Skip to content
Proyectos

shehzaadzd/MINERVA has 323 GitHub stars and 90 forks. GitHot measured 0 stars in 24 hours, 0 in 7 days, and +1 in 15 days. Last checked 2026-11-19 10:19 UTC. Last push 2026-10-27.

Cómo se miden estas cifras

Crecimiento de estrellas medido

in 24h
0
in 7 days
0
in 15 days
+1
in 1 month
Recopilando historial

Cambios netos en períodos que se solapan, no totales diarios. Los datos no disponibles se dejan en blanco. Las estrellas indican interés, no calidad.

Acerca de

Meandering In Networks of Entities to Reach Verisimilar Answers

Aprendizaje automático

README.md

Original

MINERVA

Meandering In Networks of Entities to Reach Verisimilar Answers

Code and models for the paper Go for a Walk and Arrive at the Answer - Reasoning over Paths in Knowledge Bases using Reinforcement Learning

Authors: Rajarshi Das, Shehzaad Dhuliawala, Manzil Zaheer, Luke Vilnis, Ishan Durugkar, Akshay Krishnamurthy, Alex Smola, Andrew McCallum

MINERVA is a RL agent which answers queries in a knowledge graph of entities and relations. Starting from an entity node, MINERVA learns to navigate the graph conditioned on the input query till it reaches the answer entity. For example, give the query, (Colin Kaepernick, PLAYERHOMESTADIUM, ?), MINERVA takes the path in the knowledge graph below as highlighted. Note: Only the solid edges are observed in the graph, the dashed edges are unobsrved. gif gif courtesy of Bhuvi Gupta

Requirements

To install the various python dependencies (including tensorflow)

pip install -r requirements.txt

Training

Training MINERVA is easy!. The hyperparam configs for each experiments are in the configs directory. To start a particular experiment, just do

sh run.sh configs/${dataset}.sh

where the ${dataset}.sh is the name of the config file. For example,

sh run.sh configs/countries_s3.sh

Testing

We are also releasing pre-trained models so that you can directly use MINERVA for query answering. They are located in the saved_models directory. To load the model, set the load_model to 1 in the config file (default value 0) and model_load_dir to point to the saved_model. For example in configs/countries_s2.sh, make

load_model=1
model_load_dir="saved_models/countries_s2/model.ckpt"

Output

The code outputs the evaluation of MINERVA on the datasets provided. The metrics used for evaluation are Hits@{1,3,5,10,20} and MRR (which in the case of Countries is AUC-PR). Along with this, the code also outputs the answers MINERVA reached in a file.

Code Structure

The structure of the code is as follows

Code
├── Model
│    ├── Trainer
│    ├── Agent
│    ├── Environment
│    └── Baseline
├── Data
│    ├── Grapher
│    ├── Batcher
│    └── Data Preprocessing scripts
│            ├── create_vocab
│            ├── create_graph
│            ├── Trainer
│            └── Baseline

Data Format

To run MINERVA on a custom graph based dataset, you would need the graph and the queries as triples in the form of (e1,r, e2). Where e1, and e2 are nodes connected by the edge r. The vocab can of the dataset can be created using the create_vocab.py file found in data/data preprocessing scripts. The vocab needs to be stores in the json format {'entity/relation': ID}. The following shows the directory structure of the Kinship dataset.

kinship
    ├── graph.txt
    ├── train.txt
    ├── dev.txt
    ├── test.txt
    └── Vocab
            ├── entity_vocab.json
            └── relation_vocab.json

Citation

If you use this code, please cite our paper

@inproceedings{minerva,
  title = {Go for a Walk and Arrive at the Answer: Reasoning Over Paths in Knowledge Bases using Reinforcement Learning},
  author = {Das, Rajarshi and Dhuliawala, Shehzaad and Zaheer, Manzil and Vilnis, Luke and Durugkar, Ishan and Krishnamurthy, Akshay and Smola, Alex and McCallum, Andrew},
  booktitle = {ICLR},
  year = 2018
}