Wyniki wyszukiwania - BazTech

1

Using the generalized fuzzy k-nearest neighbor classifier for biomass feedstocks classification

Mailagaha Kumbure Mahinda, Luukka Pasi

Annals of Computer Science and Information Systems

|

2021

|

Vol. 29

29--35

EN

This paper proposes a novel framework based on arecently introduced classifier called multi-local power mean fuzzyk-nearest neighbor (MLPM-FKNN) and the Minkowski distanceto classify biomass feedstocks into property-based classes. Theproposed approach uses k nearest neighbors from each class tocompute class-representative multi-local power mean vectors andthe Minkowski distance instead of the Euclidean distance to fitthe most suitable distance metric based on the properties of thedata in finding the nearest neighbors to the new data point.We evaluate the performance of the proposed approach usingthree biomass datasets collected from several articles publishedin reputable journals and the Phyllis 2 biomass database. Inputfeatures of the biomass samples include their characteristics fromthe proximate analysis and ultimate analysis. In the developedframework, we interpret the biomass feedstocks classification asa five-class problem, and the classification performance of theproposed approach is benchmarked with the results obtainedfrom classical k-nearest neighbor-, fuzzy k-nearest neighbor- andsupport vector machine classifiers. Experimental results showthat the proposed approach outperforms the benchmarks andverify its effectiveness as a suitable tool for biomass classificationproblems. It is also evident from the results that the featuresfrom both ultimate and proximate analyses can offer a betterclassification of biomass feedstocks than the features consideredfrom each of those analyses separately.

2

Luukka Pasi, Stoklasa Jan

Annals of Computer Science and Information Systems

|

2021

|

Vol. 29

45--51

EN

In this paper we present similarity based TOPSIS with OWA operators. The motivation behind this new method is the fact that in many real world problems it is more important to consider the amount of criteria that a particular alternative is able to satisfy instead of simply concentrating on the importance of particular criteria. Here with OWA operators we can tackle this problem together with multi-criteria decision making method called TOPSIS by aggregating alternatives' similarities towards positive ideal solution and negative ideal solution and aggregating these similarities using OWA. The use of linguistic quantifiers represented by OWA weights generated by a selected RIM quantifier allows for the reflection of decision-maker's attitude to risk in the calculation of the similarities of the alternative with positive and negative ideal solutions.

3

Material demand forecasting with classical and fuzzy time series models

Zakrytnoy Sergey, Luukka Pasi, Stoklasa Jan

Annals of Computer Science and Information Systems

|

2021

|

Vol. 29

1--6

EN

Direct material budgeting is an essential part of financial planning processes. It often implies the need to predict quantities and prices of hundreds of thousands of materials to be purchased by an enterprise in the upcoming fiscal period. Distortion effects in demand projections and overall uncertainty cause the enterprises to rely on internal data to build their forecasts. In this paper we are dealing with material demand forecasting and evaluate the feasibility of fuzzy time series forecasting models as compared to classical forecasting models. Relevant methods are shortlisted based on existing practice described in academic research. Three datasets from industry are used to evaluate the predictive performance of the shortlisted methods. Our findings show an improvement in prediction accuracy of up to 47% compared to naïve approach. Fuzzy time series models are reported to be the most reliable forecasting method for the analyzed intermittent time series in all three datasets.

4

Image based classification of shipments using transfer learning

Leppioja Markus, Luukka Pasi, Lohrmann Christoph

Annals of Computer Science and Information Systems

|

2021

|

Vol. 29

37--44

EN

This paper focuses on recognizing different postal shipment types from images taken by the sorting machine. Greyscale images obtained from sorting machines are used to build a classifier using transfer learning to recognize seven different classes of shipments. Three convolutional neural networks (VGG16, GoogLeNet and ResNet50), that were pretrained using the ImageNet dataset, were used as feature extractors and the extracted features were subsequently supplied to a neural network classifier. VGG16 demonstrated the best performance for six out of the seven classes and achieved an overall mean accuracy of 95.69% on the independent test set. The model accomplished F1 scores exceeding 90% for five out of seven classes, only having a lower recall for the aggregated class "Other'' and shipments from abroad. The results of this study highlight the potential of transfer learning for computer vision in the context of shipment classification.