An Approach for Predicting Hype Cycle Based on Machine Learning Zhijun Ren Shuo Xu Institute of Scientific and Technical Information of China, Institute of Scientific and Technical Information of China, Beijing 100038, China Beijing 100038, China The China Patent Information Center, the State xush@istic.ac.cn Intellectual Property Office, Beijing 100088, China renzhijun@cnpat.com.cn Xiaodong Qiao Hongqi Han Institute of Scientific and Technical Information of China, Institute of Scientific and Technical Information of China, Beijing 100038, China Beijing 100038, China qiaox@istic.ac.cn bithhq@163.com Kai Zhang The China Patent Information Center, the State Intellectual Property Office, Beijing 100088, China zhangkai@cnpat.com.cn ABSTRACT Some scholars believe that hype means advertised in publicity Analyzing mass information and supporting insight based on and exaggerated propaganda[1], and the best measure is the analysis results are very important work but it needs much expectation, which means people’s expectation for technology effort and time. Therefore, in this paper, we propose an innovation. Enterprises should use hype cycle to target the approach for predicting hype cycle based on machine learning emerging technologies, and use the concept of digital business for effective, systematic, and objective information analysis transformation to predict future business trends. and future forecasting of science and IT field. Additionally, Gartner believes that hype cycle is a qualitative we execute a comparative evaluation between the suggested decision-making tool, like other management methods, it model and Hype Cycle for Big Data, 2013 for validating the relies mainly on the judgment of experts. And to complete the suggested model and generally used for information analysis hype cycle assessment and prediction of a set of technologies and forecasting. in a certain field, we need to use a variety of evaluation methods. Other scholars have explored how to measure Keywords expectations for innovation, and quantitative indicators are mainly the number of participants[2], the number or the ratio Hype Cycle; Data Mining; Technology Predict; KNN of the technological innovation documents[3], patent statistical data[4] and the search flow of Google and other search engines[5]. When using the network measurement 1. OVERVIEW method, other tools may need to be supplemented: USPTO Hype Cycle is a conceptual model widely used by Gartner, patents, news reports of Google news archive and the official Inc., which can reveal the basic laws of the evolution of website market share of a certain product, etc., these methods technological innovation chain and is a powerful tool to get mainly create the hype cycle by adopting artificial methods. the overall grasp of technological innovation development The InSciTe[6,7] developed by Korea Institute of Science and trend, to get the objective assessment of the maturity of Technology Information adopts the decision tree and technological innovation, to make the reasonable choice on statistical feature analysis method, based on Gartner research, innovative intervention time and to seek the technological to provide the technical life cycle diagram and adopts the innovation late-mover advantage. emerging technologies discovering model to provide the key technologies on the life cycle diagram. In the particular judgment process, there might be problems of force compliance in certain stage. For example, a technology is during the plateau of productivity between the year from 2000 Copyright © 2015 for the individual papers by the papers' authors. to 2005, but its data between 2006 to 2007 is in line with the Copying permitted for private and academic purposes. stage of slope of enlightenment, so the force compliance is This volume is published and copyrighted by its editors. needed to judge the technology life cycle of the last two years[1]. Related reference didn’t include how specific Published at Ceur-ws.org technology term coordinates are obtained. Proceedings of the Second International Workshop on Patent Mining and its Applications (IPAMIN). May 27–28, 2015, Beijing, China. Therefore, the paper uses papers and patent information to realize emerging technology discovering method by machine learning; then acquires some features by feature selection and types of organizations make technology deployed. uses machine learning algorithm to classify and locate the Data from Internet between 2001 and 2014 about Gartner coordinates, and produces the hype cycle according to the Hype Cycle is manually collected on terminology, stage and prediction. coordinate to be used as training data. The paper is structured as follows. The prediction frame of technology maturity is described in Chapter 2. The learning method and model of technology maturity is introduced in 3.2 Feature Calculation Chapter 3. The prediction method and model of technology Technical features are the foundation of technology life cycle maturity is illustrated in Chapter 4. Experiment and analysis discovering model. The technical features calculation uses the are conducted in Chapter 5 and conclusion and forecasting are papers and patents as data, and uses paper made in the last chapter. index S ( Pp )  {Pp1 , Pp2 ,, Ppn } , patent index 2. TECHNOLOGY MATURITY S ( Pt )  {Pt1 , Pt 2 ,, Pt n } and combined index of PREDICTION FRAME paper and patent S ( Ppt )  {Ppt1 , Ppt 2 ,, Ppt n } as The model of approach for predicting hype cycle based on calculation objects. It can study the interaction and exclusion machine learning includes the learning part and the prediction between papers and patents, and explore the rule of part. The learning model mainly relates to data training. The development between science and technology. data acquiring and learning includes acquiring training data, data annotation and feature calculation. The prediction model Paper index includes paper growth k 1 means identification and discovering method of emerging k AN Pp  AN Pp technology in certain field and process to discover technology PpGrowthRate  k 1 system innovation by producing hype cycle and predict AN Pp maturity and discover an emerging technology, as well as to rate , paper relative k 1 make sure the input ratio of partial innovation and overall N k N Pp RelativeGrowthRate  Pp Pp innovation. It includes the term selection, feature calculation, k 1 stage classification, technology position and visible N Pp information module. The hype cycle prediction module is growth rate , paper k show in Fig. 1. A Pp AuthorRate  Pp k 1 AA Pp author rate , paper author growth k 1 A A k Pp AuthorGrowthRate  Pp Pp k 1 APp rate , paper institution k I Pp InstitutionRate  Pp k 1 AI Pp rate , and paper institution growth k 1 I k  I Pp Pp InstitutionGrowthRate  Pp k 1 I Pp rate . Patent index includes patent growth k 1 AN  AN k Pt GrowthRate  Pt Pt rate AN Ptk 1 , patent relative k 1 N N k Pt RelativeGrowthRate  Pt Pt growth rate N Ptk 1 , inventor k I Pt InventorRate  Pt k 1 rate AI Pt , inventor growth Fig.1 Technology maturity prediction frame k 1 AI  AI k Pt InventorGrowthRate  Pt Pt 3. TECHNOLOGY MATURITY rate AI Ptk 1 , application LEARNING k A Pt ApplicantRate  Pt k 1 3.1 Data Acquirement AA Pt rate , and application growth Since 1995, Gartner began to pay attention to the hype and disillusionment along with every appearance of new AA  AAPtk 1 k Pt ApplicantGrowthRate  Pt technologies and innovations, to track the trends of the technology life cycle, to study the common pattern between rate AAPtk 1 them, in order to provide guidance of when and where all Combined index of paper and patent includes paper and patent extracted from patents and papers, TLCD model can relative growth determine the stage of the emerging technology adopting the rate five stages of classification of TSKNN algorithm, and the k 1 k ( N Pp  N Pp )  ( N Ptk  N Ptk 1 ) specific stages refer to the Gartner's Hype cycle, which Ppt RelativeGrowthRate  includes Technology Trigger, Peak of Inflated, Expectations N Ptk 1  N Pp k 1 Trough of Disillusionment, Slope of Enlightenment, Plateau , of Productivity. k N SKNN algorithm is an improvement of KNN algorithm. KNN Ppt RatioRate  Pp k algorithm, by computing the distance between the training paper and patent ratio rate N Pt , paper and point from training set and test point from test set, considers patent people growth rate the closest distance having the most similar feature and can be k 1 classified into the same group, obtaining test markers k ( APp  APp )  ( I Ptk  I Ptk 1 ) characteristic points and the same tag feature training points. Ppt PeopleGrowthRate  k 1  I Ptk 1 SKNN mainly considers the time sequence issue of APp terminologies to be classified, so the data of next year need to , paper and patent people ration rate be larger than the data of last stage, in order to avoid the force k classification problem. The specific algorithm is as follows. APp Ppt PeopleRatioRate  (1) Redescribe training technology terminology and feature I Ptk , paper and patent institution vector, according to feature set. growth rate (2) When the technology terminology feature vector reaches, k 1 18 features should be calculated respectively to establish k ( I Pp  I Pp )  ( APtk  APtk 1 ) feature vector according to age. Ppt InstitutionGrowthRate  k 1 I Pp  APtk 1 (3) Select K technical terminologies which are most similar to , new technical year that is to be calculated from the training paper and patent institution ration rate technical terminology set following the formula below, k I Pp Ppt InstitutionRatioRate  M A k W  W k 1 ik jk (1) Pt . Sim (t i , t j )  M M ( Wik2 )( W jk2 ) k 1 k 1 4. TECHNOLOGY MATURITY PREDICTION (4) Among K neighbors of new terminology, weight of each classification is calculated respectively as follows, 4.1 Terminology Extraction Technology terminology refers to terms used in a certain field, which means concepts, features or relationships in the field. In     this paper, terminology extraction is based on keywords of p( x, C j )   Sim( x, d ) y(d , C ) d i STKNN i i j (2) papers and templates. The keywords are word or phrase extracted from papers to meet the needs of literature indexing or retrieval work. It is used to express the literature subject In the formula, x refers to feature vector of the new and therefore can be used as emerging technology terminology, Sim(x,di) refers to similarity calculation formula terminology. which is the same as mentioned in the last calculation formula, Template technology is a common method for terminology y(di,Cj) is categorical attributes function, if di belongs to Cj, recognition. By analyzing the characteristics of papers and then the function is equals to 1, otherwise is 0. patents, some fixed sentence structure are found, for example, (5) Compare the weight of various classifications; distribute ‘The development application of XX technology in XY ’, so the computable terminologies to the stage with the greatest ‘3D printing’ can be recognized in ‘The development weight. Record the stage. application of 3D printing technology in medical science’. After certain strings are extracted from the template, (6) Back to step 3 and calculate for next year. frequency and the subjection degree can be used to perform terminology recognition. If a collocation is found in corpus, it 4.3 Technical Position must appear more than once, thus the frequency is an Based on the classification results, PKNN algorithm important index in terminology extraction. Only when the calculates the position of terminologies on the hype cycle term frequency in the corpus exceeds a certain threshold value, curve and S-curve. The algorithm is to find K most similar it is believed that it has reached a technological terminology technologies in all classifications and the position of the standard and awareness in the field is relatively high. technical terminology is the average position of these Subjection degree refers to a relevant degree of the terminologies. terminology and its filed. It represents the degree of a term Input data includes trained technical terminology, feature belonging to a field. While meeting the frequency and vector and position under the classification, technical subjection degree at the same time, the string is a technology terminology and feature vector to be calculated. terminology. Output data includes technical terminology position. 4.2 Classify for Stage By calculating technical terminology and technical features Fig.2 Feature Extraction Algorithm steps are listed as follows. (3) Draw 5 stages line and label the text. (1) Select K technical terminologies which are most similar (4) Draw the technical terminology position according to the to the emerging technical terminology in the year to be position calculated by maximum similarity algorithm. calculated from the training technical terminology set. M 5. EXPERIMENT AND ANALYSIS  [sim( x , y) * r ] k k Gartner Emerging Technologies report describes some pred(y)  k 1 M (3) technologies that become famous because of hype, or  sim( x , y) k 1 k technologies that Gartner believes will have significant impact. In order to validate the technical life cycle discovery model, 960 training data released by ‘Hype Cycle Report’ from 2001 wherein K is 10. to 2013 is used for training technical maturity. Hype Cycle for (2) Among K neighbors of new terminologies, consider the Big Data, 2013 is used to evaluate validation set of prediction deviation between the most similar K terminologies and the results. emerging technical terminology, and calculate the prediction Using SKNN method to perform Hype Cycle model stage test, position of new terminology. Table 1 shows the experiment results of Hype Cycle for Big Data, 2013 data and technical life cycle model. Compared 4.4 Visualization of Hype Cycle Curve with Gartner, the technical life cycle model achieves the The paper uses fitting algorithm to produce Hype cycle. The precision of 67.24% and recall rate of 68.46%. The reason process is as follows. why the accuracy and recall rate in the fifth stage and the fourth stage is lower than that of other stages is that the (1) Draw coordinates of horizontal and vertical axis, and mark sample size in the fifth stage and the fourth stage is too small the expectation and time information. and the problem data has greater impact. (2) Produce Hype Cycle with curve fitting formula. Table 1. Technical life cycle discovering model experiment result Stage Result Gartner Suggested Number of Precision Recall approach results same in the both Technology Trigger 11 11 10 91% 91% Peak of Inflated 14 15 11 73.33% 78.6% Expectations Trough of 11 9 8 88.89% 72.7% Disillusionment Slope of Enlightenment 2 3 1 33% 50% Plateau of Productivity 2 2 1 50% 50% Total 40 40 33 67.24% 68.46% Using SKNN method to perform Hype Cycle model position data sets will be used to further improve the accuracy of the prediction, according to the Hype cycle visualization methods, prediction model. Hype Cycle for Big Data, 2013 prediction result is produced This paper is funded and supported by the and given as follows (This paper does not cover how each China Postdoctoral Science Foundation（2013M540125）. technology reaches the Plateau, and Gartner’s result is used in the following visualized graph ). 7. REFERENCES [1] Fenn J, Raskino M. Mastering the hype cycle: How to choose the right innovation at the right time[M]. Boston: Harvard Business School Press, 11-19, 2008. [2] Guo Daoquan. The Study of Technology Maturity Model and Assessment based on TRL[D]. Changsha: National University of Defense Technology, 2010. [3] JärvenpääH M, Mäkinen S J. An empirical study of the existence of the Hype Cycle: A case of DVD technology[C]. IEEE International Engineering Management Conference. Estoril, Europe, 257-261, 2008. [4] Budde B, Alkemadeb F, Hekkert M. On the relation between communication and innovation activities: A Fig.3 Gartner's Hype Cycle for Big Data,2013 comparison of hybrid electric and fuel cell vehicles[J]. Environmental Innovation and Societal Transitions, 101:1-15, 2013. 6. CONCLUSION AND FORCASTING [5] Steinert M, LeiferL. Scrutinizing Gartner's hype cycle[C]. The approach for predicting Hype Cycle based on machine Portland International Center for Management of learning discussed in this paper can effectively analyze paper Engineering and Technology, Portland, 2010. and patent information, forecast and produce Hype Cycle. The [6] Kim, J., Lee, S., Lee, J., Lee, M., & Jung, H. Design of approach consists of two processes of learning and training TOD Model for Information Analysis and Future and provides users with a more diverse information analysis Prediction. Communications in Computer and Information and forecasting tool. Compared with traditional methods and Science, 264(1): 301-305, 2011. services, the approach is more systematic and objective. [7] Jinhyung Kim, Myunggwon Hwang, Do Jeong , Hanmin Experimental results show that, compared to Hype Cycle for Jung.Technology trends analysis and forecasting applicati Big Data, 2013, the approach for predicting Hype Cycle based on based on decision tree and statistical on machine learning achieves the accuracy of 68.46% and feature analysis [J]. Expert Systems with Applications, 39, recall rate of 67.24% in forecasting stages, respectivly. 2012. Therefore, this model can provide more accurate information analysis and forecasting information to interested users，can [8] Wang Xin, Qiao Xiaodong, Xu Shuo, Han Hongqi, The Overview of Technology Life Cycle Analysis Method locate the position of technology accurately, and produce the Based on Factual Database[J]. Digital Library Forum. Hype Cycle automatically. [9] Bin Sun. A Summarization of Information Extraction (2) As for future work, considering the features of paper data and patent data, more features will be extracted and experiments (In Chinese). Terminology Standardization & Information in different databases to predict the experiment will be Technology, 2003. conducted; In addition, more machine learning data means more accurate results, so simulation experiments and different