| Technologies forecasting is significant for enterprise,government and other decision-making organizations’ resources allocation.Current technologies forecasting methods require long-term human and material resource investment.The thesis aims to propose a patent-based technology forecasting algorithm to offer some reference and save some investment in technologies forecasting,and design and realize a prototype to validate the feasibility in production environment.In the thesis,the problem of technologies forecasting is converted into the problem of time series forecasting.Currently,common time series forecasting methods embed all of time series in the dataset into a high-dimension time series while training and forecasting.The embedding makes the model have too much parameters.And,while new time series emerges,dimension of input increases.The entire model requires to be trained again.Meanwhile,the number and the length of time series in the dataset is much less than them of other time series forecasting problems.That makes the model’s training hard.And new technology can emerge at any time.Therefore,the number of time series in the dataset can increase at any time.The model should be able to be trained incrementally.Besides,common time series forecasting focus on the forecasting error at a point in the future.But the problem of technologies forecasting also focuses on the forecasting of the trend.So,some new evaluation should be used on the model’s validation.In the thesis,we develop a novel model that is trained by all the series to find the common patterns and generate corresponding prediction to deal with the trend.And we use a more appropriate evaluation 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.While testing and validating,the thesis uses the evaluation of both the error of the forecasting,like others does,and correlative of trend.The result indicates that the model of the thesis performs better in the forecasting of both values and trend on patent dataset.The thesis also selects some dataset containing multiple short time series to validate the model.On these dataset,the model also outperforms in the forecasting of both values and trend.That indicates the universality of the model.Finally,a technologies forecasting prototype system is designed and developed.The prototype includes searching module,algorithm module and management module.The prototype provides many kinds of data for users to analyze trends of technologies,including basic patent data and unique technology forecasting data.After tests,the performance of the prototype meets the expectation.The prototype can be used in production environment and the feasibility of the algorithm is validated. |
