Tempo: Collaborative Predictive Model Specification

Tempo is an interactive tool to help data scientists and domain experts define specifications for predictive models. It includes a simple, yet precise query language for aggregating temporal event data, as well as a built-in subgroup discovery interface powered by Divisi.

For more information and examples of how Tempo could be used in cross-functional teams, please see our paper at CHI'25 (link coming soon).

Setup Instructions

To set up, cd into the main repository directory, and run pip install -r requirements.txt. Then cd into the client directory and run npm install && npm run build.

Create a directory called datasets and place any datasets you'd like to load within that directory (see below for more details).

To run the tool, cd into the main directory and run:

python -m tempo_server.server <parent of datasets directory>

Then visit http://127.0.0.1:4999 in your browser to see the tool.

(Development) To automatically rebuild the client while changing the source code, cd into the client directory and run npm run autobuild. After making changes to the source code, you will need to reload the browser page to see the updated version.

Importing Datasets

The structure expected for the datasets directory is as follows:

<parent folder>
- datasets
  - <dataset name>
    - spec.json
    - <data files>

Below is an example of the dataset specification (spec.json):

{
  "data": {
    "sources": [
      // place data files in the dataset folder (.csv or .arrow) and provide their relative paths here
      { "type": "attributes", "path": "data/attributes.csv" },
      { "type": "events", "path": "data/events.csv" },
      { "type": "intervals", "path": "data/intervals.csv" }
    ]
  },
  "slices": {
    "sampler": {
      // subgroup discovery settings required by Divisi
      "min_items_fraction": 0.02,
      "samples_per_model": 100,
      "max_features": 2,
      "scoring_fraction": 0.5,
      "num_candidates": 20,
      "similarity_threshold": 0.5
    }
  },
  "models": {
    "default": {
      // optionally provide a default model spec
      "variables": {}
    }
  }
}

The datasets directory in this repo contains an example sourced from an e-commerce behavior dataset on Kaggle.

Citation

TBD