Hierarchical Modeling Methods And Applications For Trend Forecasting

Trend forecasting task refers to the process of predicting future trends by analyzing and mining historical data for information and potential patterns of change.These tasks typically use time series analysis and machine learning models to predict future variables and have been widely studied and applied in both academic and industrial fields in recent decades.For example,in the economics and finance field,investors can make better investment decisions by predicting stock market trends.In the field of ecommerce,by predicting the user’s preference trend,merchants can recommend more suitable products.In the marketing field,companies can adjust product strategies more effectively by predicting sales trends.In the healthcare field,doctors can adjust medication prescriptions promptly by predicting patient health status.In summary,from the development of the national economy to the individual user experience,trend forecasting affects all aspects of people’s lives.Therefore,it has become a research hotspot in computer science and related interdisciplinary fields on how to use data mining techniques to better model trend forecasting tasks.Currently,although research on trend forecasting has achieved outstanding results and a large number of practical applications have made people deeply realize its convenience,the task still faces three research challenges.First,from the input data perspective,"data complexity and variability representation difficulties" exist.A large amount of historical data contains information while being interfered with by noise.Compared to massive and high-dimensional input data,the label data is relatively sparse,making it difficult to provide enough information to help representation training.Secondly,from the modeling perspective,"lack of interpretability in modeling methods" exists.As most existing deep learning-based trend forecasting methods are black-boxes that only output prediction results and lack interpretability,making the persuasiveness of the prediction results lack,thereby affecting user decision-making.Finally,from the prediction performance perspective,"insufficient prediction accuracy" exists.The accuracy of prediction results is the core of trend forecasting.However,in many complex application scenarios,the prediction results are still unsatisfactory and need improvement.In trend forecasting tasks,data can be obtained through sampling at different granularities(such as monthly,daily,minute-level,etc.),and long-term coarse-grained data reflects the overall trend,while short-term fine-grained data contains sudden changes in a short period of time.Different granularities of data are of great importance to the final prediction results,and this dissertation proposes to integrate multi-granular data into hierarchical modeling to improve the performance of trend forecasting tasks.In view of the challenges of complex and changeable data representation,modeling methods with poor interpretability,insufficient accuracy of prediction results,etc.,this dissertation systematically carries out research on multi-granular hierarchical modeling methods and applications for trend forecasting.The key findings and contributions of this dissertation can be summarized as follows:(1)We present a trend prediction method with cross-level representation enhancement based on self-supervised pre-training to address the first challenge.The problem of effectively extracting informative representations from multi-grained data remains an unresolved issue,as training labels provide insufficient signals for extracting effective multi-grained representations.To solve this problem,we propose a cross-grained contrastive learning mechanism based on local scopes and a cross-temporal contrastive learning mechanism based on global scopes,to build additional self-supervised pretraining objectives to help learn more effective multi-grained representations.Specifically,the intrinsic mechanism of the cross-grained contrastive learning mechanism enables feature encoders to capture semantic associations between coarse and fine grained data,while the cross-temporal contrastive learning mechanism attempts to use the coherence of historical trends and current status of time-series data to construct additional self-supervised training objectives and help pre-train representations through maximizing mutual information.In the scenario of quantitative financial trend prediction,we further utilize commonly used technical indicators in the financial field to describe market state and designs a multi-grained fusion module based on gating mechanism to adaptively fuse multi-grained data based on the market state.Experiments were conducted on three real-world quantitative financial trend prediction datasets,and the results show that the proposed method has significantly improved results compared to other baseline methods,and visualization analysis of the representations further verifies the effectiveness of the proposed pre-training method.(2)We present an explainable user preference hierarchical modeling method in trend prediction task to overcome the second challenge.In the recommendation scenario,it is necessary to analyze and mine the potential user preferences in the user’s historical shopping interaction records,and predict the products that the recommended user may purchase in the future.The traditional user preference modeling method usually only models the overall preference of users with coarse granularity,and lacks the exploration of fine-grained semantic attribute preferences.Most of the previous research work projected users and historical items into a latent vector space in the modeling process.The meaning of each dimension in this space is unknown and lacks interpretability.We propose a hierarchical modeling method of explainable user preferences based on fine-grained semantic attribute modeling.Firstly,we introduce a explainable finegrained semantic attribute space.Each dimension in this space corresponds to a specific fine-grained semantic attribute.The model captures the user’s fine-grained preferences and generates explainable preference recommendation results by projecting the user and the history items into the space.This work first design a semantic attribute extraction network,which extracts the fine-grained semantic attribute representation of items in a weakly supervised manner,and projects the items into the semantic attribute space through the network.Then,in order to capture the user’s preference for each finegrained semantic attribute,we design a fine-grained semantic attribute attention module to automatically match semantic attributes and user preferences and conduct weighted aggregation.Finally,the user preference and item representation at the two levels of fine-grained semantic attribute and coarse-grained product whole are integrated.With this method,we are capable of not only providing recommendations for users,but also explaining the reason why we recommend the item through intuitive visual attribute semantic highlights in a personalized manner.Extensive experiments conducted on realworld datasets clearly demonstrate the effectiveness of our approach compared with the state-of-the-art methods.(3)To address the challenge of insufficient accuracy in prediction,we propose a general trend prediction method of multi-granularity hierarchical joint modeling.Different granularities of information play a crucial role in the accuracy of the final prediction.However,effective hierarchical modeling of multi-granularity time series data to improve prediction accuracy still faces some problems.To address the problem of severe information redundancy among multi-granularity data,we present a crossgranularity residual learning network architecture containing multiple blocks with similar structures,each responsible for learning information from data of a specific granularity.The neighboring blocks are connected using a novel residual learning approach to eliminate data redundancy.To tackle the problem of varying effectiveness of different granularities of data over time,we propose a multi-granularity confidence estimator to judge the effectiveness of each granular data on the final prediction result,constructed through self-supervised learning.Extensive experiments on real-world electricity and stock datasets show the accuracy of the proposed framework in prediction.