Long-Term Trend Prediction Algorithm Based on Neural Network for Short Time Series

In this paper, we focus on the problem of forecasting the trend of patents in different technologies. Different from other time series forecasting datasets, the length of series in our dataset is much shorter with fewer instances. So, other forecasting models which treat the time series as long high-dimension embedding do not perform well enough on this problem. Those models require a number of parameters. That is hard to trained by our dataset. And this kind of models can only applied on specified number of time series. While new technology emerges, the number of time series increases. That increases the dimension, so the model should be trained again. At the same time, not only do we value the error of forecasting, we also value the trend in our forecasting. So we develop a novel model that is trained by all the series to find the common patterns and generates corresponding prediction to deal with the trend. And we use a more appropriate index to evaluate trend that the model predicts. Treating the evolution of every technology as a time series, Convolution Neural Network (CNN) is used to capture the patterns among series. Therefore, our model requires less parameters and can be trained incrementally. Then, Recurrent Neural Network (RNN) is used to encode the information into an intermediate representation. With decoding the intermediate representation into values in multiple steps, the trend can be forecast. Finally, we test our model on other datasets. It achieves better results on some other datasets with that kind of characteristics.