Predicting future values of time series is possible using statistical approaches and machine learning. After years of research, the latter offers numerous techniques, protocols, and predictive models for predicting one as well as numerous future values of time-collected data. When predicting time series values, we usually talk about predicting one point, as predicting more than one future time point is a more complex problem. For predicting more than one point, we must face a higher accumulation of the predictive errors. It is generally accepted that the more points predicted beforehand, the greater the predictive error of the more distant points.
In the thesis, we were looking for optimal combinations of the time series processing techniques and parameterization of different predictive models. We predicted several points of predetermined events with different frequency based on online advertising data on social networks. We focused on various architectural neural networks with long short-term memory (LSTM). We wanted to compare the predictions of techniques using neural networks with the predictions of the statistical ARIMA predictive model recognized for predicting time series, as well as the predictions of the XGBoost regression model. The latter was used due to the fact that it has lately given very good results in various competitions of numerous fields of machine learning. We assumed that LSTM architectures will provide the most accurate predictions.
Based on the experiments and results analysis, we established that neural networks with long short-term memory give the most accurate predictions of all frequency of events. Considering the accuracy, the closest to the mentioned neural networks are only XGBoost model predictions, while ARIMA model predictions are on average the least accurate. Data processing plays an important role in accuracy of the results. Logarithmic transformation and normalization performed well for all window sizes, and for larger predictive windows the removal of seasonality was the best option. We also established that predictive accuracy is not lost by increasing predictive window sizes in individual groups of the predictive models.
|
