Exploration of the application of social data in the field of credit reporting

At the WOT "Internet +" era big data technology summit, Liu Lichun, a senior data mining engineer from Tencent, gave a speech entitled "Exploration of the Application of Social Data in the Credit Reporting Field". The main content consisted of five parts: social credit reporting background, Tencent social network data, individual user portrait research, community circle research, model construction and application. Let us introduce each part one by one.

Social credit background

Liu Lichun said that credit investigation is not a simple credit scoring model, but consists of three parts: data companies, credit investigation companies, and credit users. Data companies collect or do some preliminary mining of data. Such companies may have special data sources, such as courts, public security, etc. These data require in-depth industry background to obtain. Credit investigation companies have a property right connection. In addition, they will also purchase some data from some third-party data companies to enrich the dimensions of their data, and do some credit investigation based on these data, and provide some credit-level solutions. Credit users are who will use the credit investigation solutions in the end. Generally speaking, our understanding is banks and P2P lending institutions. These three parts combined form an overall industrial chain of the credit investigation industry.

Traditional credit reporting agencies

Famous American credit reporting company

Development History of Domestic Credit Reporting

Based on the data in the above four figures, if social data can be used in credit reporting , can it be a good supplement to the central bank's credit reporting system? Liu Lichun said that this was the question Tencent first thought about when doing the social credit reporting project. Social data is very large, but it is not necessarily valid data. It also depends on whether the business scenario of the specific application is relevant to the data, and whether this data can really be used in the final model or algorithm. This raises questions one after another. Is social data related to credit rating? Transaction data naturally has financial attributes, does social data have them? Social data is highly unstructured, how to mine and use it effectively?

Before talking about the composition of Tencent's social network data, Liu Lichun first introduced the analysis dimensions of traditional credit reporting. The first is the user's basic information, such as age, gender, occupation, income, marital status, years of work, work status, etc., which are basically the same as the data obtained by each bank or each credit reporting agency. The second is the credit situation, which looks at how many credit cards the user has applied for and the number of times the credit report has been queried in the past month, because we all know that the number of times the credit report has been queried can directly represent whether there have been more frequent loan applications or credit card applications recently. If the number of recent inquiries is particularly high, it means that the person is very short of money recently, which may affect the credit and directly affect the credit limit.

Tencent Social Credit SWOT Analysis

The above figure is a SWOT analysis of Tencent social credit reporting, with strengths, weaknesses, opportunities and risks clearly shown. With such a detailed analysis, it is inevitable to do personal credit reporting, but before doing credit reporting, we must clearly know what the credit reporting object looks like, so we started to study individual user portraits.

Individual User Portrait Research

Liu Lichun said that the challenges encountered in conducting individual user portrait research are mainly the following three aspects: First, how to make full use of Tencent's rich data resources and the connections between them? Second, how to make user portraits adapt to various application scenarios? Third, how to efficiently process massive user data (more than 1 billion QQ users, more than 100 billion levels of various log data)? In the face of these challenges, Liu Lichun gave the following corresponding solutions:

1. Design specific mining algorithms for different underlying data types, mine user behavior characteristics, and form underlying labels. Comprehensively consider different data sources to form higher-level abstract user labels

2. Establish a complete user portrait label system architecture to describe users from different dimensions and granularities.

3. Build a user portrait mining system based on a large-scale storage and machine learning computing platform, regularly calculate and mine all user data, and provide user tag usage and query services.

User portrait system architecture

User portrait text mining system

User portrait industry mining

User portrait mining results

The result of the personal user portrait research is to form a relatively complete portrait after mining structured data, text classification, LBS data, and social network communication and diffusion, such as some basic attributes of the population such as age, hometown, interests, etc. At the same time, a judgment will be made on the user's marital status. With these data, a lot of social credit work can be done based on these user data.

Community Circle Research

The community circle mentioned here is actually the QQ circle. Liu Lichun said that in 2012, there was a social network achievement that was very influential, which was to apply the mining results to the entire front-end QQ users. A specific example is that if you are not directly friends with a colleague of a user, Tencent will know the potential relationship during this period, or automatically group them into colleagues and add notes. This result caused a lot of controversy at the time. Some people felt that it provided convenience for them to find some potential friends, but some people felt that it touched their privacy.

In addition to its own use, QQ circles are also used in many other scenarios. For example, it is used to mine educational information. Based on the notes of friends in QQ circles, if many people note this user as an undergraduate classmate, the system may judge that my educational background is a bachelor's degree. Tencent has verified this data with some real data, and the data coverage rate is about 74%, with an accuracy of more than 90%.

Applications of social network topology

There are two main applications of social network topology: one is to determine the type of topology, and the other is to study the influence of these types in this relationship chain. The more iconic topological types are triangle and heart-shaped structures.

Model construction and application

So how to apply the research on individual user portraits and social circles to the model? Liu Lichun said that the first thing to do is to build a social model, but some basic assumptions must be made before modeling. For example, if two QQ numbers belong to the same person, there are some obvious characteristics. The first is that he often logs in from the same device, or logs in from the same IP, or has other characteristics, etc. Finally, these characteristics are used to build a model to determine whether the QQ numbers behind are the same person. The accuracy rate is about 85%, and the coverage rate is about 75%.

Variable Derivation and Model Results

Overall effect of the model

Weilidai Application

Finally, Liu Lichun introduced the specific application process of the credit investigation model in Weilidai. The above picture is a screenshot of the product. If you can see the Weilidai entrance when you open QQ, it means that it is in the whitelist selected by Tencent. As long as you click to apply for activation, it will immediately calculate a credit limit for you. If you want to borrow money, this is also very fast. As long as you bind your bank card, your loan should be transferred to your account within two minutes. In fact, compared with borrowing from traditional banks, this efficiency is a qualitative leap. But the simpler the front-end product appears, the more complex the technology behind it may be. The credit investigation model, as the technology behind Weilidai, is to screen users with good credit and provide loan services to these users.

author:
Liu Lichun is a senior data mining engineer at Tencent, the head of the data mining team of the social network business group, the architect of the first generation of QQ music recommendation system, and the builder of Tencent's customer lifecycle management system. He has been committed to combining data mining technology with business for many years and has rich practical experience and project management experience in big data analysis and mining, Internet credit reporting and other fields. He currently focuses on Internet credit reporting, user basic profiling, recommendation systems and text mining.