Mark Last
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Mark Last has a total of 90 co-authors. Here are the top co-authors by number of shared publications:
- Abraham Kandel
University of South Florida
5 shared publications - Armin Shmilovici
Ben-Gurion University of the Negev
5 shared publications - Abraham Kandel
University of South Florida
4 shared publications Abraham Kandel
Department of Computer Science and Engineering, University of South Florida, Tampa, USA
2 shared publications- Bracha Shapira
Ben-Gurion University of the Negev
2 shared publications - Chang Liu
Department of Video Algorithms, Haidian District, Xiaomi Communications Co. Ltd, Beijing, China
2 shared publications
Latest Publications
- 13 November 2026
- Published by Springer Nature
- p. 273-285
Can Induced Emotions Predict Movie Success? (Short Paper)
- 13 November 2026
- Published by Springer Nature
- p. 273-285
Can Induced Emotions Predict Movie Success? (Short Paper)
- 7 May 2026
- Published by Springer Nature
- p. 219-230
A Clustering-Based Algorithm for Fast Anomaly Detection in Hyperspectral Images
- 7 May 2026
- Published by Springer Nature
- p. 219-230
A Clustering-Based Algorithm for Fast Anomaly Detection in Hyperspectral Images
- 24 November 2025
- Published by Springer Nature
- p. 347-357
Automatically Reviewing Movie Plots with LLMs
- Open Access
- 13 January 2025
- Published by Springer Nature in Machine Learning
- Vol. 114 (1) , 1-35
Fast online feature selection in streaming data
The challenge of getting big amounts of high-quality labeled data is compounded by the fact that data labeling is often subjective and requires significant human effort. In many cases, the quality of the labeled data depends entirely on the expertise and experience of human annotators, making it challenging to ensure labeling accuracy in large and dynamic datasets. Moreover, there may be a significant delay between the arrival of a new instance and its manual labeling. This paper explores the use of fully unsupervised feature selection algorithms in non-stationary data streams, where the importance of features may change over time. We introduce a novel feature selection algorithm called Online Fast FEa-ture SELection-OFFESEL, which calculates the feature importance scores in each incoming window based on their mean normalized values and without using any class labels. We evaluate OFFESEL on 17 benchmark data streams, both stationary and non-stationary, using popular online classifiers like PerceptronMask, VFDT, Online Boosting, and Linear SVM. We compare OFFESEL to several other feature selection algorithms, including state-of-the-art supervised ones like FIRES and ABFS, as well as popular unsupervised ones like MCFS, LS, and Max Variance, which we adapted to data streams. Our results indicate that OFFESEL outperforms all supervised and unsupervised feature selection algorithms in terms of classification accuracy. Specifically, OFFESEL preserves the accuracy level of the supervised FIRES algorithm, which proved more accurate than ABFS in our experiments, while maintaining the accuracy level achieved by the unsupervised Max Variance algorithm. Moreover, OFFESEL requires even less computation time than Max Variance and shows high stability on stationary datasets. Overall, our study demonstrates the potential benefits of using unlabeled data for feature ranking and selection in dynamic data streams.
- 13 January 2025
- Published by Springer Nature in Machine Learning
- Vol. 114 (1) , 1-35
Open AccessFast online feature selection in streaming data
The challenge of getting big amounts of high-quality labeled data is compounded by the fact that data labeling is often subjective and requires significant human effort. In many cases, the quality of the labeled data depends entirely on the expertise and experience of human annotators, making it challenging to ensure labeling accuracy in large and dynamic datasets. Moreover, there may be a significant delay between the arrival of a new instance and its manual labeling. This paper explores the use of fully unsupervised feature selection algorithms in non-stationary data streams, where the importance of features may change over time. We introduce a novel feature selection algorithm called Online Fast FEa-ture SELection-OFFESEL, which calculates the feature importance scores in each incoming window based on their mean normalized values and without using any class labels. We evaluate OFFESEL on 17 benchmark data streams, both stationary and non-stationary, using popular online classifiers like PerceptronMask, VFDT, Online Boosting, and Linear SVM. We compare OFFESEL to several other feature selection algorithms, including state-of-the-art supervised ones like FIRES and ABFS, as well as popular unsupervised ones like MCFS, LS, and Max Variance, which we adapted to data streams. Our results indicate that OFFESEL outperforms all supervised and unsupervised feature selection algorithms in terms of classification accuracy. Specifically, OFFESEL preserves the accuracy level of the supervised FIRES algorithm, which proved more accurate than ABFS in our experiments, while maintaining the accuracy level achieved by the unsupervised Max Variance algorithm. Moreover, OFFESEL requires even less computation time than Max Variance and shows high stability on stationary datasets. Overall, our study demonstrates the potential benefits of using unlabeled data for feature ranking and selection in dynamic data streams.
- Open Access
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 22637-22653
Transitive self-consistency evaluation of NLI models without gold labels
Wei Wu, Mark Last. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.
- 1 January 2025
- Published by Association for Computational Linguistics (ACL)
- p. 22637-22653
Open AccessTransitive self-consistency evaluation of NLI models without gold labels
Wei Wu, Mark Last. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.
- Open Access
- 1 January 2025
- Published by Canadian Science Publishing in Drone Systems and Applications
- Vol. 13, 1-10
Early detection of small- and medium-sized drones in complex environments
Unmanned aerial vehicles, or drones, have become a weapon of choice on the modern battlefield. As military and civilian industries try to develop effective counter-drone systems, early detection of flying drones still poses multiple challenges to state-of-the-art computer vision technologies. These challenges include a growing variety of military and commercial drones, their small size compared to piloted aircraft, blending with the background, similarity to birds, etc. In our experiments on drone images of variable size, we have observed a rapid drop in the accuracy of a state-of-the-art drone detection model when applied to distant drones that take a relatively small area on the entire image. However, we show that this early detection accuracy can be significantly improved by applying the drone detection model to an image masked by the Canny edge detector. We suggest applying the model to the original and the masked images in parallel and determining a drone detection decision by the highest confidence value under the condition that the detected object is not recognized as a bird by a general-purpose object detection model. The results of our evaluation experiments confirm the effectiveness of the proposed drone detection approach.
- 1 January 2025
- Published by Canadian Science Publishing in Drone Systems and Applications
- Vol. 13, 1-10
Open AccessEarly detection of small- and medium-sized drones in complex environments
Unmanned aerial vehicles, or drones, have become a weapon of choice on the modern battlefield. As military and civilian industries try to develop effective counter-drone systems, early detection of flying drones still poses multiple challenges to state-of-the-art computer vision technologies. These challenges include a growing variety of military and commercial drones, their small size compared to piloted aircraft, blending with the background, similarity to birds, etc. In our experiments on drone images of variable size, we have observed a rapid drop in the accuracy of a state-of-the-art drone detection model when applied to distant drones that take a relatively small area on the entire image. However, we show that this early detection accuracy can be significantly improved by applying the drone detection model to an image masked by the Canny edge detector. We suggest applying the model to the original and the masked images in parallel and determining a drone detection decision by the highest confidence value under the condition that the detected object is not recognized as a bird by a general-purpose object detection model. The results of our evaluation experiments confirm the effectiveness of the proposed drone detection approach.
- Open Access
- 18 October 2024
- Published by Springer Nature in International Journal of Data Science and Analytics
- Vol. 19 (3) , 537-550
Leveraging peer-review aspects for extractive and abstractive summarization of scientific articles
This research introduces an innovative framework that supports the peer-review process by automatically extracting the following four key aspects of a scientific paper: contribution, motivation, claims, and claims support. Leveraging these extracted aspects, we generate extractive and abstractive summaries of scientific papers. Additionally, we provide a benchmarking corpus containing 1000 aspect-related sentences extracted from 40 scientific articles, which can serve as a valuable resource for evaluating various aspect extraction methods. Experimental findings reveal that our automated aspect extraction system successfully identifies between 86 and 92% of sentences related to each aspect with precision ranging from 84 to 94%. The aspect-based extractive summaries outperformed the original paper abstracts in terms of the Rouge scores as well as in Relevance, Consistency, Fluency, and Coherence dimensions. Furthermore, our study demonstrates that by prompting the LLMs using the paper itself along with the extracted aspects, the LLM-generated summaries exhibit superior performance compared to prompting with either the paper or the abstract only.
- 18 October 2024
- Published by Springer Nature in International Journal of Data Science and Analytics
- Vol. 19 (3) , 537-550
Open AccessLeveraging peer-review aspects for extractive and abstractive summarization of scientific articles
This research introduces an innovative framework that supports the peer-review process by automatically extracting the following four key aspects of a scientific paper: contribution, motivation, claims, and claims support. Leveraging these extracted aspects, we generate extractive and abstractive summaries of scientific papers. Additionally, we provide a benchmarking corpus containing 1000 aspect-related sentences extracted from 40 scientific articles, which can serve as a valuable resource for evaluating various aspect extraction methods. Experimental findings reveal that our automated aspect extraction system successfully identifies between 86 and 92% of sentences related to each aspect with precision ranging from 84 to 94%. The aspect-based extractive summaries outperformed the original paper abstracts in terms of the Rouge scores as well as in Relevance, Consistency, Fluency, and Coherence dimensions. Furthermore, our study demonstrates that by prompting the LLMs using the paper itself along with the extracted aspects, the LLM-generated summaries exhibit superior performance compared to prompting with either the paper or the abstract only.
- Open Access
- 11 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 105573-105587
Efficient Feature Ranking and Selection Using Statistical Moments
Unsupervised feature selection methods can be more efficient than supervised methods, which rely on the expensive and time-consuming data labeling process. The paper introduced skewness as a novel, unsupervised, and computationally efficient feature ranking metric, suitable for both classification and regression tasks. Its feature selection effectiveness is compared to several state-of-the-art supervised and unsupervised feature ranking and selection methods. Both theoretical analysis and empirical evaluation on several popular classification and regression algorithms show that statistical moment-based feature selection algorithms are competitive in terms of accuracy and mean squared error (MSE) with the state-of-the-art supervised approaches for feature ranking and selection, including Fast Correlation Based Filter (FCBF), Minimum Redundancy Maximum Relevance (MRMR), and Mutual Information Maximization (MIM). We also present a mathematical proof based on some common assumptions, which explains the high effectiveness of statistical moments in the feature ranking procedure. Moreover, statistical moment-based feature selection is shown empirically to run faster, on average, than the supervised approaches and the unsupervised Laplacian Score method. Additionally, skewness-based feature selection, in contrast to variance-based selection, does not depend on data normalization that requires additional computational time and may affect the feature ranking results.
- 11 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Access
- Vol. 12, 105573-105587
Open AccessEfficient Feature Ranking and Selection Using Statistical Moments
Unsupervised feature selection methods can be more efficient than supervised methods, which rely on the expensive and time-consuming data labeling process. The paper introduced skewness as a novel, unsupervised, and computationally efficient feature ranking metric, suitable for both classification and regression tasks. Its feature selection effectiveness is compared to several state-of-the-art supervised and unsupervised feature ranking and selection methods. Both theoretical analysis and empirical evaluation on several popular classification and regression algorithms show that statistical moment-based feature selection algorithms are competitive in terms of accuracy and mean squared error (MSE) with the state-of-the-art supervised approaches for feature ranking and selection, including Fast Correlation Based Filter (FCBF), Minimum Redundancy Maximum Relevance (MRMR), and Mutual Information Maximization (MIM). We also present a mathematical proof based on some common assumptions, which explains the high effectiveness of statistical moments in the feature ranking procedure. Moreover, statistical moment-based feature selection is shown empirically to run faster, on average, than the supervised approaches and the unsupervised Laplacian Score method. Additionally, skewness-based feature selection, in contrast to variance-based selection, does not depend on data normalization that requires additional computational time and may affect the feature ranking results.
- Open Access
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 846-852
Embible: Reconstruction of Ancient Hebrew and Aramaic Texts Using Transformers
Niv Fono, Harel Moshayof, Eldar Karol, Itai Assraf, Mark Last. Findings of the Association for Computational Linguistics: EACL 2024. 2024.
- 1 January 2024
- Published by Association for Computational Linguistics (ACL)
- p. 846-852
Open AccessEmbible: Reconstruction of Ancient Hebrew and Aramaic Texts Using Transformers
Niv Fono, Harel Moshayof, Eldar Karol, Itai Assraf, Mark Last. Findings of the Association for Computational Linguistics: EACL 2024. 2024.
- 4 October 2023
- Published by Springer Nature in Neural Computing and Applications
- Vol. 36 (2) , 1023-1043
Towards efficient image-based representation of tabular data
- 4 October 2023
- Published by Springer Nature in Neural Computing and Applications
- Vol. 36 (2) , 1023-1043
Towards efficient image-based representation of tabular data