Improving Performance of Multiclass Classification by Inducing Class Hierarchies

Silva-Palacios, Daniel; Ferri, Cèsar; Ramírez-Quintana, Marïa José

doi:10.1016/j.procs.2017.05.218

Cited by 56 publications

(35 citation statements)

References 12 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…This higher RAM, however, is reserved for a shorter period of time (CPU usage), consequently mitigating the memory overhead caused by the parallel execution of the distributed models. These results are in general aligned with the fact that multiclass classification problems are harder to solve than their binary decompositions, and hence a larger MLP topology may be needed for the multiclass classifiers, at the expense of higher CPU and RAM requirements [33][34][35].…”

Section: Comparing the Centralized And Distributed Modelssupporting

confidence: 55%

“…Such a study, however, constitutes an interesting future direction for providing comparative results of the possible ML methods that can be applied. Note that in general, despite the ML method applied, multiclass classification problems (centralized framework) are harder to solve than their binary decompositions [33][34][35] (distributed framework). Hence, even though different ML methods are expected to perform differently regarding their achievable accuracy and RAM and CPU requirements, the centralized approach will still be harder to solve (with higher CPU and RAM requirements) than its distributed decompositions.…”

Section: Ml-based Qot Estimationmentioning

confidence: 99%

“…• We train both the centralized and distributed QoT models according to neural network (NN) topologies consisting of different numbers of hidden layers and units. Specifically, as the centralized framework is based on a multiclass classification problem, which is in general harder to solve than its binary decompositions [33][34][35] (distributed QoT problem), the centralized framework is examined for a larger NN topology.…”

Section: Introductionmentioning

confidence: 99%

See 2 more Smart Citations

Decentralizing machine-learning-based QoT estimation for sliceable optical networks

Panayiotou

Savva

Tomkos

et al. 2020

J. Opt. Commun. Netw.

View full text Add to dashboard Cite

Dynamic network slicing has emerged as a promising and fundamental framework for meeting 5G's diverse use cases. As machine learning (ML) is expected to play a pivotal role in the efficient control and management of these networks, in this work, we examine the ML-based quality-of-transmission (QoT) estimation problem under the dynamic network slicing context, where each slice has to meet a different QoT requirement. Specifically, we examine ML-based QoT frameworks with the aim of finding QoT model/s that are fine-tuned according to the diverse QoT requirements. Centralized and distributed frameworks are examined and compared according to their model accuracy, routing and spectrum allocation (RSA) accuracy, and CPU (training time) and RAM (memory) requirements. We show that the distributed QoT models outperform the centralized QoT model in accuracy and CPU usage. The RSA accuracy, i.e., measuring the accuracy of the models with regard to the QoT-aware RSA decisions, is sufficiently high for both frameworks. Regarding the RAM usage, as the distributed framework has to train in parallel several QoT models, it may require higher memory, especially as the number of diverse QoT requirements increases. This memory, however, tends to be reserved for a shorter period of time. Moreover, this work develops a dynamic multi-slice QoT-aware (RSA) framework that integrates the ML-based QoT models. The aim is to examine the network performance when the diverse QoT models are considered, as opposed to the state-of-the-art single-slice QoT-aware RSA approach where all connections/slices are provisioned according to a single QoT requirement. We show that the multi-slice QoT-aware RSA approach significantly improves network performance, a clear indicator that the commonly considered single-slice QoT-aware RSA approach may lead to connection overprovisioning.few years adopted network slicing as an emerging technology and business model [2,3].In general, the implementation of network slicing is conceived as an end-to-end feature spanning over different technology domains. While relevant work focuses mostly on the implementation of network slicing in the radio access network (RAN) segment [1,2,4,5], recent work also focuses on extending the scope of network slicing beyond the RAN to include both mobile and transport (optical) network segments [6][7][8]. In particular, recent work deals with SDN/NFV demonstrations, network orchestration implementations, and traffic demand analysis, towards the vision of end-to-end network slicing. In general, SDN/NFV, network orchestration, and analytics are considered as the key ingredients for successful implementation of end-to-end network slicing [1]. However, regarding the efficient control and management of these slices (i.e., enabling not only their creation but also their 1943-0620/20/070146-17 Journal smart deployment at the right place, at the right time, with optimized resources), little has been done so far.Without doubt, machine learning (ML) will be instrumental in the efficient control an...

show abstract

Section: Comparing the Centralized And Distributed Modelssupporting

confidence: 55%

Section: Ml-based Qot Estimationmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Decentralizing machine-learning-based QoT estimation for sliceable optical networks

Panayiotou

Savva

Tomkos

et al. 2020

J. Opt. Commun. Netw.

View full text Add to dashboard Cite

show abstract

“…Decision Tree is used for classifying the data while Regression Tree is used for predicting continuous variables. The Decision Tree algorithm has the potential of obtaining better result than other machine learning methods on multi-class prediction due to its tree-like model (Silva-Palacios et al, 2017). A decision tree is a tree where each node represents an independent variable, each link (branch) represents a condition/decision (rule) and each leaf represents an outcome(i.e.…”

Section: Decision Treementioning

confidence: 99%

A Machine Learning Approach on Occupant Number Prediction for Indoor Spaces

Işıkdağ¹,

Sahin²,

Cansiz³

2018

Int. Arch. Photogramm. Remote Sens. Spatial Inf. Sci.

View full text Add to dashboard Cite

<p><strong>Abstract.</strong> The knowledge about the occupancy of an indoor space can serve to various domains ranging from emergency response to energy efficiency in buildings. The literature in the field presents various methods for occupancy detection. Data gathered for occupancy detection, can also be used to predict the number of occupants at a certain indoor space and time. The aim of this research was to determine the number of occupants in an indoor space, through the utilisation of information acquired from a set of sensors and machine learning techniques. The sensor types used in this research was a sound level sensor, temperature/humidity level sensor and an air quality level sensor. Based on data acquired from these sensors six automatic classification techniques are employed and tested with the aim of automatically detecting the number of occupants in an indoor space by making use of multi-sensor information. The results of the tests demonstrated that machine learning techniques can be used as a tool for prediction of number of occupants in an indoor space.</p>

show abstract

“…Based on these observations, we examined frequently altered genes in large tumor cohorts by systematically interrogating mutation subgroupings associated with improved classification performance. Instead of focusing on a single gene or pathway of interest, we sought to create a framework inspired by class-grouping approaches [28][29][30] which would generalize well to the population of somatic alterations recurrent in cancer. Specifically, we test each gene for the existence of at least one good multinomial classifier by searching over a hierarchy of one-vs-rest binary classifiers.…”

Section: Introductionmentioning

confidence: 99%

Systematic interrogation of mutation groupings reveals divergent downstream expression programs within key cancer genes

Grzadkowski

Manning

Somers

et al. 2020

Preprint

View full text Add to dashboard Cite

6Genes implicated in tumorigenesis often exhibit diverse sets of genomic variants in the tumor 7 cohorts within which they are frequently mutated. We sought to identify the downstream expres-8 sion effects of these perturbations and to find whether or not this heterogeneity at the genomic 9 level is reflected in a corresponding heterogeneity at the transcriptomic level. Applying a novel 10 hierarchical framework for organizing the mutations present in a cohort along with machine learn-11 ing pipelines trained on sample expression profiles we systematically interrogated the signatures 12 associated with combinations of perturbations recurrent in cancer. This allowed us to catalogue 13 the mutations with discernible downstream expression effects across a number of tumor cohorts 14 as well as to uncover and characterize a multitude of cases where subsets of a genes mutations are 15 clearly divergent in their function from the remaining mutations of the gene. 16 Each tumor faces a common set of obstacles arising from internal dynamics and external de-18 fense mechanisms 1 . Tumor cohorts, however, are replete with diverse yet recurrent tactics for 19 overcoming these shared obstacles. Tumorigenesis can thus be perceived as a landscape within 20 which each tumor navigates a unique, multidimensional path, weaving between segments trodden 21 by other tumors. A number of the early breakthroughs in cancer treatment directly resulted from 22 coarse demarcations of these paths into distinct subtypes based on "landmarks"-usually defined 23 by mutations and/or markers derived from proteomic or transcriptomic data-that were then used 24 to engineer subtype-specific treatments [2][3][4] . 25 Although these biomarker-based treatment matching criteria have proven effective in some 26 precision medicine applications, there is a sizable subset of patients whose tumors harbor no dis-27 cernible drug targets, thus diminishing their likelihood of successful treatment and survival [5][6][7][8] . 28 Developing a more thorough understanding of the downstream effects of landmark events could 29 therefore improve tailored treatment design outcomes. In particular, we envision a tactic which de-30 tects whether two genomic alterations (or combinations thereof) have a shared downstream effect, 31 and can therefore be grouped together when weighing treatment options. This type of approach 32 should also be able to detect whether two such alterations or groupings result in divergent tran-33 scriptional programs and can therefore be considered distinct. Despite recent efforts to profile the 34 downstream effects of mutations recurrent in cancer, for most mutations we still know little about 35 the programs they trigger. As a result, most clinical guidelines depend on only a limited sub-36 set of specific perturbations within a gene or on other coarse biomarker-based demarcations 9,10 . 37A clearer discernment of the convergences and divergences between the downstream programs 38 present within cancer genes is thus a crucial prerequisi...

show abstract

Improving Performance of Multiclass Classification by Inducing Class Hierarchies

Cited by 56 publications

References 12 publications

Decentralizing machine-learning-based QoT estimation for sliceable optical networks

Decentralizing machine-learning-based QoT estimation for sliceable optical networks

A Machine Learning Approach on Occupant Number Prediction for Indoor Spaces

Systematic interrogation of mutation groupings reveals divergent downstream expression programs within key cancer genes

Contact Info

Product

Resources

About