Data Scientists at CRED, Ravi Kumar and Samiran Roy explained the nature of the use of graph neural networks and how the emerging technology of CRED was based on the recently held Deep Learning DevCon 2021. The duo explained how neural graph networks should be modeled and which key factors separate graph data from the traditional tabular data in neural network models.
CRED is an online payment app linked to credit cards. The app was developed by Kunal Shah, the founder of FreeCharge. The CRED app aims to automate the use of credit cards. The app also offers many rewards for use in the form of CRED coins, which can later be redeemed for cash or various offers.
In the opening minutes, Ravi Kumar explained what Graph Analytics and Graph Neural Networks are and discussed the problems related to traditional neural networks currently in use in the market.
“Graphs are a general language for describing and analyzing entities with relationships and interactions. In a graph network, nodes are entities that define a user, dealer, or other similar element. Edges describe the relationships between two nodes, and the properties define the information associated with nodes or edges, ”he said.
Image source: DevCon 2021
“Data from the real world is dynamic and grows steadily over time. A graph database gives the processed data a deeper context and provides great value for the relationship between the entities. Tabular data becomes sparse as the volume of data increases, ”explains Kumar, explaining why graph databases are used dynamically.
There are several types of graph networks; some examples are wiki networks, flight networks, underground networks, and social networks. He also demonstrated how graphs represent data and the difference between two of the most popular representations, RDF (Resource Description Framework) and LPG (Labeled Property Graph). In RDF the corners and edges of the graph network have no internal structure, while in LPG the corners and edges present in the graph have an internal structure and properties. RDF does not support the same pair of nodes and relationships more than once, but LPG does.
Image source: DevCon 2021
Ravi later went into the basic problems of traditional neural networks, such as interactivity between data points, disadvantages in logically separating nodes, problems encountered when nodes go beyond 3, and more.
“In general, we can assume id as data points in machine learning, but we cannot say that in the case of graphical neural networks,” added Ravi to continue the discussion.
Samiran Roy later took over the presentation and continued speaking on the subject by explaining the functionality and framework of a graph-neural network. A neural graph network comprises three main blocks: the edge block, the node block and the global block. These blocks work together with aggregator functions towards the required target goal.
Samiran said, “From a data science perspective, we have the flexibility to embed functionality in any of the graph network components. We can embed functions at the edge, node or graph level. “
He also described how the points in a graph network influence each other and which methods can be used to solve problems in a graph network. In his lecture he explained the interactions between network blocks and their aggregators in graph networks and their different variants.
Variants for graphic neural networks include the following types:
- Full GN block
- Independent recurring block
- Message delivery neural network
- Non-local neural network
- Relationship network
- Deep set
Regarding the problems of neural graph networks in practice, Roy added: “The graphs that we encounter in the real world are heterogeneous graphs; There are several types of knots. The aggregation functions used are to be assigned the greatest importance when defining our target output. When working with large graphs, we do not have to calculate for all edges and nodes; Instead, we can convert nodes into subsamplings and use them to train our neural graph networks. “
Graph neural networks have several use cases, such as classifying users on social networks and predicting molecular properties, to name a few.
Current applications of graphic neural networks at the CRED include:
- Product targeting model
- Community discovery
- Graph completion
- Ranking of propensity to recommend
Use cases that CRED is investigating are embedding models for downstream use cases and creating user-to-user or user-to-merchant affinity models.
Graphic neural networks are emerging but are rapidly evolving as a field. Today, industry data requires a lot of research on graph networks as the current research is repeated on the same standard data sets. The flexibility in defining features and goals sets graphical neural networks apart from other typical neural network types.
Join our Discord server. Become part of a dedicated online community. Join here.
Subscribe to our newsletter
Get the latest updates and relevant offers by sharing your email.