Snapcase

This repository contains the source code for our demo submission "Snapcase - Regain Control over Your Predictions with Low-Latency Machine Unlearning" to VLDB'24.

Running the demonstration yourself

1. Make sure you have a recent version of Rust installed.
2. Clone this repository locally via git clone https://github.com/amsterdata/snapcase-demo and change into the snapcase-demo folder
3. Download the prebuilt top-k index and the purchase database from Google Drive. The __instacart-index.bin file must be placed directly in the snapcase-demo folder, and the *.parquet files must be placed in the datasets/instacart/ subfolder.
4. Start the demo with the following command: cargo run --release --bin service
5. You should see some console output from DuckDB and Differential Dataflow, after which the demo will be served at http://localhost:8080 , which you can open in a browser

Inspecting the source code

We provide a few pointers for researchers interested in understanding the source code:

• Our implementation of the TIFU model for next-basket recommendation in Differential Dataflow is available in the src/tifuknn folder.
• Our data structure and algorithm for efficiently maintaining a top-k aggregation over the result of a sparse matrix multiplication is available in the src/caboose folder.
• The code to coordinate the incremental view maintenance of the recommendation model is available in the src/materialised subfolder.

Attribution

This demonstration uses icons from the Flaticon platform.