Haydar Tuna

Manual classification of algorithmic problems is costly and inconsistent, although it remains essential for curriculum sequencing, recommendation systems, and quality control on competitive programming platforms. To address this limitation, this study investigates whether consensus-based, zero-shot classification using large language models can replace manual annotation. Five architecturally diverse models were evaluated: Claude Sonnet v4.6 (CS4.6), DeepSeek v3.2 (DS3.2), Gemini v3 Flash (G3F), GPT v5.4 (GPT5.4), and Grok v4.1 Fast (GRK4.1). Each model independently classified 2911 LeetCode problems into 13 mutually exclusive algorithm categories to generate predicted labels and confidence vectors. These outputs served as the basis for systematically comparing eight consensus aggregation methods and five disagreement-resolution strategies. The five-model ensemble achieved at least 4/5 agreement on 85.3% of the problems, demonstrating substantial overall inter-rater reliability. While consensus weakened as problem difficulty increased, it was notably lowest for the Algorithmic Paradigms group because category boundaries depend on computational properties that are not directly visible in source code. Six aggregation methods converged within a narrow 90.0–90.1% average-agreement band. A Friedman omnibus test with Holm–Bonferroni-corrected pairwise comparisons showed that this band separates further into two statistically distinguishable trios differing by only 0.07–0.08 percentage points, illustrating that statistical significance at this sample size does not imply practical importance. An unsupervised adaptation of pairwise ranking from LLM-Blender reached 89.1%, whereas Borda Count underperformed at 83.1%. Among disagreement-resolution strategies, progressive ensemble pruning proved most effective by resolving 92.5% of cases. Calibration varied substantially across the panel, with GPT5.4 showing the largest calibration gap (0.124) and GRK4.1 showing the smallest (0.023). This variation suggests that confidence-weighted aggregation is effective only when applied selectively to well-calibrated models. These findings support consensus-based ensembles of these five models as a scalable alternative to manual algorithmic annotation and provide actionable guidance for designing robust aggregation pipelines in unsupervised settings.

Supervised emotion classification for Turkish is limited by a scarcity of large, balanced, human-annotated corpora. Large language models (LLMs) can generate labeled sentences on demand; however, data quality varies across providers, and few studies compare multiple LLMs under a leakage-free design. To address this gap, this study compares seven LLMs as synthetic-data sources for Turkish emotion classification: Claude Opus 4.8, DeepSeek 3.2, Gemini 3.5 Flash, GPT 5.4, Grok 4.3, Llama 4 Scout, and Mistral Large 3. Each model generated approximately 2,000 sentences for each of six emotion categories: happiness, fear, sadness, disgust, surprise, and anger. For each corpus, a BERTurk classifier was fine-tuned and evaluated on a balanced, held-out subset of the human-annotated TREMO corpus. Each corpus was characterized using lexical diversity and self-similarity metrics and compared against TREMO on inter-emotion vocabulary overlap.Grok 4.3 produced the most diverse and least repetitive sentences, while TREMO retained the highest lexical diversity without the highest inter-emotion overlap. On the TREMO test set, macro-F1 ranged from 0.749 for Mistral Large 3 to 0.813 for Claude Opus 4.8. Classification performance did not track corpus diversity: Grok 4.3, the most diverse source, produced the sixth-best classifier, while DeepSeek 3.2, a repetitive source, produced the second-best. Every model over-predicted disgust, and most under-recalled sadness; statistical tests confirmed significant differences between most model pairs.These results indicate that diversity metrics are not a reliable guide for selecting a generator LLM. A small, human-labeled sample, validated against each candidate LLM before full-scale generation, offers a direct alternative.

This study aims to define new simple shape descriptors to analyze airfoils. The ImageJ platform is used to calculate twelve different shape descriptors such as area, convex hull, contour temperature and solidity by performing image processing. One of the most important findings is that an increase in the thickness of an airfoil leads to corresponding increases in its area, perimeter, area of minimum enclosing area, and convex hull area. Another noteworthy discovery is that the values derived from these basic features, either increasing or decreasing. Simple shape features in the study are not used independently, as they do not possess distinct characteristics that set them apart from one another. Machine learning and deep learning applications can achieve greater success when these features are combined with other elements. The combination of these features with other shape attributes, such as chain code histograms, shape signatures, and central moments, can enhance the success of machine learning and deep learning applications.

Şekil bir nesnenin doğal sınırlarının insan beyninde oluşturduğu algı olarak tanımlanabilir. Şekil tanıma ise herhangi bir nesnenin ait olduğu sınıfın daha önceden karşılaşılan ve hafızaya kaydedilen algılarla karşılaştırılarak bulunmasıdır. Bilgisayarda şekil tanıma nesnelerin sınır veya bölge tabanlı şekil temsil yöntemleriyle elde edilen özellik vektörlerinin karşılaştırılarak sınıflarının tespitidir. Sınır tabanlı özellik çıkarma metotlarından biri olan zincir kodu sayısal görüntüdeki bir nesnenin sınır noktaları takip edilerek üretilen sembol dizisidir. İlgili sembol kümesinin elemanları her yön için daha önceden belirlenmelidir. Şekil temsilinde kullanılan zincir kodlarının en temel problemi ölçekleme veya döndürme işlemlerine karşı yeterince güçlü olmamalarıdır. Başka bir ifade ile şekiller ölçeklendiğinde ya da döndürüldüğünde zincir kodlarının uzunluklarının ve içeriklerinin değişmesidir. Bu nedenle şekillerin benzerlik karşılaştırılmasında farklı uzunluklardaki sembol dizisinden oluşan zincir kodları yerine, normalize edilmiş zincir kod histogramı tercih edilmektedir. Böylece sınır bilgileri sembol çeşidi ile orantılı olan sabit uzunlukta vektörlere dönüştürülerek benzerlik hesaplaması yapılmaktadır. Bu çalışmada nesnelerin sınır noktalarında bulunan piksellerin kabuk numaraları kullanılarak ölçeklenme ve döndürme işlemlerine karşı dayanıklı yeni bir zincir kod histogramı önerilmiştir. Önerilen yöntemin döndürmeye karşı duyarlılığı Freeman 8 (FR8) zincir kod histogramıyla deneysel olarak karşılaştırılmış ve elde edilen sonuçlar verilmiştir.