Scribd
Upload
7 views5 pages

Notes on MiDA

Uploaded by

hapet27813
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
7 views5 pages

Notes on MiDA

Uploaded by

hapet27813
Copyright
© © All Rights Reserved
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 5

Detailed Notes on "A Multi-View Deep Learning Approach for Predictive

Business Process Monitoring"

Authors:

 Vincenzo Pasquadibisceglie
 Annalisa Appice
 Giovanna Castellano
 Donato Malerba

Abstract

 Objective: The article addresses the task of predicting the next trace activity in a
running trace based on historical event logs. This predictive capability can help
companies act proactively and improve their business processes.
 Approach: Combines multi-view learning and deep learning to enhance predictive
accuracy by considering the variety of information recorded in event logs.
 Results: Experiments with various benchmark event logs show the effectiveness of
the proposed approach compared to other state-of-the-art methods.

Introduction

 Event Logs: Collect business process executions recorded as traces, which are
ordered lists of events.
 Predictive Business Process Monitoring: Analyzing event log data to predict future
events in a process. This includes predicting the next activity, completion cycle time,
and trace outcomes.
 Importance: Predicting the next activity in a running trace allows analysts to
intervene proactively and prevent undesired behaviors, thereby improving business
processes.
 MiDA Framework: The proposed Multi-view Deep learning based approach for next
Activity prediction (MiDA) extracts a predictive model from event log data and
applies it to real-time event data to predict future trace behavior.

Proposed MiDA Framework (Page 2)

 Input Representation: Event logs include activities, timestamps, and resources,


treated as multiple views of the same trace.
 Architecture:
o Separate neural networks process each view independently.
o A concatenation layer merges these views.
o A final neural network learns the predictive model from the combined views.
 Training Process: Involves minimizing the cross-entropy loss using mini-batch
stochastic gradient descent (SGD) with adaptive moment estimation (Adam) as the
optimization algorithm.
 Evaluation:
o Metrics used include accuracy, precision, recall, F-score, area under the curve
(AUC), and area under the precision-recall curve (AUCPR).
o Results show MiDA outperforms other methods in terms of F-score,
demonstrating its effectiveness in predicting the next activity in a running
trace.
 Feasibility and Comparison: The multi-view learning nature of MiDA is compared
with a counterpart approach called MiDAPre, highlighting MiDA's superior
performance.

Benchmark Event Logs (Page 3)

 Datasets: BPI 2012, BPI 2013, and BPI 2014, which are widely used in the process
mining community.
 Characteristics: Description of each dataset including the number of traces, number
of events, and average trace length.
 Preprocessing: Steps include filtering out incomplete traces and splitting the data into
training and testing sets.

Experimental Setup (Page 4)

 Evaluation Strategy: A 10-fold cross-validation strategy is used, where the event


logs are divided into 10 subsets, and the model is trained and tested on each subset.
 Hyperparameters Optimization: Hyperparameters such as the number of hidden
units, learning rate, and dropout rate are optimized using a Bayesian optimizer.
 Evaluation Metrics: Metrics include accuracy, precision, recall, F-score, AUC, and
AUCPR.
 Baseline Methods for Comparison: Includes LSTM, CNN, and SVM models.

Empirical Evaluation Results (Page 5)

 Performance Comparison: MiDA is compared to baseline methods across multiple


metrics.
 Results: MiDA achieves higher scores in terms of accuracy, precision, recall, F-score,
AUC, and AUCPR.
 Discussion: Highlights the strengths and limitations of the proposed approach,
analyzing the impact of factors such as the number of views and the length of prefix
traces on MiDA’s performance.

Conclusion (Page 6)

 Effectiveness: The MiDA framework effectively predicts the next activity in a


running trace.
 Advantages: The multi-view learning approach allows for the integration of diverse
information recorded in event logs.
 Limitations: The research acknowledges the need for further investigation into the
scalability of the approach and its generalizability to different domains.
 Future Work: Suggests potential directions for extending and improving the
proposed approach.
Related Work (Page 7)

 Predictive Analytic Approaches: Discusses various works related to predictive


analytics in process management, classified based on the type of target variable
predicted (continuous versus discrete).
 Explicit Representations: Approaches that use explicit process models such as state-
transition models and hidden Markov models, which rely on prior assumptions about
the model and have limitations in considering diverse information observed in event
logs.
 Implicit Model Representations: Recent studies using deep learning approaches
with neural network architectures that process activities and timestamps but pay little
attention to additional data perspectives.
 Multi-View Learning: Discusses the relevance of multi-view learning in improving
predictive accuracy by simultaneously processing multiple perspectives (e.g., activity,
time, resource).

MiDA Framework Introduction (Page 8)

 Limitations of Existing Approaches: Current methods face issues related to


scalability, flexibility, and accuracy.
 Embedding Dimension Analysis: Analysis of optimal embedding dimensions for the
MiDA framework.
 Framework Architecture: Detailed description of the MiDA architecture,
highlighting the use of separate neural networks for each view followed by a
concatenation layer and a final predictive model.
 Comparison with Existing Approaches: Emphasizes the advantages of MiDA's
multi-view deep learning approach over traditional methods.

Deep Learning Trends (Page 9)

 Recent Trends: Use of deep learning with implicit model representations in


predictive process monitoring.
 Limitations: Some deep learning approaches face challenges such as overfitting and
high computational costs.
 Multi-View Learning: Reintroduction of the concept, emphasizing its benefits in
processing multiple perspectives simultaneously.
 Empirical Validation: Presentation of empirical results validating the effectiveness
of the MiDA framework.
 Evaluation of View Combinations: Impact of different combinations of views on the
performance of the MiDA framework.

Performance Comparison and Conclusions (Page 10)

 Performance Comparison: MiDA's performance is compared with several state-of-


the-art methods, showing superior results.
 Results Discussion: Detailed discussion on the results, highlighting MiDA’s
strengths in predictive accuracy.
 Summary and Contributions: Recap of the main findings and contributions of the
research.
 Research Limitations: Discussion on areas needing improvement.
 Future Directions: Suggestions for further research and potential improvements.

Explainability and Experimental Setup (Page 11)

 Explainability: Introduction to the concept of explainability in predictive business


process mining, highlighting its importance in making predictive models more
interpretable.
 Outcome-Oriented Predictive Monitoring: Introduction to the use of image
encoding and convolutional neural networks (CNNs) for outcome-oriented predictive
process monitoring.
o ORANGE Framework: Describes the ORANGE framework, which predicts
the outcome of a running process instance based on the visual representation
of the process.
o Advantages: Discusses the benefits of using image encoding and CNNs for
monitoring process outcomes.
 Multi-View Learning: Explains the relevance of multi-view learning in predictive
process mining, highlighting challenges such as feature selection and model fusion.

Evaluation Metrics and ORANGE Framework (Page 12)

 Evaluation Metrics: Detailed explanation of the metrics used to assess performance,


including accuracy, precision, recall, F-score, AUC, and AUCPR.
 Future Research Directions: Discussion on potential research directions for
enhancing predictive process monitoring with explanation capabilities.
 ORANGE Framework:
o Process: Conversion of process event logs into images and the architecture of
the CNNs used.
o Results: Presentation of empirical evaluation outcomes for the ORANGE
framework.
 MiDA Framework:
o Experimental Setup: Detailed setup and evaluation on different view
variants.
o Results: Performance on benchmark event logs and comparison with several
baseline methods.
o Factors Impact: Analysis of various factors affecting MiDA's performance.

Evaluation and Results (Page 13)

 MiDA Evaluation:
o Metrics Explanation: Detailed computation of evaluation metrics used for
assessing performance.
o Experimental Setup: Detailed description of the experimental setup for both
ORANGE and MiDA frameworks.
o Benchmark Event Logs: Use of BPI 2012, BPI 2013, and BPI 2014 datasets,
including preprocessing steps applied.
o Results: Presentation of empirical results for different view variants and their
impact on performance.
o Discussion: Detailed analysis of the experimental results, emphasizing the
impact of various factors on MiDA’s performance.
Significance and Implications (Page 14)

 Experimental Results Significance: Discussion on the implications of the


experimental findings for predictive process mining.
 ORANGE Framework Contributions: Highlighting the significance and
contributions of the ORANGE framework.
 MiDA Framework Strengths and Limitations: Analysis of MiDA’s strengths and
areas needing improvement based on empirical evaluation.
 Implications for Predictive Business Process Monitoring: Discussion on the
broader implications of the research findings.
 Conclusion: Emphasizes the MiDA framework's ability to handle raw data from any
view without additional implementation effort, distinguishing it from existing
baselines.
 Future Research: Suggestions for addressing challenges such as activity imbalance
and further extending the MiDA framework’s capabilities.

You might also like