Gowalla Network: Location and Social Intelligence

Inspiration

I am passionate about traveling, knowledge graphs and and AI. When I found the Gowalla dataset I had an idea to combine all my interests.

What it does

• Allows graph exploration
• Finds popular users/places
• Finds commonalities between users in terms of common friends or locations visited
• Finds other places of interest (e.g. universities/restaurants/etc.) in the same area
• Finds common clientele between places of interest
• Provides friend recommendations
• Provides places to visit recommendations in terms of areas/neighborhoods and points of interest

How we built it

• Loaded and enriched Gowalla data with Nominatim and Google Search
• Built networkx graph and persisted it into ArangoDB
• Implemented and executed Graph Neural Network over graph
  • generated embeddings for amenity node snippets using OpenAIEmbeddings
  • for all other nodes used a vector of zeros
  • ran Message Passing algorithm for 5 timesteps (5 iterations)
  • saved learned embeddings
• Implemented below tools
  • ArangoDB AQL tool
  • Networkx code tool
  • Amenities vector search using text query and optionally id filters
  • Similar node search using input node id, learned embeddings, and optionally include/exclude ids and node_type filter
• Instantiated react agent with above tools
• Created visualization agent which takes user query, agent answer and generates networkx and matplotlib code to visualize result

Challenges we ran into

• Getting the LLM to adhere to the structure of the graph when generating AQL or networkx code.
• Challenges with storing and processing learned embeddings - it was necessary to use sparse matrices to avoid out of memory error

Accomplishments that we're proud of

• Effective use of other APIs to augment data
• Implementation of the Graph Neural Networks from scratch
• Implementation of vector search and embedding search tools
• Prompt engineering to get the LLM to produce correct queries/execution code.

What we learned

• About NetworkX and cuGraph
• About ArangoDB, AQL and ArangoGraphQAChain
• How to implement message passing GNN
• How to describe a graph and provide relevant examples so that LLM produces correct queries/code

What's next for Gowalla Network: Location and Social Intelligence

• Recommendations for new users
  • Allow users to add themselves to graph, who they are friends with and the locations they have visited
  • Re run GNN on new user's subgraph
  • Provide insights and recommendations for new users
• Get user-location ratings to enhance GNN
• Fine tune LLM to produce better AQL and networkX code
• Fine tune LLM to produce more accurate queries/code/tool use for our specific use case/graph

Built With

• apis
• aql
• arangodb
• cugraph
• faiss
• gcn
• gnn
• langchain
• langgraph
• matplotlib
• networkx
• nominatim
• plotly
• rag
• scipy
• search