Improving the explainability of Random Forest classifier – user centered approach

Petković, Dragutin; Altman, Russ B.; Wong, Mike; Vigil, Arthur

doi:10.1142/9789813235533_0019

Cited by 58 publications

(46 citation statements)

References 14 publications

Supporting

Mentioning

Contrasting

Unclassified

Order By: Relevance

“…A number of researchers have leveraged RF's ability to determine feature importance for addressing the explainability challenge or to transform RF into sets of rules, as covered in [16] and also applied in e.g. [21,22].…”

Section: Our Approach To Random Forest Model and Sample Explainabilitmentioning

confidence: 99%

“…Specific RFEX design goals were developed in consultation and conversations with our target users namely domain experts (but not ML experts) attempting to use ML technologies and were initially evaluated in [16] with 13 users. RFEX design goals include: a) user centered design: we specifically focused on developing explainers that are driven by the specific needs of our intended users; and b) simplicity and familiarity: we enforced simplicity and familiarity with common ways and measures our users analyze the data such as medical tests e.g.…”

Section: Our Approach To Random Forest Model and Sample Explainabilitmentioning

confidence: 99%

“…RFEX Model Explainer presented in this paper is a significant improvement of the original RFEX approach published in [16]. It provides one page RFEX Model Summary table using the following seven steps (steps 1-3 are the same as in [16], steps 4-7 are new). These steps directly answer users' model explainability questions 1-5 as outlined above but use some novel, more standard and institutive measures than the original RFEX.…”

Section: Rfex Model Explainermentioning

confidence: 99%

“…"User in the loop" paradigm for improvement of ML systems are gaining attention [4] with a good example in [26] where authors demonstrate benefits of this approach (they call it "explanatory debugging") to improving the quality of ML solution and increasing user trust, with explainability as its very important component. Jointly with Stanford Bioengineering we have been working for a number of years in applying ML to biomedical problems [16,17,18] and naturally we got involved in ML explainability, especially of well known Random Forest (RF) ML method [13]. We originally developed RF Explainability Enhancement pipeline (RFEX) as RF Model explainer and applied it to Stanford FEATURE data [16] where we also performed usability experiment with 13 users and demonstrated that RFEX increased their understanding of RF classification.…”

Section: Introductionmentioning

confidence: 99%

“…Jointly with Stanford Bioengineering we have been working for a number of years in applying ML to biomedical problems [16,17,18] and naturally we got involved in ML explainability, especially of well known Random Forest (RF) ML method [13]. We originally developed RF Explainability Enhancement pipeline (RFEX) as RF Model explainer and applied it to Stanford FEATURE data [16] where we also performed usability experiment with 13 users and demonstrated that RFEX increased their understanding of RF classification. In this paper we present several key contributions: a) significant improvements in RFEX Model explainer compared to the previously published version [16]; b) a new RFEX Sample explainer which provides simple explanation of how RF classifies a particular data sample and is designed to directly relate to RFEX Model explainer; and c) an RFEX Model and Sample explainer case study on the data and findings from J. Craig Venter Institute (JCVI) and Allen Institute for Brain Science [21,22,27] and on the Stanford FEATURE data described in [16].…”

Section: Introductionmentioning

confidence: 99%

See 4 more Smart Citations

RFEX: Simple Random Forest Model and Sample Explainer for non-Machine Learning experts

Petković

Alavi

Cai

et al. 2019

Preprint

View full text Add to dashboard Cite

Machine Learning (ML) is becoming an increasingly critical technology in many areas. However, its complexity and its frequent non-transparency create significant challenges, especially in the biomedical and health areas. One of the critical components in addressing the above challenges is the explainability or transparency of ML systems, which refers to the model (related to the whole data) and sample explainability (related to specific samples). Our research focuses on both model and sample explainability of Random Forest (RF) classifiers. Our RF explainer, RFEX, is designed from the ground up with non-ML experts in mind, and with simplicity and familiarity, e.g. providing a one-page tabular output and measures familiar to most users. In this paper we present significant improvement in RFEX Model explainer compared to the version published previously, a new RFEX Sample explainer that provides explanation of how the RF classifies a particular data sample and is designed to directly relate to RFEX Model explainer, and a RFEX Model and Sample explainer case study from our collaboration with the J. Craig Venter Institute (JCVI). We show that our approach offers a simple yet powerful means of explaining RF classification at the model and sample levels, and in some cases even points to areas of new investigation. RFEX is easy to implement using available RF tools and its tabular format offers easy-to-understand representations for non-experts, enabling them to better leverage the RF technology.

show abstract

Section: Our Approach To Random Forest Model and Sample Explainabilitmentioning

confidence: 99%

Section: Our Approach To Random Forest Model and Sample Explainabilitmentioning

confidence: 99%

Section: Rfex Model Explainermentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

Section: Introductionmentioning

confidence: 99%

See 3 more Smart Citations

RFEX: Simple Random Forest Model and Sample Explainer for non-Machine Learning experts

Petković

Alavi

Cai

et al. 2019

Preprint

View full text Add to dashboard Cite

show abstract

Understanding high‐involvement product purchase through an innovative machine learning approach: A case of housing type choice

Arachchige

Quach

Roca

et al. 2022

J of Consumer Behaviour

View full text Add to dashboard Cite

Adopting the conservation of resources theory, this research explores the joint influence of multiple factors affecting the choice of housing, an instance of a high involvement product. We develop an innovative Machine Learning approach to identify the most significant set of factors affecting consumers' housing choices. It was found that housing choices were primarily bound by energy resources and secondarily by other personal, conditional, and object resources. The study identified 25 key factors, many of which previous studies have not explored. The new factors include government payments, vehicle possession, time living with children, marriage duration, current location duration, and health conditions. The results suggest that joint effects of factors are more prominent than individual effects in influencing the choice of high‐involvement products. Further, the study suggests that ML methods are more robust than traditional methods and can be applied to analyse other types of high involvement products. The findings can assist real estate investors, policymakers and other stakeholders to understand sophisticated market behaviours to develop better and tailored housing strategies.

show abstract

DNA methylation‐based profiling of bone and soft tissue tumours: a validation study of the ‘DKFZ Sarcoma Classifier’

Lyskjær

Noon

Tirabosco

et al. 2021

The Journal of Pathology CR

View full text Add to dashboard Cite

Diagnosing bone and soft tissue neoplasms remains challenging because of the large number of subtypes, many of which lack diagnostic biomarkers. DNA methylation profiles have proven to be a reliable basis for the classification of brain tumours and, following this success, a DNA methylation-based sarcoma classification tool from the Deutsches Krebsforschungszentrum (DKFZ) in Heidelberg has been developed. In this study, we assessed the performance of their classifier on DNA methylation profiles of an independent data set of 986 bone and soft tissue tumours and controls. We found that the 'DKFZ Sarcoma Classifier' was able to produce a diagnostic prediction for 55% of the 986 samples, with 83% of these predictions concordant with the histological diagnosis. On limiting the validation to the 820 cases with histological diagnoses for which the DKFZ Classifier was trained, 61% of cases received a prediction, and the histological diagnosis was concordant with the predicted methylation class in 88% of these cases, findings comparable to those reported in the DKFZ Classifier paper. The classifier performed best when diagnosing mesenchymal chondrosarcomas (CHSs, 88% sensitivity), chordomas (85% sensitivity), and fibrous dysplasia (83% sensitivity). Amongst the subtypes least often classified correctly were clear cell CHSs (14% sensitivity), malignant peripheral nerve sheath tumours (27% sensitivity), and pleomorphic liposarcomas (29% sensitivity). The classifier predictions resulted in revision of the histological diagnosis in six of our cases. We observed that, although a higher tumour purity resulted in a greater likelihood of a prediction being made, it did not correlate with classifier accuracy. Our results show that the DKFZ Classifier represents a powerful research tool for exploring the pathogenesis of sarcoma; with refinement, it has the potential to be a valuable diagnostic tool.

show abstract

Improving the explainability of Random Forest classifier – user centered approach

Cited by 58 publications

References 14 publications

RFEX: Simple Random Forest Model and Sample Explainer for non-Machine Learning experts

RFEX: Simple Random Forest Model and Sample Explainer for non-Machine Learning experts

Understanding high‐involvement product purchase through an innovative machine learning approach: A case of housing type choice

DNA methylation‐based profiling of bone and soft tissue tumours: a validation study of the ‘DKFZ Sarcoma Classifier’

Contact Info

Product

Resources

About