01 introduction to graph data science
1 like289 views
This document discusses the value of graph data science and provides examples of its applications. It notes that graph data science can improve predictions by leveraging network structure and connections in data. Specific applications mentioned where graphs add value include fraud detection, recommendations, customer segmentation, and predictive maintenance. It also profiles three large companies (a bank, social network, and logistics firm) that successfully apply graph techniques at scale.
01 introduction to graph data science
- 2. ● Why should you care about Graph Data Science? ● What does better predictions with graphs mean? ● Where can/is it applied? ● Who would have made the start of a nice fourth line, but all of these two days are aimed at you, so that's already answered!
- 4. Maximus, let us whisper now, together, you and I ... How exactly did you get to your current job position?
- 5. ● Has been in IT for >25 years ● Has experience in retail, banking, government, … ● Has a flawless track record, known for delivering on time and under budget against all odds ● Has experience as an expat and communicates well in three languages ● Is up to date in his field and regularly reboots his knowledge ● Is loyal ● …
- 6. ● He pops up on the radar of every headhunter on the planet for almost any job in IT ● Is labeled as a top candidate by almost any job matching algorithm for anything he would possibly want to do ● Frequently gets rejected on job interviews with you are the ideal candidate but my gut feeling tells me ... ● Got his current position by reaching out to an old contact on LinkedIn with hardly any jobinterview following that, effectively landing him the type of job he wanted for over a decade ...
- 7. It's not about what you know, it's about who you know and where they are in the organization you are looking at … it's about the connections.
- 8. Consider the following graph ...
- 9. Which one of the nodes is most likely to get the highest pay raise? ❏ The node with the highest degree (that's what Google does with pageranking). Using the graph to tell you! ❏ You can't tell based on this information, you'd need information (features) about performance, experience, age, … ❏ The nodes that connect the two parts of the graph.
- 10. ● It's the bridging nodes. ● In some organizations known as organizational misfits because they rarely fit well in their own/current department. ● They go from ladder to ladder, sometimes even going down a rung or two in the process, in order to get ahead. And it pays ... By the way, it is the graph that tells you, but here you needed centrality rather than pageranking … you don't need to tell Google though, they already know ...
- 11. ● Both examples are counter intuitive within the context of regular data science where you focus mostly one what you know about a given thing. ● It is only when you take into account the connections and the positions in the network, that it becomes obvious. ● And that's why Graph Data Science is so powerful.
- 12. Nihil nove sub sole … our parents and grandparents lamented the fact that some got carried into a job and we all have stories of job hoppers and mavericks that passed us by in our careers … Those may be depressive thoughts, but a graph at least enables us to show that there was indeed truth to those stories ...
- 15. You can do better predictions with graphs … and Neo4j is not the only company saying this ● As sort of shown by the fact that this training filled up almost as soon as we announced it (bit of a circular argument there) ● As shown by the boom of AI Research Papers featuring graph. See next slide, data from https://dimensions.ai searched for "graph neural network" OR "graph convolutional" OR "graph embedding" OR "graph learning" OR "graph attention" OR "graph kernel" OR "graph completion"
- 18. That's quite a good marketing line … but what does it mean? Well, here are a few statements that almost any data scientist agrees with in the context of almost any given project/problem ● Give me more data and more features and the predictions will be better. ● Getting more data is simply not practical for the problem I'm working on, I've done the best I could.
- 19. ● Current data science models largely ignore network structure even when it is available ● Graphs add highly predictive features to machine learning models, increasing accuracy Machine Learning Pipeline
- 20. This is the kicker. The network structure and connections are often there in the data already, but ● not used or heavily underused ● not stored in way that makes using them easy Graphs are the natural way to go to leverage this additional source of information within your existing data. Different from what you are used to, not difficult!
- 22. A few definitions so we are clear on what is meant in the next couple of slides … Vertical (market) - a specific business niche (e.g. banking) Horizontal (application) - concept that applies to many (sometimes all) verticals (e.g fraud detection) I know we're all supposed to know stuff like that, but I'll admit I didn't know so maybe you didn't either. Feel free to quietly thank me later, you're most welcome!
- 23. What are the Graph Data Science sweet spots? Fraud Detection Disambiguation & Segmentation Personalized Recommendations Churn Prediction Search & Master Data Mgmt. Predictive Maintenance Cybersecurity
- 24. ● Synthetic identity fraud ● Fraud rings ● Money laundering ● Recommendations ● Customer segmentation ● Churn prediction
- 25. ● Drug repurposing ● Patient journey ● Contact tracing ● Regulatory compliance
- 26. ● Logistics & Routing ● Supply chain ● Recommendations ● Customer segmentation
- 27. Top 10 Bank
- 28. Why did I pick those three (yes yes, there are more)? ● They cover the three verticals I explored a minute ago. ● The stories are publicly verifiable/documented. ● ... ● All three run Graph Data Science at massive scale. That is a key element.