Potential for Improvement of Sentence Translations

Ergun Biçici, Potential for Improvement of Sentence Translations. 2023 8th International Conference on Computer Science and Engineering (UBMK), Burdur, Turkiye, 2023, pp. 482-485, doi: 10.1109/UBMK59864.2023.10286778. URL: https://ieeexplore.ieee.org/document/10286778

Natural language processing tasks like machine translation can be parallelized, corrected, and repeated at the sentence level since they are data-centric and highly parallel. Each sentence's translation quality can vary, thus any problems must be fixed on an individual basis. At the same time, translation quality should be predicted before further processing and shipment since we cannot be sure without knowing the target language. We identify the potential for improvement (pfi) of sentence translations, whether a translation has potential to be improved by further processing or retranslation, and achieve overall performance improvements. We describe a method for ranking the obtained translations to fix translation errors by retranslating the sentences using a computationally more demanding translation model. In the first step we identify the potential regardless of prediction errors and later we use triangular uninorm combination to sort instances in our decision making process about sorting sentences for retranslation. By retranslating the top 200 sentences that have the most pfi we achieve up to 2.31 BLEU improvements.

Instance Weighting in Neural Networks for Click-Through Rate Prediction

Ergun Biçici, Instance Weighting in Neural Networks for Click-Through Rate Prediction. 2023 Innovations in Intelligent Systems and Applications Conference (ASYU), Sivas, Turkiye, 2023, pp. 1-5, doi: 10.1109/ASYU58738.2023.10296657. URL: https://ieeexplore.ieee.org/abstract/document/10296657

The instances on which a learning algorithm is most undecided can be weighted more during training to guide the learning model towards spending more effort on the difficult instances. We introduce three instance weighting algorithms to weigh the loss obtained. All of these instance weighting methods improve loss, AUC, and F1 results. We demonstrate the improvements on four different classifiers and on two different datasets. The improvements in loss reach 2.8% and in AUC reach 0.73% for Masknet on the Avazu dataset.

Neural Network Calibration for CTR Prediction

Ergun Biçici and Hasan Saribaş, Neural Network Calibration for CTR Prediction. 2023 8th International Conference on Computer Science and Engineering (UBMK), Burdur, Turkiye, 2023, pp. 473-476, doi: 10.1109/UBMK59864.2023.10286733. URL: https://ieeexplore.ieee.org/document/10286733

After a machine learning model has been trained, calibration techniques correct its prediction errors to produce predictions that are more robust and confident. As calibration techniques, we implement isotonic regression, Platt's scaling, neural networks, spline regression, and temperature scaling and test them on the prediction of click-through rate (CTR), which is an unbalanced task. We use 3 neural network based CTR prediction models on a publicly available CTR dataset and measure the improvements. Our findings show that isotonic regression is the fastest method whereas isotonic regression and spline regression are the two techniques that improves the performance the most.

Efficiently Sampling in Neural Network Training for Click-Through Rate Prediction

Ergun Biçici and Serdarcan Dilbaz. Efficiently Sampling in Neural Network Training for Click-Through Rate Prediction. 2023 8th International Conference on Computer Science and Engineering (UBMK), Burdur, Turkiye, 2023, pp. 469-472, doi: 10.1109/UBMK59864.2023.10286811. URL: https://ieeexplore.ieee.org/document/10286811

Finding efficient downsampling techniques has become more crucial as the training datasets for advertisement click-through rate (CTR) prediction models are growing to billions in size. We present efficient downsampling to sample CTR datasets with goals of faster training and limited decrease in the performance. We present encouraging results demonstrating the effectiveness of our approach on two publicly available CTR prediction datasets and compare efficient downsampling with stratified random downsampling.

Residual Fusion Models with Neural Networks for CTR Prediction

Ergun Biçici. Residual Fusion Models with Neural Networks for CTR Prediction. 2023 8th International Conference on Computer Science and Engineering (UBMK), Burdur, Turkiye, 2023, pp. 01-04, doi: 10.1109/UBMK59864.2023.10286706. URL: https://ieeexplore.ieee.org/document/10286706

No single prediction model achieves the best performance on all datasets and we are better off combining the strengths of different models for each task. Residual fusion learning is a two step combination method that trains a second model on the residual of the target from the first model's prediction. In the final phase, the predictions of both of these models are added. We use gradient boosting decision trees (GBDT) and neural networks for the initial model in residual fusion and compare three GBDT models and four neural network models. We introduce residual fusion with two different neural network models and show that we can achieve AUC gains that reach 0.95%.