Credit Footprint: Credit Assessment Based On User Trajectory Data

Facilitating inclusive financial services for increasing overall social good is one of the hottest worldwide topics nowadays.Vigorous development of inclusive finance is also an inevitable requirement for building a moderately prosperous society in all respects,which also is conducive to the sustainable and balanced development of the financial industry and widespread Entrepreneurship and Innovation,boosting the transformation and upgrading of economic development mode and contributing to social equity and justice.Credit investigation is critical for financial services.Whereas,traditional methods are often restricted as most of the employed data are strong financial attributes,such as the career,revenue and Historical consumption and loan history.Nowadays,lowincome people in rural areas and special groups such as the disabled and the elderly have less access to financial services and fewer historical data records.So traditional methods hardly provide sufficient,timely,and reliable information to a large coverage of people,thus not being able to support inclusive finance well.How to promote the development of inclusive finance and study the issue of credit investigation from the perspective of inclusive finance still needs to find new directions and methods.With the advent of the 5G era and the popularity of smart mobile devices,the mobile Internet has grown rapidly in recent years.The rich data generated by various types of mobile devices has become a research hotspot in the field of data mining and machine learning.The research directions for this kind of data are diverse and cover a wide range of fields,including urban management,geographic location-based services,etc.Different from the existing research problems,this dissertation studies the relationship between user behavior patterns and their credit levels based on users' trajectory data,and uses deep learning methods to learn trajectory embeddings for credit prediction.This is an extension of the existing credit evaluation system and an exploration of new application scenarios of trajectory data.This paper studies two users' trajectory datasets collected from Hangzhou and Shanghai,China and contain 4,779 and 7,015 users,respectively.The total track number is over 25 million.Therefore,we first generalized the data from two dimensions of time and space during preprocessing.In space,a uniform grid generalization method is used to convert the precise latitude and longitude of the user's location at a certain moment into fixed-size region on the premise of determining the range of all trajectory data.In terms of time,taking into account the uneven sampling of trajectory data caused by the users' habits of using mobile phones and behavior patterns,the user's daily trajectory is normalized into a time series of 48 trajectory points with half an hour as the sampling frequency.On this basis,we conducted further exploratory analysis on trajectory datasets and visualized the analysis results.The analysis results showed that the users' credit levels are related to the regions they visited.Inspired by this result,we designed a two-stage credit investigation framework based on deep learning methods with users' geographic footprints,namely Credit Print.In the first stage,we need to obtain region credit-aware embeddings.Considering that regions don't have explicit features which can reflect visiting people's credit levels,we generated different types of graphs,including distance graph,interaction graph and correlation graph.Then we explored regions' credit characteristics and learns a credit-aware embedding for each region by considering both regions' individual characteristics and cross-region relationships with graph convolutional networks.In the second stage,a classification model is built to predict users' credit levels.A hierarchical attention-based credit assessment network is proposed to aggregate the credit indications from a user's multiple trajectories covering diverse regions.Then we concatenate the trajectory embedding and manually extracted features to predict users' credit levels.Finally,we use users' trajectory data to conduct a large number of experiments to verify the effectiveness of Credit Print.In the experiments,we first compared Credit Print with some machine learning methods using manually extracted features.Since Credit Print incorporates deep learning embedding,it increases AUC of credit prediction by more than 10%.We then verified important components in the framework one by one,for instance,multi-graph fusion,loss functions,attention mechanisms,and region and trajectory embedding' dimensions.According to the results,it can also be found that removing or replacing any component would reduce the performance of the Credit Print model.At last,we selected two specific users for further observation and analyzed the interpretability of the model based on the attention mechanism.We found that regular activities in the daily trajectories of users are more important for the prediction of users' credit,especially activities at night and users' family addresses.This is also in line with common sense that the regular activities of users can better reflect the long-term stable status of users,and the location of family can also reflect the economic situation of users to a certain extent.