Search

Stock forecasting is challenging because of stock volatility and dependability on external factors, such as economic, social, and political factors. This motivates investors to seek tools to identify stock trends to reap profits.In this research, we compared several heterogeneous ensembles for financial forecasting, including averaging, weighted, stacking, and blending ensembles. In addition, we used a random forest regressor as the baseline.Regression was used to predict the next day’s closing stock price. We used classification to label closing stock value as HIGH or LOW by comparing with the opening stock value of a particular company. We used Long Short Term Memory (LSTM) models, Linear Regression, and Support Vector Machines (SVM) as individual models. Further, we analyzed 10 years of historical data of the most active 20 companies of the NASDAQ stock exchange for implementing ensemble models.In conclusion, experimental results depict blending ensembles perform the best out of compared ensembles in financial forecasting. Further, they reveal SVM is under-performing, LSTM outputs are satisfactory, while linear regression produced promising results.Data: Data for this research was gathered from online available sources from the NASDAQ American stock exchange.We gathered data for most active 20 companies and 10 years of historical data from 21st September 2019 backwards. We used 40044 data points in total.

In Olympic-style weightlifting athlete attempts to lift the weight plates on a barbell and scores are determined by a combination of the successful highest weight achieved in snatch and the clean-and-jerk actions. However, the current method does not objectively measure the player techniques. In this paper, we introduce a novel method to objectively measure player performance on weightlifting using human action recognition in videos. We introduce a method to assess player techniques in weightlifting by using skeleton-based human action recognition. In order to achieve our goal, we further introduce a new video dataset for action recognition in weightlifting which is annotated to frame level and introduce an automated scoring system through action recognition. We conclude our paper with qualitative and quantitative experimental results using non-Olympic players and 2016 Gold, Silver, and Bronze medalists in different weight categories (both men and women).

k-Nearest Neighbor (k-NN) is a well-known classification algorithm used in many machine learning applications. When the input data size of k-NN algorithm increases, the execution time of the algorithm increases significantly. Therefore, this becomes a bottleneck for practical usage. Since k-NN algorithm is using searching, sorting and other parallelly executable tasks, we have implemented the k-NN algorithm on a GPU using CUDA utilizing the parallel tasks. The results were promising as expected and for 43,500 training records and 14,500 testing records with nine attributes, the GPU execution showed about a 100× faster execution time compared to a CPU version.

The focus of this research was to use Educational Data Mining (EDM) techniques to conduct a quantitative analysis of students interaction with an e-learning system through instructor-led non-graded and graded courses. This exercise is useful for establishing a guideline for a series of online short courses for them. A group of 412 students' access behaviour in an e-learning system were analysed and they were grouped into clusters using K-Means clustering method according to their course access log records. The results explained that more than 40% from the student group are passive online learners in both graded and non-graded learning environments. The result showed that the difference in the learning environments could change the online access behaviour of a student group. Clustering divided the student population into five access groups based on their course access behaviour. Among these groups, the least access group (NG-41% and G-42%) and the highest access group (NG-9% and G-5%) could be identified very clearly due to their access variation from the rest of the groups.

Plagiarism is one of the growing issues in academia and is always a concern in Universities and other academic institutions. The situation is becoming even worse with the availability of ample resources on the web. This paper focuses on creating an effective and fast tool for plagiarism detection for text based electronic assignments. Our plagiarism detection tool named AntiPlag is developed using the tri-gram sequence matching technique. Three sets of text based assignments were tested by AntiPlag and the results were compared against an existing commercial plagiarism detection tool. AntiPlag showed better results in terms of false positives compared to the commercial tool due to the pre-processing steps performed in AntiPlag. In addition, to improve the detection latency, AntiPlag applies a data clustering technique making it four times faster than the commercial tool considered. AntiPlag could be used to isolate plagiarized text based assignments from non-plagiarised assignments easily. Therefore, we present AntiPlag, a fast and effective tool for plagiarism detection on text based electronic assignments.

Image retrieval in general and content based image retrieval in particular are well-known research fields in information management. A large number of methods have been proposed and investigated in both areas but satisfactory general solution have still not been developed. An image contains several types of visual information which are difficult to extract and combine manually by humans. In this paper, we propose a content based image retrieval system based on three major types of visual information: colour, texture and shape, and their distances to the origin in a three dimensional space for the retrieval. We experimentally investigated several feature extraction methods and learning algorithms for content based image retrieval. The results show that 5-Nearest Neighbour yield the highest accuracy for the chosen feature extraction methods.

Using computers to answer natural language questions is an interesting and challenging problem. Generally such problems are handled under two categories: open domain problems and close domain problems. This paper presents a system that attempts to solve close domain problems. Typically, in a close domain, answers to questions are not available in the public domain and therefore they cannot be searched using a search engine. Hence answers have to be stored in a database by a domain expert. Then, the challenge is to understand the natural language question so that the solution could be matched to the respective answer in the database. We use a template matching technique to perform this matching. In addition, given that our target is to use this system with non-native English speakers, we developed a method to overcome the mismatches we might encounter due to spelling mistakes. The system is developed such that the questions can be asked using short messages from a mobile phone and therefore the system is designed to understand SMS language in addition to English. One of the main contributions of this paper is the outcome presented of a deployment of this system in a real environment.

The computational cost of using nearest neighbor classification often prevents the method from being applied in practice when dealing with high-dimensional data, such as images and micro arrays. One possible solution to this problem is to reduce the dimensionality of the data, ideally without loosing predictive performance. Two different dimensionality reduction methods, principle component analysis (PCA) and random projection (RP), are investigated for this purpose and compared w.r.t. the performance of the resulting nearest neighbor classifier on five image data sets and five micro array data sets. The experiment results demonstrate that PCA outperforms RP for all data sets used in this study. However, the experiments also show that PCA is more sensitive to the choice of the number of reduced dimensions. After reaching a peak, the accuracy degrades with the number of dimensions for PCA, while the accuracy for RP increases with the number of dimensions. The experiments also show that the use of PCA and RP may even outperform using the non-reduced feature set (in 9 respectively 6 cases out of 10), hence not only resulting in more efficient, but also more effective, nearest neighbor classification.