gdemelo / cafe

Cafe: Coarse-to-Fine Neural Symbolic Reasoning for Explainable Recommendation

[CIKM 2020] Cafe: Coarse-to-Fine Neural Symbolic Reasoning for Explainable Recommendation

This repository contains the source code of the CIKM 2020 paper "Cafe: Coarse-to-Fine Neural Symbolic Reasoning for Explainable Recommendation" [2].

Data

The Amazon Beauty dataset is available in the "data/" directory and the split is consistent with [1]. Data files include:

• kg_entities.txt.gz: KG entities where each row is in the format of <entity_global_id>\t<entity_type>\t<entity_local_id>\t<entity_value>. <entity_local_id> numbers the entities of the same type while <entity_global_id> assigns a unique id to each entity in KG.
• kg_relations.txt.gz: KG relations including reverse relation, where each row is in the format of <relation_global_id>\t<relation_name>.
• kg_triples.txt.gz: all KG triples including reverse edges, where each row is in the format of <head_entity_global_id>\t<relation_global_id>\t<tail_entity_global_id>.
• kg_rules.txt.gz: KG rules of length at most 3. Each row is a sequence of <relation_global_id>'s.
• kg_embedding.ckpt: The pretrained KG embeddings based on the work [1].
• train_labels.txt.gz: user-item training pairs, where in each row, the first number is a <user_id> and the rest numbers are <item_id>'s. These pairs are already contained in the KG triples, e.g., if a user-item pair (0, 100) is in the training labels, there exists a corresponding triple (user_0, purchase, item_100) in KG.
• test_labels.txt.gz: user-item test pairs. They are the groundtruth to be predicted by models, and are NOT part of the KG.

How to Run

1. Data preprocessing.

python preprocess.py --dataset <dataset_name>

2. Train neural-symbolic model.

python train_neural_symbol.py --dataset <dataset_name> --name <model_name>

The model checkpoint can be located in the directory "tmp/<dataset_name>/<model_name>/symbolic_model_epoch*.ckpt".

3. Do path inference by the trained neural-symbolic model.

python execute_neural_symbol.py --dataset <dataset_name> --name <model_name> --do_infer true

4. Execute neural program (tree layout given by user profile) for profile-guided path reasoning.

python execute_neural_symbol.py --dataset <dataset_name> --name <model_name> --do_execute true

References

[1] Qingyao Ai, et al. "Learning Heterogeneous Knowledge Base Embeddings for Explainable Recommendation." In Algorithms. 2018.
[2] Yikun Xian, et al. "Cafe: Coarse-to-Fine Neural Symbolic Reasoning for Explainable Recommendation." In CIKM. 2020.