Agus Widodo

Combination of time series forecasts using neural network

Forecast combination, which is a method to combine the result of several predictors, offers a way to improve the forecast result. Several methods have been proposed to combine the forecasting results into single forecast, namely the simple averaging, weighted average on validation performance, or non-parametric combination schemas. Recent literature uses dimensional reduction method for individual prediction and employs ordinary least squares for forecast combination. Other literature combines prediction results from neural networks using dimensional reduction techniques. Thus, those previous combination schemas can be categorized into linear combination methods. This paper aims to explore the use of non-linear combination method to perform the ensemble of individual predictors. We believe that the non-linear combination method may capture the non linear relationship among predictors, thus, may enhance the result of final prediction. The Neural Network (NN), which is widely used in literature for time series tasks, is used to perform such combination. The dataset used in the experiment is the time series data designated for NN5 Competition. The experimental result shows that forecast combination using NN performs better than the best individual predictors, provided that the predictors selected for combination have fairly good performance.

Automatic lag selection in time series forecasting using multiple kernel learning

This paper reports the feasibility of employing the recent approach on kernel learning, namely the multiple kernel learning (MKL), for time series forecasting to automatically select the optimal lag length or size of sliding windows. MKL is an approach to choose suitable kernels from a given pool of kernels by exploring the combination of multiple kernels. In this paper, we extend the MKL capability to select the optimal size of sliding windows for time series domain by adopting the data integration approach which has been previously studied in the domain of image processing. In this study, each kernel represents the different lengths of time series lag. In addition, we also examine the feasibility of MKL for decomposed time series. We use the dataset from previous time series competitions as our benchmark. Our experimental results indicate that our approaches perform competitively compared to the previous methods using the same dataset. Furthermore, MKL may predict the detrended time series without explicitly computing the seasonality. The advantage of our method is in its ability in automatically selecting the optimal size of sliding windows and finding the pattern of time series.

Technology forecasting in the field of Apnea from online publications: Time series analysis on Latent Semantic

Analysis of technological trends and developments has been attempted in a number of previous publications using a quantitative method to measure the growth of science. Previous studies on this subject, however, put more emphasis on the frequency either to categorize the technological classes or to find the most prominent technology for a given period of time but less analysis on the future trends. This paper presents the time series analysis of the technological trends while employing the Latent Semantic Analysis to associate each technological term. In current study, we are interested on analyzing the correlation among different trends of terms in the area of biomedical technology to deal with Apnea sleep disorder (difficulty in breathing during sleep). We assume that the study will also applicable to the other areas of research. The technological terms within the concept having the highest trend identified in our experiment are the movement disorders, cystic fibrosis, dental prosthesis, microvascular angina, and esophageal sphincter. Performance evaluation during the experiment indicates that the Support Vector Regression outperform the other techniques, while statistical techniques such as Holts and Winters yield above average performance and comparable to the Polynomial method.

Clustering patent document in the field of ICT (Information & Communication Technology)

The current classification of patent data that refers to the IPC (International Patent Classification) of the WIPO (World Intellectual Property Organization), deemed not reflect the classification of the field of ICT (Information & Communication Technology). ICT applications are usually included in sections G (Physics) and H (Electricity). This paper will evaluate the eight groupings of patents based on the IPC classes (G01, G06, G09, G11, H01, H03, H04, and H06) of patents registered in the Directorate General of Intellectual Property Rights in Indonesia, from the year 1991 to 2000. The algorithm used to grouping is KMeans, KMeans++, Hierchical Clustering, and a combination of these three algorithms with SVD (Singular Value Decomposition). For external validation, Purity and F-Measure are used, whereas Silhouette is used for internal validation. From the experimental results it can be concluded that SVD provides improvements to the clustering results. In addition, the use of abstract does not necessarily improve the performance of clustering, and the use of phrase does not always yield better cluster than the use of the word as index. Moreover, no cluster has purity measure greater than 50%, which means that the existing IPC classification has not been able to accommodate the field of ICT appropriately.

Nonlinearly weighted multiple kernel learning for time series forecasting

Machine Learning methods such as Neural Network (NN) and Support Vector Regression (SVR) have been studied extensively for time series forecasting. Multiple Kernel Learning (MKL) which utilizes SVR as the predictor is yet another recent approaches to choose suitable kernels from a given pool of kernels by means of a linear combination of some base kernels. However, some literatures suggest that this linear combination of kernels cannot consistently outperform either the uniform combination of base kernels or simply the best single kernel. Hence, in this paper, other combination method is devised, namely the squared combination of base kernels, which gives more weight on suitable kernels and vice versa. We use time series data having various length, pattern and horizons, namely the 111 time series from NN3 competition, 3003 of M3 competition, 1001 of Ml competition and reduced 111 of Ml competition. Our experimental results indicate that our new forecasting approaches using squared combination of Multiple Kernel Learning (MKL) may perform well compared to the other methods on the same dataset.

Feature enhancement for model selection in time series forecasting

Selecting the most appropriate forecasting model for certain time series may utilize the similarity between time series. Previous literature defined several global characteristics of time series as similarity measure. This paper attempts to enhance those characteristics by the coefficients of polynomial function. Considering that not all features may be useful for categorization, we employ feature selection to choose the most discriminating features. In addition, we select a forecasting method based on its previous performance on similar dataset. Hence, there is no need to train the current dataset against all predictors. The pool of predictors ranges from simple to sophisticated ones, namely polynomial interpolation, automatic ARIMA, and Multiple Kernel Learning. The dataset used for experiment is the 3003 records from M3 competition to construct the historical database and 111 records from the M1 competition as testing dataset. Our experimental results indicate that our feature enhancement for model selection may improve the forecasting performance.