Jointly Structure Learning And Pattern Prediction For Time Series Data

Time-series data analysis is an important technology in the field of data mining. In the real-world applications, this technology is applied to not only discovering the latent developmental pattern or the evolutionary role of the time-series data, but also predicting the future developmen-tal trend of the given testing data. Time-series data analysis is usually composed of two research tasks:structure learning and pattern prediction. In detail, the objective of structure learning is investigating the relationship of the time-series data samples to extract the novel sample features and the novel sample labels; the objective of pattern prediction is building an effective mapping between the sample features and the sample labels. Compared with the typical approaches of time-series data analysis often regard the above two research tasks as the independent ones, our literature considers that there exists the latent correlation between structure learning and pattern prediction. Further, this literature demonstrates that these two research tasks are mutually supportive and mu-tually reinforcing in the views of the theoretical analysis and the model design. Based on the detailed analysis of the problems and the challenges in the real-world and the deep survey of the existing advanced efforts in the related literatures, our literature proposes several novel approaches of time-series data analysis, which are all capable of jointly implementing the structure learning model and the pattern prediction model. The major content and advantages of this literature are demonstrated as follows,1. This literature proposes a novel business model mining approach based on the analysis of the consumers’social influence powers and then applies this approach to solve a real problem of the women-cloth sales prediction in Alibaba E-commerce platform. Specifically, this ap-proach first uses the commodity data and the consumer data of the E-commerce platform to extract the commodity features and the consumer features and then analyzes the relationships among the commodities and the social influence powers among the consumers. Next, con-sidering that the real sales change both exists the smoothing varying and the drastic varying, this approach splits the sales into two parts (a true part and a noise part) and then designs the corresponding two regression models. Finally, this approach builds an effective optimization model to jointly implement the mining of the social influence powers and the prediction of the women-cloth sales.2. This literature proposes a novel age estimation approach based on the technology of the structure-aware slow feature analysis. Specifically, this approach first constructs a structural face sequence based on combining the face temporal structure and the face content structure, which is capable of representing the relationships among the face samples. Next, this ap-proach extracts the slow-varying features of the samples based on the above structural face sequence. Finally, this approach builds an effective optimization model to jointly implement the construction of the structural face sequence and the prediction of the age labels.3. This literature proposes a novel age estimation approach based on learning the data-dependent label distribution. Specifically, this approach first uses the subspace learning technology to discover the face content structure. Next, this approach builds the data-dependent label dis-tributions based on the learned face content structure. In addition, our label distribution is helpful in understanding the cross-age correlation among the content-neighboring samples and analyzing the problem of the age label ambiguity. Finally, this approach builds an effec-tive optimization model to jointly implement the discovery of the face content structure and the prediction of the label distributions.