Modeling Techniques For Context Based Trend Forecasting

Trend forecasting is a kind of task to predict the future by extracting and analyz-ing useful information and patterns from data.It has long been popular in industrial and academia,and has played a crucial role in many application scenarios.For ex-ample,in the online consumption model that has emerged in recent years,merchants will pay more attention to how much inventory should be maintained to meet customer demand.In the financial field,investors hope to predict the rise and fall of stocks in advance to make investment and financial strategies.In healthcare field,doctors hope to judge in advance which treatment plan will have the best effect on patients' recov-ery in the future.In a word,the research and application of trend forecasting are of great significance for detonating the Internet economy,formulating business strategies for intelligent commerce,and improving service satisfaction.Since different applica-tion scenarios have different concerns on the accuracy,stability,and interpretability of trend prediction,how to use data mining techniques to analyze and model different sit-uations to meet the needs of trend prediction in specific situations has become a hot research topic in computer science and related interdisciplinary subjects.Recently,although a variety of trend forecasting methods have been applied and achieved certain results,the research of trend forecasting methods in a specific con-text still faces the challenges of multiple data relationships,uncertainty,and poor in-terpretability.To this end,this dissertation systematically carries out the exploratory research on the trend prediction modeling method with the integration of situational in-formation.In particular,in terms of accuracy,a demand prediction algorithm based on situational multiple relationships is proposed to solve the problem that multiple rela-tionships are difficult to be effectively integrated and represented.In terms of stability,a time-series prediction algorithm is proposed to solve the problem of a complex and changeable future state and uncertainty in the prediction process.In terms of inter-pretability,a time-series prediction algorithm based on situational causality is proposed to solve the problem of data bias.The main contributions of this dissertation can be summarized as follows:First,from the perspective of accuracy,we study the demand prediction algorithm of situational multiple relations.With the abundance of data acquisition channels nowa-days,trend prediction of products is often faced with multiple data relationships,which is called situational multi-relation.Based on the situation of product demand prediction,such a multi-relationship includes the internal relation and the external relation of the product.Since the hierarchical relationship and cooperative and competitive relation-ship within products are directly related to the future trend,how to model the role of internal relation in the future trend predictions has become one of our research points.In addition,there is a relationship between the product and the market in the external relations of the product,and the high liquidity of the market will have an impact on the future trend of the product.How to model the trend under the change of the en-tire market environment has become another focus of our research.To this end,in the multi-relation context-based trend forecasting,we first analyzed the relationship be-tween products,proposed a regression prediction model based on multi-task learning,and introduced the portfolio theory in economics by considering the characteristics of application scenarios,which effectively optimized the allocation of product quantity.Then,we constructed a dynamic graph structure of products and markets and proposed a recursive graph neural network algorithm based on attention mechanism and adap-tive technology,so to achieve the purpose of real-time product demand forecasting.The experimental results showed that by modeling the internal and external relations of products,the accuracy of trend prediction has been significantly improved,and the rational allocation of resources has been improved.Second,from the perspective of stability,we study the time-series prediction model of situational uncertainty.Trend prediction is to predict the future situation based on past historical data,as for trend forecasting in real-time changing scenarios,an intuitive challenge is that there are many uncertainties.We call it situational uncertainty.The most direct cause of uncertainty lies in the data level.On the one hand,it is the fact that future data cannot be observed.On the other hand,the randomness of observable sequence data itself leads to the uncertainty of the data level in trend forecasting.Be-sides,in addition to the uncertainty in the data level,the prediction model itself also has a certain degree of uncertainty.For example,the most widely used supervised model,whose training characteristics of relying on future target data introduces uncertainty to the model.To this end,aiming at the uncertainty in the real-time changing situation,we both consider from the perspective of data uncertainty and model uncertainty.Specif-ically,we proposed a self-supervised representation model based on the technique of coupling functions in statistics,then provided a high-dimensional representation of the future state of the product,which further improved the stability of downstream trend prediction tasks.Third,from the perspective of interpretability,we study the time-series prediction algorithm of situational causality.With the continuous development and application of deep learning methods in trend prediction,people also have certain requirements on the interpretability of prediction models,especially in the face of trend forecasting tasks in decision-assisted situations.In addition to the interpretability of the model itself,a leading concern in trend forecasting is the desire for the model to understand the impact of different decisions on future trends.Especially in decision aid,we need to under-stand the predicted causality,which we call situational causality.Therefore,from the perspective of causal inference,we aim to estimate the individual effect on trend fore-casting for inferencing the sequential data.Different from the method that only focuses on the prediction loss,the proposed method can predict the potential outcome under the action of multiple intervention schemes,and solve the selection bias problem in causal inference and the time-varying bias problem in the time-series data.Experiments were carried out on the data of games and the medical field,whose results verified that the proposed model was helpful to understand the changes of future trends under different decisions,and provided an auxiliary role for the making of intelligent decisions.