We look forward to presenting Transform 2022 in person again on July 19 and virtually from July 20 to 28. Join us for insightful conversations and exciting networking opportunities. Register today!
Graphs are simply a universal way of representing relationships between entities – starting with immediate connections, then “jumping” to connections of connections of connections. The further out they go, the wider the tree gets.
Graph Neural Networks (GNNs) are often used to understand this. These deep learning models specialize in understanding graphs.
However, when it comes to today’s social networks, GNNs are far from ideal. When determining connections between friends, acquaintances, and work colleagues, they often fail to calculate the nuances and complex degrees of relationships. This makes it difficult for platforms like LinkedIn, Twitter, Facebook and Instagram to provide accurate recommendations – a core part of their mission.
LinkedIn takes a PASS at GNN
To overcome such inherent challenges with GNNs and improve its recommendation skills, LinkedIn has developed a process it calls Performance-Adaptive Sampling Strategy (PASS). It uses AI to select the most relevant neighbors in charts, improving prediction accuracy.
After applying the new GNN model to its own recommendation engines, the professional networking platform has just released PASS to the open source community.
“We want to compare our methods to other researchers’ datasets,” said Jaewon Yang, senior software engineer at LinkedIn who led PASS. “We hope that they can expand our networks.”
At a high level, LinkedIn uses GNNs to understand the relationships between individual members, groups, skills and interests – at the primary, secondary, tertiary level and beyond – to inform recommendations.
PASS’s unique AI neighbor selection model refines this by using its attributes to decide whether to select a particular neighbor. It can also help detect if any of those neighbors is actually a bot or a fake account by determining the authenticity of their connections. This adaptive model learns how to choose neighbors that increase its accuracy.
“Sometimes people overlook other titles that can be very relevant to a job posting or other recommendation,” Yang said. “We want to understand exactly who this user is following, we want to understand what other users are following users.”
Traditional GNNs are difficult to scale to social networks because they represent so many potential relationships, not all of which are relevant to specific tasks, Yang said. For example, a member’s connections may be personal friends who work in different fields, reducing recommendation accuracy.
Meanwhile, an influencer or prominent public figure could have connections in the hundreds of millions — flouting the sociological “Dunbar number” theory that each person can only have a certain number of friends, Yang pointed out — and it is impossible to calculate them everything.
“These represent an explosive number of data points to consider,” he said. “We can’t take everyone into account, we have to take a few samples.”
Some existing methods have attempted to overcome scaling challenges by sampling a fixed number of “neighbors”, thereby reducing the inputs to the GNN. But such samplers aren’t fully representative, Yang said, and don’t take into account which neighbors might prove most relevant.
IBM and Yale platforms promote GNNs
Other organizations are launching similar platforms trying to strengthen existing GNNs. For example, Yale University and IBM recently proposed a concept they call Kernel Graph Neural Networks (KerGNNs) that integrate graph kernels into GNN messaging. This is the process by which vector messages are exchanged between nodes in a graph and then updated. According to Yale researchers, using this KerGNN method has resulted in improved model interpretability compared to traditional GNNs.
Similarly, Google has released TensorFlow Graph Neural Networks, a library intended to make it easier to work with graphically structured data in its TensorFlow machine learning (ML) framework. Twitter, Pinterest, Airbnb and others are also researching and releasing tools to remove GNN restrictions.
PASS has been shown to achieve higher prediction accuracy despite using fewer neighbors than other GNN models. In experiments using seven public benchmark graphs and two LinkedIn graphs, PASS outperformed GNN methods by up to 10.4%. It also showed up to three times more accuracy compared to base methods by adding so-called “noisy edges”.
With the open-sourcing PASS, the hope is that other researchers will find new ways to apply the platform, Yang said, making it more expressive, flexible and easy to model and break its limitations to continuously expand its use for a variety of applications extend .
“This technology is evolving very quickly,” said Romer Rosales, Senior Director for AI at LinkedIn. “We’re just scratching the surface of all the uses this can have. There is a lot of room for us to grow and for the wider community to grow in this area.”
LinkedIn researchers will continue to refine PASS to handle larger and larger datasets without losing expressiveness, he said. The goal is to eventually automate certain processes that still require human delivery—such as B. specifying parameters of how hop samples are taken and whether the system should identify two hops, three hops or further down the chain.
“This is fertile ground to try these new ideas,” Rosales said. “We hope other communities will join us as well, and we will be joining other communities to try and share these experiences.”
PASS is one of several that LinkedIn has released as open source this year, he pointed out. Another is FastTreeSHAP, a package for the Python programming language. This helps interpret the results of algorithms more efficiently to improve transparency in AI, including explainable AI, to build trust and improve decision-making – such as: It also helps modelers with debugging and general improvements.
Another project is Feathr, a feature store that simplifies the management of ML features at scale and improves developer productivity. Dozens of applications use the Feature Store to define features, calculate them for training, deploy them to production, and share them with teams. Feathr users have reported that the time required to add new functions to model training workflows is significantly reduced and runtime performance is improved compared to previous application-specific function pipeline tools.
“PASS is one example in a long line of AI projects that we have brought to the community,” said Rosales, “in an effort to share our experiences and help create more scalable, expressive, and accountable AI algorithms and to develop tools.”
VentureBeat’s mission is intended to be a digital marketplace for technical decision makers to acquire knowledge about transformative enterprise technology and to conduct transactions. Learn more about membership.
https://venturebeat.com/2022/04/18/linkedin-creates-pass-to-tailor-graph-neural-networks-for-social-media/ LinkedIn creates PASS to customize neural graph networks for social media