Table 5. Limitations of ML techniques...

Table 5.

Limitations of ML techniques mapped to objectives and tasks

ML techniques	Limitations	Qualitative accounting objective	Tasks in the accounting process	Source
Transfer learning decision forests and random forests	Poor interpretability	Verifiability and understandability	Adaptability of OCR Account allocation	Dataiku (2017)
Transfer learning decision forests and random forests	Overfitting	Relevance	Adaptability of OCR Account allocation	Dataiku (2017)
Support vector machine	Poor interpretability	Verifiability and understandability	Image classification Validation of document information Removing duplicate entries and linking documents Matching records or record-linkage	Kotsiantis (2007), Karamizadeh et al. (2014), Witten et al. (2016)
Convolutional neural networks	Poor interpretability	Verifiability and understandability	Deep document feature extraction Irregular document layout classification using NLP	Dataiku (2017), Tarawneh et al. (2019)
Convolutional neural networks	Takes a long time to train	Timeliness		Dataiku (2017), Tarawneh et al. (2019)
k-Nearest neighbour	Poor interpretability	Verifiability and understandability	Image classification Text classification	Kotsiantis (2007), Witten et al. (2016)
k-Nearest neighbour	Complex, which makes it slow	Timeliness	Image classification Text classification	Kotsiantis (2007), Witten et al. (2016)
Semi-supervised clustering	Training rate may be slow depending on available labelled data	Timeliness	Text classification	Zheng, Zhou, Deng, et al. (2017)
Naïve Bayes	Requires independent variables	Faithful representation	Validation of document information Removing duplicate entries and linking documents Matching records or record-linkage Account allocation Error detection in financial data and fraud detection	Samoil (2015)
Naïve Bayes	Training set sensitive	Materiality and faithful representation		Witten et al. (2016), Samoil (2015), Larsson and Segerås (2016)
Artificial neural network	Poor interpretability	Verifiability and understandability	Removing duplicate entries and linking documents Matching records or record-linkage	Kotsiantis (2007), Dataiku (2017), SMACC (2017)
Artificial neural network	Computing intensive	Cost-saving		Kotsiantis (2007), Dataiku (2017), SMACC (2017)
Bayesian Belief network	Computing intensive	Cost-saving	Error detection in financial data and fraud detection	Heckerman (2008), Niedermayer (2008)
Bayesian Belief network	Training set sensitive	Materiality and faithful representation	Error detection in financial data and fraud detection	Heckerman (2008), Niedermayer (2008)
Association rules	Excessive output	Relevance	Error detection in financial data and fraud detection	Witten et al. (2016), Kaur (2014)
Association rules	Requires lots of time	Timeliness	Error detection in financial data and fraud detection	Witten et al. (2016), Kaur (2014)
K-means clustering	Requires lots of time	Timeliness	Account allocation Error detection in financial data and fraud detection	Ayodele (2010b), Witten et al. (2016)
Self-organising maps	Computing intensive	Cost-saving	Error detection in financial data and fraud detection	Ayodele (2010b), SMACC (2017)
Self-organising maps	Requires adequate data	Materiality and faithful representation	Error detection in financial data and fraud detection	Ayodele (2010b), SMACC (2017)
Conditional random fields	Trade-off between accuracy, which requires memory and overfitting	Relevance and faithful representation	Report descriptions	Witten et al. (2016), Sutton and Mccallum (2007)

ML techniques	Limitations	Qualitative accounting objective	Tasks in the accounting process	Source
Transfer learning decision forests and random forests	Poor interpretability	Verifiability and understandability	Adaptability of OCR Account allocation	Dataiku (2017)
Transfer learning decision forests and random forests	Overfitting	Relevance	Adaptability of OCR Account allocation	Dataiku (2017)
Support vector machine	Poor interpretability	Verifiability and understandability	Image classification Validation of document information Removing duplicate entries and linking documents Matching records or record-linkage	Kotsiantis (2007), Karamizadeh et al. (2014), Witten et al. (2016)
Convolutional neural networks	Poor interpretability	Verifiability and understandability	Deep document feature extraction Irregular document layout classification using NLP	Dataiku (2017), Tarawneh et al. (2019)
Convolutional neural networks	Takes a long time to train	Timeliness		Dataiku (2017), Tarawneh et al. (2019)
k-Nearest neighbour	Poor interpretability	Verifiability and understandability	Image classification Text classification	Kotsiantis (2007), Witten et al. (2016)
k-Nearest neighbour	Complex, which makes it slow	Timeliness	Image classification Text classification	Kotsiantis (2007), Witten et al. (2016)
Semi-supervised clustering	Training rate may be slow depending on available labelled data	Timeliness	Text classification	Zheng, Zhou, Deng, et al. (2017)
Naïve Bayes	Requires independent variables	Faithful representation	Validation of document information Removing duplicate entries and linking documents Matching records or record-linkage Account allocation Error detection in financial data and fraud detection	Samoil (2015)
Naïve Bayes	Training set sensitive	Materiality and faithful representation		Witten et al. (2016), Samoil (2015), Larsson and Segerås (2016)
Artificial neural network	Poor interpretability	Verifiability and understandability	Removing duplicate entries and linking documents Matching records or record-linkage	Kotsiantis (2007), Dataiku (2017), SMACC (2017)
Artificial neural network	Computing intensive	Cost-saving		Kotsiantis (2007), Dataiku (2017), SMACC (2017)
Bayesian Belief network	Computing intensive	Cost-saving	Error detection in financial data and fraud detection	Heckerman (2008), Niedermayer (2008)
Bayesian Belief network	Training set sensitive	Materiality and faithful representation	Error detection in financial data and fraud detection	Heckerman (2008), Niedermayer (2008)
Association rules	Excessive output	Relevance	Error detection in financial data and fraud detection	Witten et al. (2016), Kaur (2014)
Association rules	Requires lots of time	Timeliness	Error detection in financial data and fraud detection	Witten et al. (2016), Kaur (2014)
K-means clustering	Requires lots of time	Timeliness	Account allocation Error detection in financial data and fraud detection	Ayodele (2010b), Witten et al. (2016)
Self-organising maps	Computing intensive	Cost-saving	Error detection in financial data and fraud detection	Ayodele (2010b), SMACC (2017)
Self-organising maps	Requires adequate data	Materiality and faithful representation	Error detection in financial data and fraud detection	Ayodele (2010b), SMACC (2017)
Conditional random fields	Trade-off between accuracy, which requires memory and overfitting	Relevance and faithful representation	Report descriptions	Witten et al. (2016), Sutton and Mccallum (2007)

Source: Compiled by author from multiple sources as indicated above

[ViewLarge]