Learn Machine Learning with Graph in Hyperconnected Data (Part 2)

Hi everyone,

We are Bitnine Global, a dedicated research and development company in graph databases. Today, we are excited to dive deeper into the applications of graph-based machine learning, specifically focusing on e-commerce data.

This article will explore unsupervised learning-based 'Graph Clustering' and supervised learning-based 'Graph Node/Link Prediction' techniques, showcasing their significance in storing and utilizing results within a graph database. By understanding these powerful approaches, we can uncover valuable insights and optimize decision-making processes in the dynamic world of e-commerce.

Missed the first part? Read it here

Graph Projection and Clustering Analysis

Figure 1. Graph DB Modeling with Payment Sample Data Table

Once we understand the intricate product-purchase relationships within our e-commerce data, the next step is to cluster customers based on their shared properties. We employ the 'Graph Projection' technique, which constructs a unified network by projecting a graph along specific relationships. This helps visualize and analyze distinct customer clusters, uncovering valuable insights that can drive personalized marketing strategies and enhance customer experience.

Figure 2. Graph Database Modeling with Payment Sample Data Table

The relationship between Customer A and Customer B, connected through the purchase of the same product, creates a view that links the edges of the graph projection.

Querying Single Relationships

First, we can ask simple questions to find out what products customers buy and what services they use.

Essentially, we construct a network of customers who have purchased the same product. From this network, we use a graph clustering algorithm to group customers based on structural similarities. This helps identify distinctive characteristics within each cluster, leading to personalized marketing strategies.

Figure 3. Creating Homogeneous Network and Graph Cluster in Apache AGE

Querying Features of Clusters

Focusing on products reveals a single network with distinctive characteristics for each cluster. This network offers valuable insights and personalized recommendations to help customers find what they need. It also unveils new meta-information beyond traditional categorization, suggesting previously unseen product characteristics and facilitating deeper customer understanding.

Figure 4. Exploration of Cluster Graph Merging with Meta Info in Apache AGE

Node/Link Prediction

The results above demonstrate the meaningfulness of item purchases within the identified clusters. For example, in Cluster 2, the purchase of cleaning supplies is grouped with the keyword 'Moving and Cleaning,' enabling targeted product recommendations.

Similarly, in Cluster 4, furniture supplies are associated with 'Interior,' allowing for precise suggestions to enhance interior design needs.

Beyond traditional product categories, we can utilize Node/Link Prediction techniques to predict and establish connections for items not yet categorized. This approach helps offer tailored recommendations based on genuine customer needs.

Node prediction operates at the individual node level, predicting characteristics based on relationships with neighboring nodes. Link Prediction predicts the likelihood of a connection between two nodes without existing connections. This enhances the utility of network information.

Figure 5. Example of Node Prediction

(Source: CS224W Chapter 6 Message passing and Node Classification)

Figure 6. Example of Link Prediction in Apache AGE

(Source: Missing Link Prediction using Common Neighbour and Centrality based Parameterized Algorithm)

Conclusion

This article explored the concept of graph hyper-connectivity, connecting relational database table data to a graph structure. By harnessing hyperconnectivity, data can be managed flexibly, leading to new value creation through advanced analytics.

For efficient data modeling and management, an enterprise DBMS like Bitnine Global is ideal.

It combines relational and graph databases, facilitating hyper-connectivity in e-commerce and other industries. Our solution is built on PostgreSQL, offering flexibility and compatibility with SQL queries. This integration empowers businesses to gain comprehensive insights, understand data flows, and create value across diverse domains.

Want to leverage both relational and graph databases? Bitnine Global is your solution for enabling hyper-connectivity.

Unlock valuable insights, enhance data exploration, and drive informed decision-making now with us !

Learn more from our experts at marketing@bitnineglobal.com