Arsim Susuri

Arsim Susuri

Auto-generated profile, based on publicly available information

Arsim Susuri

Faculty of Computer Science, University “Ukshin Hoti” Prizren, Shkronjat, 20000 Prizren, Kosovo

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
17 December 2024 in International Journal of Interactive Mobile Technologies (iJIM)
Verified
Journal Article
Open access
PDF file
EFEndrit Fetahi
Endrit Fetahi
MHMentor Hamiti
Mentor Hamiti
ASArsim Susuri
Arsim Susuri
JAJaumin Ajdari
Jaumin Ajdari
XZXhemal Zenuni
Xhemal Zenuni
This paper aims to advance AI-based hate speech (HS) detection in the Albanian language, which is resource-limited in natural language processing (NLP). Addressing the challenge of limited data, we developed a human-annotated dataset of over 11,000 comments, carefully curated from various Albanian social media platforms, containing a substantial number of HS instances. The dataset was annotated using a detailed two-layer taxonomy to capture the complex dimensions of HS. To ensure high-quality annotations, three expert annotators applied a majority voting system, achieving a substantial Fleiss’s kappa coefficient of 0.62, underscoring the reliability and consistency of the annotations. We conducted a comparative analysis of several machine learning (ML) algorithms, including support vector machine (SVM), Naïve Bayes (NB), XGBoost, and random forest (RF), paired with various text vectorisation techniques and pre-processing methods. In binary classification, the NB model with term frequencyinverse document frequency (TF-IDF) vectorization achieved the highest performance, with an F1 score of 0.80. For multiclass classification, XGBoost outperformed other models, achieving an F1 score of 0.77. Interestingly, our experiments revealed that pre-processing steps generally reduced model performance, suggesting that raw text inputs work better for the Albanian language. Through error analysis using local interpretable model-agnostic explanations (LIME), we identified key challenges, such as polysemy and irony, which contributed to misclassifications. To demonstrate the practical applicability of our work, we developed a user-friendly mobile web application based on the best-performing model, providing realtime HS detection with the potential for integration into social media platforms.

22 October 2026 in Internet Technology Letters
Verified
Journal Article
Closed access
LSLamir Shkurti
Lamir Shkurti
MSMennan Selimi
Mennan Selimi
ASArsim Susuri
Arsim Susuri

31 March 2026 in International Journal of Interactive Mobile Technologies (iJIM)
Verified
Journal Article
Open access
PDF file
LSLamir Shkurti
Lamir Shkurti
ASArsim Susuri
Arsim Susuri
VSVehebi Sofiu
Vehebi Sofiu
BQBesnik Qehaja
Besnik Qehaja
Federated learning (FL) enhances data privacy but faces significant deployment challenges on resource-constrained edge devices. Building on our previous development of the AdaptiveMesh algorithm, this study presents an end-to-end implementation of a resourceaware FL framework for pneumonia detection from chest X-ray images. The system is optimized for heterogeneous hardware and unstable network conditions. Validated on two independent datasets, the proposed global model achieved high diagnostic reliability, reaching up to 95% accuracy and critical recall performance for clinical use. A key contribution of this work is the practical transition from decentralized training to real-time prediction; we integrated the trained model into a lightweight Flask-based web application deployed on devices such as Raspberry Pi 4 and Intel Atom MiniPCs. Our results confirm that the AdaptiveMesh-based system provides a stable, privacy-preserving, and feasible solution for medical diagnostics in low-resource environments, bridging the gap between algorithmic optimization and real-world application.

9 January 2026 in Mathematics
Verified
Journal Article
Open access
PDF file
Aybeyan SelimAybeyan Selim
Aybeyan Selim
Muzafer SaračevićMuzafer Saračević
Muzafer Saračević
ASArsim Susuri
Arsim Susuri
Grammar induction runs into a serious problem due to the exponential growth of the number of possible derivation trees as sentence length increases, which makes unsupervised parsing both computationally demanding and highly indeterminate. This paper proposes a mathematics-based approach that alleviates this combinatorial complexity by introducing structural constraints based on Catalan and Fuss–Catalan numbers. By limiting the depth of the tree, the degree of branching and the form of derivation, the method significantly narrows the search space, while retaining the full generative power of context-free grammars. A filtering algorithm guided by Catalan structures is developed that incorporates these combinatorial constraints directly into the execution process, with formal analysis showing that the search complexity, under realistic assumptions about depth and richness, decreases from exponential to approximately polynomial. Experimental results on synthetic and natural-language datasets show that the Catalan-constrained model reduces candidate derivation trees by approximately 60%, improves F1 accuracy over unconstrained and depth-bounded baselines, and nearly halves average parsing time. Qualitative evaluation further indicates that the induced grammars exhibit more balanced and linguistically plausible structures. These findings demonstrate that Catalan-based structural constraints provide an elegant and effective mechanism for controlling ambiguity in grammar induction, bridging formal combinatorics with practical syntactic learning.

6 October 2026 in Mathematics
Verified
Journal Article
Open access
PDF file
Aybeyan SelimAybeyan Selim
Aybeyan Selim
Omer AydinOmer Aydin
Omer Aydin
Muzafer SaračevićMuzafer Saračević
Muzafer Saračević
Komlen LalovićKomlen Lalović
Komlen Lalović
ASArsim Susuri
Arsim Susuri
This study presents a structural cryptanalysis of a Fiat–Shamir-type digital-signature candidate based on class-2 nilpotent Lie algebras over finite fields. We show that the public transformation is linear in the signing secret and that public-key inversion reduces to a consistent linear system over the underlying finite field, solvable in polynomial time by Gaussian elimination. Because the associated linear map has a non-trivial kernel, the public key determines an affine family of equivalent signing secrets rather than a unique private key. Any compatible representative can therefore be used to generate a valid signature for a fresh message without access to a signing oracle, yielding a polynomial-time no-message forgery and failure of EUF-CMA security. The baseline evaluation considered n=16, 32, 64 and 128, using ten independently generated Lie algebra instances per dimension and ten cryptographic trials per instance. In the 400 baseline trials, none of the recovered compatible secrets coincided with the corresponding originally sampled private key, yet every recovered representative was sufficient to generate an accepted forgery, yielding a 100% experimental forgery rate. Additional sensitivity experiments on field size, central fraction, and bracket density showed that these parameters affect practical attack cost and rank-nullity structure but do not remove the vulnerability. The results demonstrate that the examined class-2 transformation is unsuitable as a hardness basis for a secure digital-signature construction or a meaningful post-quantum security claim.