ºÝºÝߣ

ºÝºÝߣShare a Scribd company logo

Bigdata and ai in p2 p industry: Knowledge graph and inference

?
?
S
sfbiganalytics

The document discusses how Puhui Finance, a Chinese P2P lending company, uses big data and AI techniques for risk control. It introduces their Feature Compute Engine, which converts unstructured user data into structured features, and their Knowledge Graph, which connects entities and analyzes relationships. Specific use cases discussed include anti-fraud detection using rules, contact recovery by building phone networks, and detecting high-risk individuals via search engines. Challenges around unstructured data, name disambiguation, reasoning and lack of training data are also covered.

1 of 59
Downloaded 112 times
Big Data and AI in P2P Industry Wenzhe Li nadalwz1115@gmail.com Feb 1, 2016
Puhui Finance (www.puhuifinance.com) Services °®Ç® ½ø ÆÕ»Ý ÐÅ´û ´´Ð ×ʲú ÆÕ»Ý ²Æ¸» ? Internet Financing P2P company, headquarters in Beijing ? Founded in July 2013 ? $50M series A funding in Dec 2014 ? ~5500 employees, 100+ offline stores Offline Financing Service Online Financing Service Online Lending Service Offline Lending Service
Puhui Finance (cont.) Fastest growing p2p company. Big data technology is the key
In this talk, I will mainly focus on the techniques used in lending side risk control. Similar techniques can be applied to the financing side. What the talk is about
Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
? Credit system is not mature in China ? Targeting at under-served market, those who don¡¯t have enough credit to borrow from bank ? The data solely from credit history is not enough to build the scoring models ? More efficient application reviewing process is needed as we move more transactions from offline to online Why big data & AI
Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
The central problem is risk control The solution is to use big data
Measure the risk for a person Individual Feature Analysis Relation Analysis £¿ Knowledge Graph Feature Compute(FC) Engine
? User explicitly input data (i.e. application form) ? Authorized* user data ? Mobile History ? Purchasing History ? ¡­¡­ ? Open Search ? ? 360.com ? Others (i.e. craigslist) ? 3rd- party data (i.e. blacklist) Data Unstructured Data * User authorizes us to use their data
Feature Compute Engine The goal is to convert unstructured data to structured features
Feature Compute Engine Credit Card Mobile History Purchasing ...... Precision Marketing Fraud Score Risk Score FeatureCompute Engine Feature Container (tens of thousands) Data .... .... Data Credit Card History Mobile History Purchasing History Feature Compute Engine Data Scoring Model
Purchasing History i.e. Purchasing History Total amount spent during the last 6 months User level (i.e. Prime, Normal¡­) Total number of transactions during the last 6 months The length of time he/she uses the account Total number of transactions related to virtual products Total number of transactions related to luxury products ¡­¡­¡­ Few thousand features
? It is a semantic network ? Based on graph data structure, consists of points and edges. Point represents entity, edge represents relationship. ? Knowledge graph connects heterogeneous information. It provides the ability to analyze the data from the perspective of relationship. What is knowledge graph
Some knowledge graphs
Knowledge graph ¨C search engine
Knowledge graph ¨C search engine
Knowledge graph ¨C recommendation [1]
Storing Knowledge graph Ranking DBMS 21 Neo4j (Graph Database) 32 MarkLogic (XML) 42 Titan (Graph Database) 46 OrientDB (Graph Database) 61 Virtuoso (RDF) 80 Jena (RDF) 88 Sesmae (RDF) 90 ArangoDB (GraphDatabase) 120 AllegroGraph (RDF) Trends for different types of database [2] Graph/RDF database ranking [3]
? Logic-based approach ? Probabilistic approach (i.e. distributed representation) ? Hybrid approach Key techniques for knowledge graph Link Prediction
Simple Approach: Pre-define some rules i.e. (Peter FatherOf Tom) -> (Tom SonOf Peter) (Peter ColleagueOf Tom), (Sarah ColleagueOf Peter) -> (Peter ColleaugeOf Sarah) Logic-based approach
Methods based on distributed representation ? Translating Embedding [4] ? Tensor Factorization (RESCAL) Hybrid approach [5] ? Neural Tensor Network (NTN) [6]
Hybrid Approach ¨C Logic + Probabilistic Simple Approach: 1. Generating all the new links using pre-define rules 2. Apply Statistical Learning Advanced Approach (i.e.): ? Incorporation of Rules into Embeddings [7] ? Injecting Logical Background [8]
Use Cases
Connects person, phone, address, email, company¡­¡­ Domain-specific knowledge graph
? 10 types of entities ? ~50 types of relations ? ~50M entities ? 0.2B relations We expect that it will become ~20 times bigger by the end of this year due to the business growth Domain-specific knowledge graph
Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
Applicant shares the same personal phone with other applicant Phone Applicant Other applicant Personal Phone Personal Phone Antifraud - rules
Applicant and other applicant share the same colleague phone, but with different company names Phone Applicant Other applicant Colleague phone Company 1 Company 2 Colleague phone Antifraud ¨C rules (cont.)
Phone Applicant Personal phone Phone Phone Phone Phone Phone Overdue Overdue Some of the applicant¡¯s contacts didn¡¯t pay back the loan on time Antifraud ¨C rules (cont.)
Person 2 Person 1 Triangle relationship Person 3 Antifraud ¨C cycle detection
Applicant Applicant 2 Parent of Parent of Applicant 1 Spouse Inconsistent relations Antifraud ¨C inconsistent relationship
Antifraud ¨C suspicious group Person 2 Person 1 Person 3 Share a lot of common attributes
Knowledge Graph Visualization ? Visualize entities and relationships ? Design anti-fraud rules via observational study Antifraud ¨C design by observation
Rapid change of relationship structure within short time period Antifraud ¨C evolution of graph structure
LR Decision Tree Random Forest SVM ANN Models Prediction Extracted Features from Raw Data Results from anti-fraud rules User direct attributes Variables DNN Score is used to directly reject or accept the loan Antifraud ¨C fraud score score
Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
The borrowers disappear, all the contact information they explicitly provided become invalid. How to reach them? Lost contact recovery ¨C what is it Implicitly infer potential contact information
Phone Applicant Personal phone Phone Phone Phone Phone Phone Rank the phone numbers, and predict relationship Building phone network ¨C 1st order extension
Building phone network ¨C 2nd order extension Phone Applicant Personal phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Rank the phone numbers, and predict relationship
3rd order .. Phone Applicant Personal phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Phone
Simple Ranking Criteria ? The total length of time ? The frequency of calls Advanced Approach ? Learning the ranking score using machine learning approach Building phone network ¨C Rank
? Total # of times of calling ? Total length of time of calling ? Total # of times of being called ? Total # of times of calling ? Average time per call ? Maximum length of time ? # of times of calling between 0-4am ? # of times of calling between 4-8am ? ¡­¡­ Building phone network ¨C Predict the relation LR Decision Tree Random Forest SVM ANN Models Prediction of relation ~100 Features DNN Relation With very limited training data, our model provides ~30% accuracy
Person Applicant Personal phone Person Other applicant knows£¿ Other approach ¨C Link prediction (on-going work) Link Prediction
Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
Detect Bad People via Search From the search results, we label each entities in the knowledge graph i.e. black, green etc.
? ? 360.com ? other public websites Search for basic information¡­. ? Phone number ? Email ? QQ ? Other IDs Search Fields Search Engines & Public Site
Search for phone ²Ô³Ü³¾²ú±ð°ù¡­
Search for Email¡­ Fraud
? Clustering analysis ? Precision marketing ? ¡­¡­ Other Applications we are working on
Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
Challenges : Unstructured Data Unstructured Data Images Text AudioVideo Machine Learning Natural Language Processing Data Mining
Challenges : Name Disambiguation Applicant Other applicant Puhui Finance Ltd. Puhui Finance Same company, can we merge? It is a very important problem to deal with!
Challenges : Reasoning However, It is still an open problem ? Logic-based approach ? Probabilistic approach (i.e. distributed representation) ? Hybrid approach Link Prediction
Challenges : Insufficient Samples Big data, but small samples
? Senior/Lead Machine Learning/NLP Engineers ? Senior/Lead Data Engineer/Scientist ? Senior/Lead Architect ? Senior/Lead Software Engineer liwenzhe@puhuifinance.com zhaopin@puhuifinance.com We are hiring! (in Beijing) Open positions, but not limited to¡­. Contact Company Website www.puhuifinance.com
Email£º nadalwz1115@hotmail.com nadalwz1115@gmail.com Wechat£¨Î¢ÐÅ£©£º liwenzhe595675 Thanks!
[1] http://www.datapop.com/ [2] http://db-engines.com/en/blog_post//43 [3] http://db-engines.com/en/ranking [4] Bordes, Antoine, et al. "Translating Embeddings for Modeling Multi- relational Data." Advances in Neural Information Processing Systems(2013):2787-2795. [5] Nickel, Maximilian, V. Tresp, and H. P. Kriegel. "A Three-Way Model for Collective Learning on Multi-Relational Data.." International Conference on Machine Learning 2011:809-816. References
[6] Richard Socher, Danqi Chen, Christopher D. Manning, Andrew Ng. Reasoning With Neural Tensor Networks for Knowledge Base Completion. Advances in Neural Information Processing Systems(2013) [7] Wang, Quan, Wang, Bin, and Guo, Li. "Knowledge base completion using embeddings and rules." Proceedings of the 24th International Conference on Artificial Intelligence AAAI Press, 2015. [8] T Rockt?schel£¬S Singh£¬S Riedel. Injecting Logical Background Knowledge into Embeddings for Relation Extraction http://talks.cam.ac.uk/talk/index/58360 References

More Related Content

Bigdata and ai in p2 p industry: Knowledge graph and inference

  • 1. Big Data and AI in P2P Industry Wenzhe Li nadalwz1115@gmail.com Feb 1, 2016
  • 2. Puhui Finance (www.puhuifinance.com) Services °®Ç® ½ø ÆÕ»Ý ÐÅ´û ´´Ð ×ʲú ÆÕ»Ý ²Æ¸» ? Internet Financing P2P company, headquarters in Beijing ? Founded in July 2013 ? $50M series A funding in Dec 2014 ? ~5500 employees, 100+ offline stores Offline Financing Service Online Financing Service Online Lending Service Offline Lending Service
  • 3. Puhui Finance (cont.) Fastest growing p2p company. Big data technology is the key
  • 4. In this talk, I will mainly focus on the techniques used in lending side risk control. Similar techniques can be applied to the financing side. What the talk is about
  • 5. Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
  • 6. ? Credit system is not mature in China ? Targeting at under-served market, those who don¡¯t have enough credit to borrow from bank ? The data solely from credit history is not enough to build the scoring models ? More efficient application reviewing process is needed as we move more transactions from offline to online Why big data & AI
  • 7. Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
  • 8. The central problem is risk control The solution is to use big data
  • 9. Measure the risk for a person Individual Feature Analysis Relation Analysis £¿ Knowledge Graph Feature Compute(FC) Engine
  • 10. ? User explicitly input data (i.e. application form) ? Authorized* user data ? Mobile History ? Purchasing History ? ¡­¡­ ? Open Search ? ? 360.com ? Others (i.e. craigslist) ? 3rd- party data (i.e. blacklist) Data Unstructured Data * User authorizes us to use their data
  • 11. Feature Compute Engine The goal is to convert unstructured data to structured features
  • 12. Feature Compute Engine Credit Card Mobile History Purchasing ...... Precision Marketing Fraud Score Risk Score FeatureCompute Engine Feature Container (tens of thousands) Data .... .... Data Credit Card History Mobile History Purchasing History Feature Compute Engine Data Scoring Model
  • 13. Purchasing History i.e. Purchasing History Total amount spent during the last 6 months User level (i.e. Prime, Normal¡­) Total number of transactions during the last 6 months The length of time he/she uses the account Total number of transactions related to virtual products Total number of transactions related to luxury products ¡­¡­¡­ Few thousand features
  • 14. ? It is a semantic network ? Based on graph data structure, consists of points and edges. Point represents entity, edge represents relationship. ? Knowledge graph connects heterogeneous information. It provides the ability to analyze the data from the perspective of relationship. What is knowledge graph
  • 16. Knowledge graph ¨C search engine
  • 17. Knowledge graph ¨C search engine
  • 18. Knowledge graph ¨C recommendation [1]
  • 19. Storing Knowledge graph Ranking DBMS 21 Neo4j (Graph Database) 32 MarkLogic (XML) 42 Titan (Graph Database) 46 OrientDB (Graph Database) 61 Virtuoso (RDF) 80 Jena (RDF) 88 Sesmae (RDF) 90 ArangoDB (GraphDatabase) 120 AllegroGraph (RDF) Trends for different types of database [2] Graph/RDF database ranking [3]
  • 20. ? Logic-based approach ? Probabilistic approach (i.e. distributed representation) ? Hybrid approach Key techniques for knowledge graph Link Prediction
  • 21. Simple Approach: Pre-define some rules i.e. (Peter FatherOf Tom) -> (Tom SonOf Peter) (Peter ColleagueOf Tom), (Sarah ColleagueOf Peter) -> (Peter ColleaugeOf Sarah) Logic-based approach
  • 22. Methods based on distributed representation ? Translating Embedding [4] ? Tensor Factorization (RESCAL) Hybrid approach [5] ? Neural Tensor Network (NTN) [6]
  • 23. Hybrid Approach ¨C Logic + Probabilistic Simple Approach: 1. Generating all the new links using pre-define rules 2. Apply Statistical Learning Advanced Approach (i.e.): ? Incorporation of Rules into Embeddings [7] ? Injecting Logical Background [8]
  • 25. Connects person, phone, address, email, company¡­¡­ Domain-specific knowledge graph
  • 26. ? 10 types of entities ? ~50 types of relations ? ~50M entities ? 0.2B relations We expect that it will become ~20 times bigger by the end of this year due to the business growth Domain-specific knowledge graph
  • 27. Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
  • 28. Applicant shares the same personal phone with other applicant Phone Applicant Other applicant Personal Phone Personal Phone Antifraud - rules
  • 29. Applicant and other applicant share the same colleague phone, but with different company names Phone Applicant Other applicant Colleague phone Company 1 Company 2 Colleague phone Antifraud ¨C rules (cont.)
  • 30. Phone Applicant Personal phone Phone Phone Phone Phone Phone Overdue Overdue Some of the applicant¡¯s contacts didn¡¯t pay back the loan on time Antifraud ¨C rules (cont.)
  • 31. Person 2 Person 1 Triangle relationship Person 3 Antifraud ¨C cycle detection
  • 32. Applicant Applicant 2 Parent of Parent of Applicant 1 Spouse Inconsistent relations Antifraud ¨C inconsistent relationship
  • 33. Antifraud ¨C suspicious group Person 2 Person 1 Person 3 Share a lot of common attributes
  • 34. Knowledge Graph Visualization ? Visualize entities and relationships ? Design anti-fraud rules via observational study Antifraud ¨C design by observation
  • 35. Rapid change of relationship structure within short time period Antifraud ¨C evolution of graph structure
  • 36. LR Decision Tree Random Forest SVM ANN Models Prediction Extracted Features from Raw Data Results from anti-fraud rules User direct attributes Variables DNN Score is used to directly reject or accept the loan Antifraud ¨C fraud score score
  • 37. Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
  • 38. The borrowers disappear, all the contact information they explicitly provided become invalid. How to reach them? Lost contact recovery ¨C what is it Implicitly infer potential contact information
  • 39. Phone Applicant Personal phone Phone Phone Phone Phone Phone Rank the phone numbers, and predict relationship Building phone network ¨C 1st order extension
  • 40. Building phone network ¨C 2nd order extension Phone Applicant Personal phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Rank the phone numbers, and predict relationship
  • 41. 3rd order .. Phone Applicant Personal phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Phone Phone
  • 42. Simple Ranking Criteria ? The total length of time ? The frequency of calls Advanced Approach ? Learning the ranking score using machine learning approach Building phone network ¨C Rank
  • 43. ? Total # of times of calling ? Total length of time of calling ? Total # of times of being called ? Total # of times of calling ? Average time per call ? Maximum length of time ? # of times of calling between 0-4am ? # of times of calling between 4-8am ? ¡­¡­ Building phone network ¨C Predict the relation LR Decision Tree Random Forest SVM ANN Models Prediction of relation ~100 Features DNN Relation With very limited training data, our model provides ~30% accuracy
  • 44. Person Applicant Personal phone Person Other applicant knows£¿ Other approach ¨C Link prediction (on-going work) Link Prediction
  • 45. Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
  • 46. Detect Bad People via Search From the search results, we label each entities in the knowledge graph i.e. black, green etc.
  • 47. ? ? 360.com ? other public websites Search for basic information¡­. ? Phone number ? Email ? QQ ? Other IDs Search Fields Search Engines & Public Site
  • 48. Search for phone ²Ô³Ü³¾²ú±ð°ù¡­
  • 50. ? Clustering analysis ? Precision marketing ? ¡­¡­ Other Applications we are working on
  • 51. Outline ? Why need Big data and AI ? Intro to FC Engine and Knowledge Graph ? Case 1: Anti-Fraud ? Case 2: Lost Contact Recovery ? Case 3: Detect Bad People via Search ? More use cases ? Challenges
  • 52. Challenges : Unstructured Data Unstructured Data Images Text AudioVideo Machine Learning Natural Language Processing Data Mining
  • 53. Challenges : Name Disambiguation Applicant Other applicant Puhui Finance Ltd. Puhui Finance Same company, can we merge? It is a very important problem to deal with!
  • 54. Challenges : Reasoning However, It is still an open problem ? Logic-based approach ? Probabilistic approach (i.e. distributed representation) ? Hybrid approach Link Prediction
  • 55. Challenges : Insufficient Samples Big data, but small samples
  • 56. ? Senior/Lead Machine Learning/NLP Engineers ? Senior/Lead Data Engineer/Scientist ? Senior/Lead Architect ? Senior/Lead Software Engineer liwenzhe@puhuifinance.com zhaopin@puhuifinance.com We are hiring! (in Beijing) Open positions, but not limited to¡­. Contact Company Website www.puhuifinance.com
  • 58. [1] http://www.datapop.com/ [2] http://db-engines.com/en/blog_post//43 [3] http://db-engines.com/en/ranking [4] Bordes, Antoine, et al. "Translating Embeddings for Modeling Multi- relational Data." Advances in Neural Information Processing Systems(2013):2787-2795. [5] Nickel, Maximilian, V. Tresp, and H. P. Kriegel. "A Three-Way Model for Collective Learning on Multi-Relational Data.." International Conference on Machine Learning 2011:809-816. References
  • 59. [6] Richard Socher, Danqi Chen, Christopher D. Manning, Andrew Ng. Reasoning With Neural Tensor Networks for Knowledge Base Completion. Advances in Neural Information Processing Systems(2013) [7] Wang, Quan, Wang, Bin, and Guo, Li. "Knowledge base completion using embeddings and rules." Proceedings of the 24th International Conference on Artificial Intelligence AAAI Press, 2015. [8] T Rockt?schel£¬S Singh£¬S Riedel. Injecting Logical Background Knowledge into Embeddings for Relation Extraction http://talks.cam.ac.uk/talk/index/58360 References