Use of Explainable AI to Re ne Arti cial Immune System Algorithms?

Abstract

Arti cial Immune System (AIS) algorithms o er distributed Arti cial Intelligence (AI) that can be e ectively used to solve real-world data segregation problems. Current unmet needs include (i) re ning the AIS algorithm using representative patterns based on deterministic features of the data and (ii) interpretability of the AIS based classi ers. Model agnostic eXplainable AI (XAI) approaches may cater to both of these unmet needs. Here, we propose building an AIS based classi er pipeline using XAI algorithms to identify deterministic feature patterns from data that can be used to improve the performance and interpretability of AIS classi ers.

Keywords

• AIS
• NSA
• CSPRA
• DCA
• XAI
• LIME
• SHAP

Introduction

Arti cial Immune System (AIS) is a branch of biologically-inspired computation for developing computational models and methodologies based on the principles of the biological immune system (BIS). Self discrimination by distinguishing own body cells (self) from foreign or transformed entities (non-self) is one of the primary abilities of BIS [ 1 ]. AIS is relatively a young eld of research where each molecular player of BIS can potentially inspire at least one AIS model. The molecular function of BIS, including self-discrimination, feature extraction, pattern recognition, learning, memory, and distributive yet cooperative nature, have been adapted in AIS to develop immune-inspired models. The AIS models have a wide range of applications such as anomaly detection, pattern recognition, clustering bases, data analyses, function optimisation, and computer security [ 1 ]. Basic AIS algorithms such as the Negative Selection Algorithm (NSA) that di erentiates self (normal) from non-self (abnormal) are ideal for real-world problems where normal data is well de ned, and any alterations can be deemed as anomalies. For example, cancer detection or labelling device performance as good or bad.

Although useful, the simple NSA algorithm su ers from a high false-positive rate [ 2 ]. The proposed re nements include extending NSA to include an additional check for pattern recognition. Brie y, in addition to the non-self detectors, NSA can be re ned as Conserved Self Pattern Recognition Algorithm (CSPRA) to include self pattern detectors that con rm the non-self classi cation. Such an approach has previously been shown to decrease the false positive rate and improve accuracy [ 2 ]. Similarly, the overall performance of AIS classi ers can be improved by using both self and non-self patterns during training. Dendritic Cell Algorithm (DCA) is one such algorithm that improves performance by including both self and non-self data during classi er training [ 3 ]. AI algorithms, including AIS algorithms, su er from a lack of transparency. Hence, the eXplainable Articial Intelligence (XAI) approaches are essential to improve the understanding and interpretability of AIS classi ers. Among the XAI approaches available, the model agnostic approaches are easier to incorporate into AI/ML pipeline. XAI algorithms such as Local Interpretable Model-agnostic Explanations (LIME) and SHapley Additive exPlanations (SHAP) can be used for the identi cation of deterministic features [ 4 ] [ 5 ] [ 6 ]. Thus, LIME and SHAP based parallel approaches can be used to understand how AIS classi ers segregate data as self and non-self. 2

Proposed Idea

The current unmet need for AIS classi ers is two-fold- (i) re nement of algorithms to include appropriate self and non-self patterns and (ii) apply XAI to improve AIS interpretability. Based on the current limitation of AIS, we propose the use of XAI to identify self and non-self patterns to re ne AIS algorithms. Brie y, the idea includes creating a generalisable classi cation pipeline for datasets with clearly anticipated (normal or self) data. The pipeline will include NSA, CSPRA and DCA as the three AIS algorithms. Each AIS based classi er is used to overcome the limitations of another classi er. For example, CSPRA will be used to reduce the anticipated false positive rate of NSA. Similarly, DCA will be used to overcome the limitation of using only one class to train NSA and CSPRA. Both CSPRA and DCA use data patterns as the deterministic factor for classi cation. We propose the use of LIME and SHAP to identify deterministic features governing the self and non-self classi cation. Once the deterministic features are identi ed, well-informed self and non-self patterns can be generated for CSPRA and DCA. Thus, the XAI directed self and non-self patterns will be used to re ne AIS algorithms to improve the classi er performance and interpretability.

The two signi cant anticipated limitations for the proposed idea include (i) generalisability and (ii) coverage of the self and non-self patterns. Since LIME and SHAP uses only a subset of samples, it is possible that the pattern identi ed from the limited number of samples may not be representative of the entire dataset and therefore not generalisable. Similarly, the number of samples to be analysed by LIME and SHAP may not o er su cient coverage for the entire dataset to identify appropriate self and non-self patterns. Thus, the number of samples analysed using LIME and SHAP with NSA to identify self and non-self patterns will be an anticipated challenge of the proposed approach, which is shown in Figure 1.

Use of Explainable AI to Re ne Arti cial Immune System Algorithms Overall, we propose the combination of XAI methods to re ne AIS algorithm based classi ers as a new way to improve data segregation to tackle real-world problems. Further work on the proposed idea will enable us to identify best approach to select deterministic features for AIS re nements.

References

1. D. Dasgupta and N. L. Fernando , Immunological computation; theory and applications , First. CRC Press Taylor Fransis Group, 2009 .

2. S. Yu and D. Dasgupta , \ Conserved Self Pattern Recognition Algorithm with Novel Detection Strategy Applied to Breast Cancer Diagnosis," J. Artif. Evol. Appl. , vol. 2009 , pp. 1 { 12 , 2009 , doi: 10.1155/ 2009 /130498.

3. J. Greensmith , U. Aickelin , and S. Cayzer , \ Introducing Dendritic Cells as a Novel Immune-Inspired Algorithm for Anomaly Detection .," in Jacob C., Pilat M.L. , Bentley P.J. , Timmis J.I. (eds) Arti cial Immune Systems . ICARIS 2005. Lecture Notes in Computer Science , C. Jacob, M. L. Pilat , P. J. Bentley , and J. I. Timmis , Eds. Berlin, Heidelberg: Springer., 2005 .

4. A. Adadi and M. Berrada , \ Peeking Inside the Black-Box: A Survey on Explainable Arti cial Intelligence (XAI)," IEEE Access , vol. 6 , pp. 52138 { 52160 , 2018 , doi: 10.1109/ACCESS. 2018 . 2870052 .

5. M. T. Ribeiro , S. Singh , and C. Guestrin , \ `Why should i trust you?' Explaining the predictions of any classi er," Proc. ACM SIGKDD Int. Conf. Knowl. Discov. Data Min. , vol. 13 -17-Augu, pp. 1135 { 1144 , 2016 , doi: 10.1145/2939672.2939778.

6. P. Linardatos , V. Papastefanopoulos , and S. Kotsiantis , \ Explainable ai: A review of machine learning interpretability methods," Entropy , vol. 23 , no. 1 , pp. 1 { 45 , 2021 , doi: 10.3390/e23010018.