Although deep learning (DL) algorithms have been proved to be effective in diverse research domains, their application in developing models for tabular data remains limited. Models trained on tabular data demonstrate higher efficacy using traditional machine learning models than DL models, which are largely attributed to the size and structure of tabular datasets and the specific application contexts in which they are utilized. Thus, the primary objective of this paper is to propose a method to use the supremacy of Stacked Bidirectional LSTM (Long Short-Term Memory) deep learning algorithms in pattern discovery incorporating tabular data with customized 3D tensor modeling in feeding neural networks. Our findings are empirically validated using six diverse, publicly available datasets each varying in size and learning objectives. This paper proves that the proposed model based on time-sequence DL algorithms, which were generally described as inadequate when dealing with tabular data, yields satisfactory results and competes effectively with other algorithms specifically designed for tabular data. An additional benefit of this approach is its ability to preserve simplicity while ensuring fast model training also with large datasets. Even with extremely small datasets, models can be applied to achieve exceptional predictive results and fully utilize their capacity.

Implementation of credit scoring models is a demanding task and crucial for risk management. Wrong decisions can significantly affect revenue, increase costs, and can lead to bankruptcy. Together with the improvement of machine learning algorithms over time, credit models based on novel algorithms have also improved and evolved. In this work, novel deep neural architectures, Stacked LSTM, and Stacked BiLSTM combined with SMOTE oversampling technique for the imbalanced dataset were developed and analyzed. The reason for the lack of publications that utilize Stacked LSTM-based models in credit scoring lies exactly in the fact that the deep learning algorithm is tailored to predict the next value of the time series, and credit scoring is a classification problem. The challenge and novelty of this approach involved the necessary adaptation of the credit scoring dataset to suit the time sequence nature of LSTM-based models. This was particularly crucial as, in practical credit scoring datasets, instances are not correlated nor time dependent. Moreover, the application of SMOTE to the newly constructed three-dimensional array served as an additional refinement step. The results show that techniques and novel approaches used in this study improved the performance of credit score prediction.

Time-aware recommender systems extend traditional recommendation methods by revealing user preferences over time or observing a specific temporal context. Among other features and advantages, they can be used to provide rating predictions based on changes in recurring time periods. Their underlying assumption is that users are similar if their behavior is similar in the same temporal context. Existing approaches usually consider separate temporal contexts and generated user profiles. In this paper, we create user profiles based on multidimensional temporal contexts and use their combined presentation in a user-based collaborative filtering method. The proposed model provides user preferences at a future point in time that matches temporal profiles. The experimental validation demonstrates that the proposed model is able to outperform the usual collaborative filtering algorithms in prediction accuracy.

Credit scoring is one the most important parts of credit risk management in reducing the risk of client defaults and bankruptcies. Deep learning has received much attention in recent years, but it has not been implemented so intensively in credit scoring compared to other financial domains. In this article, stacked unidirectional and bidirectional LSTM (long short‐term memory) networks as a complex area of deep learning are applied in solving credit scoring problems for the first time. The proposed robust model exploits the full potential of the three‐layer stacked LSTM and BDLSTM (bidirectional LSTM) architecture with the treatment and modeling of public datasets in a novel way since credit scoring is not a time sequence problem. Attributes of each loan instance were transformed into a sequence of the matrix with a fixed sliding window approach with a one‐time step. Our proposed models outperform existing and much more complex deep learning solutions thus we succeeded in preserving simplicity. In this article, measures of different types are employed to carry out consistent conclusions. The results by applying three hidden layers on the German Credit dataset showed an accuracy of 87.19%, for Kaggle dataset accuracy reached 93.69%, and for Microcredit dataset accuracy of 97.80%.

Social media is an important source of real-world data for sentiment analysis. Hate speech detection models can be trained on data from Twitter and then utilized for content filtering and removal of posts which contain hate speech. This work proposes a new algorithm for calculating user hate speech index based on user post history. Three available datasets were merged for the purpose of acquiring Twitter posts which contained hate speech. Text preprocessing and tokenization was performed, as well as outlier removal and class balancing. The proposed algorithm was used for determining hate speech index of users who posted tweets from the dataset. The preprocessed dataset was used for training and testing multiple machine learning models: k-means clustering without and with principal component analysis, naïve Bayes, decision tree and random forest. Four different feature subsets of the dataset were used for model training and testing. Anomaly detection, data transformation and parameter tuning were used in an attempt to improve classification accuracy. The highest F1 measure was achieved by training the model using a combination of user hate speech index and other user features. The results show that the usage of user hate speech index, with or without other user features, improves the accuracy of hate speech detection.

Vehicle Routing Problem (VRP) is the process of set selection of the most convenient route in a network of roads vehicles are supposed to drive along when serving customers. Although vehicle problems solutions are being researched and improved in science, this problem is also important in industry, and the reason is the potential reduction of the shipping cost. Transport management is the central problem in logistics of one company, and the choice of optimal routes is one of the crucial functions in that process. However, as much as routes are algorithmically optimal, and as much as they include predefined limitations, there are some factors in the realistic environment which perhaps are not adequately treated during the creating the given routes. The innovative approach of adjustment of most of the parameters and factors necessary for the VRP algorithms being used in reality is presented in this work. It is based on the principle of successful feasibility of the given routs in realistic environment. The feasibility of the routes on the realistic example of one of the greatest distribution companies in Bosnia and Herzegovina has been significantly increased by introducing the realistic settings and improvements by comparative results before and after the introduction of the suggested modifications.

Vehicle routing problem as the generalization of the Travelling Salesman Problem (TSP) is one of the most studied optimization problems. Industry itself pays special attention to this problem, since transportation is one of the most crucial segments in supplying goods. This paper presents an innovative cluster-based approach for the successful solving of real-world vehicle routing problems that can involve extremely complex VRP problems with many customers needing to be served. The validation of the entire approach was based on the real data of a distribution company, with transport savings being in a range of 10-20 %. At the same time, the transportation routes are completely feasible, satisfying all the realistic constraints and conditions.

Identifying at-risk students is a crucial step in different learning settings. Predictive modeling technique can be used to create an early warning system which predicts students’ success in courses and informs both the teacher and the student of their performance. In this paper we describe a course-specific model for prediction of at-risk students. The proposed model uses the case-based reasoning (CBR) methodology to predict at-risk students at three specific points in time during the first half of the semester. In general, CBR is an approach of solving new problems based on solutions of similar previously experienced problem situation encoded in the form of cases. The proposed model classifies students as at-risk based on the most similar past cases retrieved from the casebase by using the k-NN algorithm. According to the experimental evaluation of the model accuracy, CBR model that is being developed for a specific course showed potential for an early prediction of at-risk students. Although the presented CBR model has been applied for one specific course, the key elements of predictive model can be easily reused by other courses.

This paper presents a framework capable of accurately forecasting future sales in the retail industry and classifying the product portfolio according to the expected level of forecasting reliability. The proposed framework, that would be of great use for any company operating in the retail industry, is based on Facebook's Prophet algorithm and backtesting strategy. Real-world sales forecasting benchmark data obtained experimentally in a production environment in one of the biggest retail companies in Bosnia and Herzegovina is used to evaluate the framework and demonstrate its capabilities in a real-world use case scenario.