# KG_Rec:

- New Alignment Method:
- Alignment By:
- (BM25) + BERT (Anserini)
- IMDB(ml1m)(query) <-> dbpedia(document) (predict.sh)
- ranking method 1.5h
- top-1
- top-10
- top-20
- top-50
- top-100
- fine-tuned BERT alignment
- IMDB -> dbpedia
- how many 1?
- standard of alignment 1.5h
- finished
- find score of ranking 4h
- range(-1 ~ 1)
- < -0.9 rate
- === 9/15 ====
- BM25 Baseline
- recall with top alignment (Jacky)
- Recommendation Result (Jacky + Ivy)
- Random
- Fully Random
- random * 1
- random * 10
- random * 20
- random * 50
- random * 100
- Topic Alignment + IR
- focus on unlinked
- BM25+BERT
- BM25
- Topic Alignment + Random
- focus on unlinked
- Random
- seed alignment fine-tuned BERT (Jeanie)
- manipulate threshold
- Case Study
- Number of Nodes & Edges
- Merge all aligned nodes / different nodes

- Recommendation on Ripple Net (Jeanie)
- BM25 Baseline

- SoTA of KG + Recommendation (Jeanie)
- Leaderboard: https://paperswithcode.com/task/recommendation-systems

- Graph Construction: (Ivy 8/29)
- 1. connect only Items & Entities
- 2. Connect Entities with Users
- Connect all Entity neighbor to the user
- Connect some Entity neighbor to the user
- 0.3
- 0.5
- 0.7
- Recommendation By:
- KGAT (freebase -> dbpedia) 4h
- ripple net
- Evaluate By:
- Recall (recommendation)
- NDCG (recommendation)
- hit@10 (recommendation) 1h
- f1 (recommendation) 1h
- Topic-based Alignment (Alignment) 1h
- IMDB(ml1m) <-> dbpedia
- Code problem
- crawl
- crawl by anserini? [freebase_anserini](https://github.com/castorini/anserini/blob/master/docs/freebase.md)
- source code modification:
- Anserini: convert_to_jsonl -> add a line to admit null string.

- UI-BERT:
- Sample
- user review for item?
- loss & math
- baseline (BPR, ...)
- survey to avoid dulplicate idea
- how to predict

- Measure the importance of each item to user:

- 1. PageRank -> better version from pprgo (Jacky 8/29)

- 2. Sample a subgraph with similar users
- KL-Divergence / cross entropy (Jeanie 8/29)
- pick a user i
- user adj by average / by rating+softmax (every score) / by rating+softmax (only >= 3)
- KL by user i and others
- top-100 similarity
- Reformer(local nearest neighbor)
- Group Recommendation

- 3. Train on Embedding (Baseline: TPR)

# KG_Rec 8/27:
- summarzation:
- http://nlpprogress.com/english/summarization.html
- https://arxiv.org/pdf/1808.10792.pdf?fbclid=IwAR36a58Jis1df1zaj9eZGa6qI8gsFqCfqa7UJNw8cffR6IhSE3Ut_lxfIaM
- https://www.aclweb.org/anthology/P17-1099.pdf?fbclid=IwAR031urH_V-ZVpk3jYK1EiLrF5iWq0bb7eeIsjuU-vZrDMZDAM_Lvv-ZLOI