Scalable K-Means++

Bahmani, Bahman; Moseley, Benjamin; Vattani, Andrea; Kumar, Rajiv; Vassilvitskii, Sergei

doi:10.48550/arxiv.1203.6402

Cited by 16 publications

(19 citation statements)

References 0 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…al. [12] is one of the most popular clustering techniques. This algorithm divides the network into k clusters, where each cluster is defined by a reference node (centroid).…”

Section: Literature Reviewmentioning

confidence: 99%

NICASN: Non-negative Matrix Factorization and Independent Component Analysis for Clustering Social Networks

Tadi¹,

Rueda²,

Alhadidi³

2022

Proceedings of the Canadian Conference on Artificial Intelligence

View full text Add to dashboard Cite

Discovering clusters in social networks is of fundamental and practical interest. This paper presents a novel clustering strategy for large-scale highly-connected social networks. We propose a new hybrid clustering technique based on non-negative matrix factorization and independent component analysis for finding complex relationships among users of a huge social network. We extract the important features of the network and then perform clustering on independent and important components of the network. Moreover, we introduce a new k-means centroid initialization method by which we achieve higher efficiency. We apply our approach on four well-known social networks: Facebook, Twitter, Academia and Youtube. We experimentally show that our approach achieves much better results in terms of the Silhouette coefficient compared to well-known counterparts such as Hierarchical Louvain, Multiple Local Community detection, and k-means++.

show abstract

“…al. [12] is one of the most popular clustering techniques. This algorithm divides the network into k clusters, where each cluster is defined by a reference node (centroid).…”

Section: Literature Reviewmentioning

confidence: 99%

NICASN: Non-negative Matrix Factorization and Independent Component Analysis for Clustering Social Networks

Tadi¹,

Rueda²,

Alhadidi³

2022

Proceedings of the Canadian Conference on Artificial Intelligence

View full text Add to dashboard Cite

show abstract

“…The Point-Net and ResNet-50 network weights are updated accordingly. After MoCo v2 training, k-means clustering [3] is performed on the trained 3072-d object segment features (with C clusters). Each object segment is assigned with the corresponding cluster index as its initial category label, y i ∈ {1, 2, ..., C}.…”

Section: Iterative Segment Labelingmentioning

confidence: 99%

“…Summary Figure 1 illustrates the iterative segment labeling process. First, we get the initial guess of the labels {y i } n i=1 via k-means clustering [3] based on features learned with self-supervised training. Then, the network training and segment labeling process are applied iteratively for several rounds.…”

Section: Segment Labeling Network Design and Trainingmentioning

confidence: 99%

“…the annotated ground-truths. However, we have no idea which discovered cluster 3 in training corresponds to which semantic category annotated, without annotations or human examination. Besides, it is also possible that a discovered cluster captures just a subgroup of an annotated category.…”

Section: Object Detector Training and Evaluationmentioning

confidence: 99%

“…During the MoCo feature learning, all hyperparameters follow the MoCo v2 [16]. After that, k-means clustering [3] is performed on the 3072-d features, which is the concatenation of the 1024-d 3D shape feature and 2048-d 2D image appearance feature.…”

Section: A2 Implementation Detailsmentioning

confidence: 99%

See 2 more Smart Citations

Unsupervised Object Detection with LiDAR Clues

Huang¹,

Chen²,

Dai³

et al. 2020

Preprint

View full text Add to dashboard Cite

Despite the importance of unsupervised object detection, to the best of our knowledge, there is no previous work addressing this problem. One main issue, widely known to the community, is that object boundaries derived only from 2D image appearance are ambiguous and unreliable.To address this, we exploit LiDAR clues to aid unsupervised object detection. By exploiting the 3D scene structure, the issue of localization can be considerably mitigated. We further identify another major issue, seldom noticed by the community, that the long-tailed and open-ended (sub-)category distribution should be accommodated. In this paper, we present the first practical method for unsupervised object detection with the aid of LiDAR clues. In our approach, candidate object segments based on 3D point clouds are firstly generated. Then, an iterative segment labeling process is conducted to assign segment labels and to train a segment labeling network, which is based on features from both 2D images and 3D point clouds. The labeling process is carefully designed so as to mitigate the issue of long-tailed and open-ended distribution. The final segment labels are set as pseudo annotations for object detection network training. Extensive experiments on the large-scale Waymo Open dataset suggest that the derived unsupervised object detection method achieves reasonable accuracy compared with that of strong supervision within the LiDAR visible range. Code shall be released.* Equal contribution. † This work is done when Hao Tian and Yuntao Chen are interns at SenseTime Research.

show abstract

Spatiotemporal variability of soil moisture over Ethiopia and its teleconnections with remote and local drivers

Jimma

Demissie

Diro

et al. 2023

Theor Appl Climatol

View full text Add to dashboard Cite

Soil moisture is one of the essential climate variables with a potential impact on local climate variability. Despite the importance of soil moisture, studies on soil moisture characteristics in Ethiopia are less documented. In this study, the spatiotemporal variability of Ethiopian soil moisture (SM) has been characterized, and its local and remote influential driving factors are investigated. An empirical orthogonal function (EOF) and KMeans clustering algorithm have been employed to classify the large domain into homogeneous zones. Complex maximum covariance analysis (CMCA) is applied to evaluate the covariability between SM and selected local and remote variables such as rainfall (RF), evapotranspiration (ET), and sea surface temperature (SST). Inter-comparison among SM datasets highlight that the FLDAS dataset better depicts the country’s SM spatial and temporal distribution (i.e., a correlation coefficient $$r=0.95$$ r = 0.95 , $$rmsd=0.04{\mathrm{m}}^{3}{\mathrm{m}}^{-3}$$ r m s d = 0.04 m 3 m - 3 with observations). Results also indicate that regions located in northeastern Ethiopia are drier irrespective of the season (JJAS, MAM, and OND) considered. In contrast, the western part of the country consistently depicted a wetter condition in all seasons. During summer (JJAS), the soil moisture variability is characterized by a strong east–west spatial contrast. The highest and lowest soil moisture values were observed across the country’s central western and eastern parts, respectively. Furthermore, analyses indicate that interannual variability of SM is dictated substantially by RF, though the impact on some regions is weaker. It is also found that ET likely drives the SM in the eastern part of Ethiopia due to a higher atmospheric moisture demand that ultimately invokes changes in surface humidity and rainfall. A composite analysis based on the extreme five wettest and driest SM years revealed a similar spatial distribution of wet SM with positive anomalies of RF across the country and ET over the southern regions. Remote SSTs are also found to have a significant influence on SM distribution. In particular, equatorial central Pacific and western Indian oceans SST anomalies are predominant factors for spatiotemporal SM variations over the country. Major global oceanic indices: Oceanic Nino Index (ONI), Indian Ocean Dipole (IOD), Pacific warm pool (PACWARMPOOL), and Pacific Decadal Oscillations (PDO) are found to be closely associated with the SM anomalies in various parts of the country. The associationship between these remote SST anomalies and local soil moisture is via large-scale atmospheric circulations that are linked to regional factors such as precipitation and temperature anomalies.

show abstract

Scalable K-Means++

Cited by 16 publications

References 0 publications

NICASN: Non-negative Matrix Factorization and Independent Component Analysis for Clustering Social Networks

NICASN: Non-negative Matrix Factorization and Independent Component Analysis for Clustering Social Networks

Unsupervised Object Detection with LiDAR Clues

Spatiotemporal variability of soil moisture over Ethiopia and its teleconnections with remote and local drivers

Contact Info

Product

Resources

About