Ongoing Research
DEVELOPING MACHINE LEARNING ASSISTED MODELS AND SOLUTION METHODS FOR INTEGRATED CONTAINER LOADING AND ROUTING PROBLEMS IN LOGISTICS OPERATIONS
Funded by TUBITAK (The Scientific and Technological Research Council)
dsOpt Researchers: U. Mahir Yıldırım, Doruk Sen
This project aims to enhance customer satisfaction in logistics by integrating models that address routing problems with loading constraints, crucial for efficient route planning. Focused on the logistics industry’s specific challenges, such as inefficiencies and uncertainties in order data leading to suboptimal loading plans, the project will utilize industry data from Ekol Logistics to perform descriptive analyses on past customer and order information. It will develop prediction models using Supervised Machine Learning algorithms like Support Vector Machines and Artificial Neural Networks for forecasting uncertain or missing cargo details. By hybridizing these algorithms with meta-heuristics, the project intends to pioneer in the logistics literature for predicting freight details. Additionally, it will tackle the logistics companies’ main issue of reliable, cost-effective delivery by addressing container loading
and vehicle routing with time windows, areas rarely solved concurrently due to their NP-hard nature.
The project will present a mathematical model and develop hybrid solution methods to solve real-life problems effectively. It seeks to provide a holistic methodology that reduces cost increases from inefficient planning and contributes to decreasing transportation cost-induced indirect inflation, enriching academic and practical experiences in logistics.
A HYBRID ARTWORK VALUATION MODEL PROPOSITION USING NATURAL LANGUAGE PROCESSING, MACHINE LEARNING AND METAHEURISTIC ALGORITHMS
Funded by ISTANBUL BILGI UNIVERSITY
Researchers: Doruk Sen, U. Mahir Yıldırım, Nil Yagmur Ilba
The research aims to propose a hybrid model using machine learning, natural language processing, and metaheuristic algorithms to address the issue of a discrepancy between the estimated price range for the artworks and the auction hammer price. Artist and artwork information for the valuation model will be obtained via web scraping techniques from various internet sources to implement NLP and ML techniques. Metaheuristics, however, will be used to improve the prediction performance of the model by determining the factors that influence the price. A realistic and systematic price estimation model for artworks will be developed using the aforementioned techniques. After the successful completion of this study, it aims to publish the methodology to guide the industry for increased investor diversity.
Setting a project frame within the scope of the United Nations Sustainable Development Goals, researchers aim to contribute to the establishment of effective, accountable, and transparent institutions at all levels by determining the factors affecting the price of the artworks with a transparent valuation model to society and the industry.
Determining Risk LevelS for Musical Instruments Using Machine Learning and Metaheuristic Algorithms
Researchers: Doruk Sen, Nil Yagmur Ilba, Barensel Bas
Insurance risk premiums are calculated over a certain percentage of the musical instrument price. In this project, instead of making calculations based on the price of an instrument, The risk level for musical instruments was determined by obtaining product reviews, technical details of products, damage records from Turkey’s leading retail music market, and price information.
The model is developed using a semi-supervised machine learning model and a tabu search algorithm for hyperparameter tuning are developed based on the Support Vector Machine algorithm. While the accuracy value was 87.2% in the basic model, it increased to 92.7% with the method developed using the tabu search algorithm. The developed model was fed to the self-training classifier and the risk level was determined for 1373 products that were excluded from the sample with gradual estimations. Currently, we are developing a new hybridized ML-MH algorithm to enhance the model’s predictive performance.
The technical details of the products and user comments were obtained by data scraping from Amazon and Sweetwater websites, which contain the largest information about musical instruments. In the obtained texts, the semantic similarity was focused on by applying natural language processing methods. In this context, semantic similarity scores were determined using BERT, Sentence-BERT, Spacy, and BERTopic, and comments showing similar characteristics were clustered. The obtained data were converted into numerical features and used in the risk level estimation model. In addition, risk level information was obtained for the sample (295 products) determined by the experts working in the music market to be used in the prediction model. All this information is available for the semi-supervised machine learning model developed to predict the risk levels of products excluded from the sample with data preprocessing.
A REINFORCEMENT LEARNING BASED ANT COLONY OPTIMIZATION ALGORITHM WITH APPLICATION IN COURSE SCHEDULING
Researchers: U. Mahir Yıldırım, Doruk Sen, Beril Güney
TBA