination of graph convolutional networks and deep learning have attracted extensive attention for their superior performance in graph clustering tasks. However, the existing models have the following limitations: (1) Existing models are limited by the calculation method of graph convolution, and their computational cost will increase exponentially as the graph scale grows. (2) Stacking too many convolutional layers causes the over-smoothing issue and neglects the local graph structure. (3) Expanding the range of the neighborhood and the model depth together is difficult due to the orthogonal relationship between them. Inspired by personalized pagerank and auto-encoder, we conduct the node-wise graph clustering task in the undirected simple graph as the research direction and propose a Scalable Deep Network (SDN) for graph clustering via personalized pagerank. Specifically, we utilize the combination of multi-layer perceptrons and linear propagation layer based on personalized pagerank as the backbone network (i.e., the Quasi-GNN module) and employ a DNN module for auto-encoder to learn different dimensions embeddings. After that, SDN combines the two embeddings correspondingly; then, it utilizes a dual self-supervised module to constrain the training of the embedding and clustering process. Our proposed Quasi-GNN module reduces the computational costs of traditional GNN models in a decoupled approach and solves the orthogonal relationship between the model depth and the neighborhood range. Meanwhile, it also alleviates the degraded clustering effect caused by the over-smoothing issue. We conducted experiments on five widely used graph datasets. The experimental results demonstrate that our model achieves state-of-the-art performance.