Scott Campbell
arXiv recommender

Paper recommender, OpenAlex computer science corpus

Find related arXiv CS papers, ranked by four recommenders

Pick an arXiv computer science paper from 2019 on. A tuned hybrid, a sentence transformer, TF-IDF and a citation graph model each return their closest papers, side by side, against a popularity baseline.

The question
Given one paper, which method best finds the papers it actually cites?
The data
OpenAlex records for computer science papers on arXiv since 2019, with their citation graph.
The read
The tuned hybrid ranks first on MAP@10. The MiniLM sentence transformer lost to plain TF-IDF. Read the case study.
Best MAP@10 · hybrid
0.199
20x the popularity baseline
Hit-rate@10 · hybrid
0.43
finds a citation 14x as often
p95 latency · hybrid
59.4ms
top-10, on the evaluation set
Evaluation set
28,436
CS arXiv papers since 2019
  • Popularity(baseline)
  • TF-IDF
  • Neural
  • Citation ALS
  • Hybrid
Step 1

Choose a seed paper

Type a title or author. The FastAPI service does the lookup.

Step 2

Recommendations, side by side

pick a seed paper above

Hybrid

-

TF-IDF + neural + ALS + popularity, weights tuned on held-out NDCG

Pick a seed to see results

Neural

-

MiniLM sentence-transformer over title + abstract, cosine

Pick a seed to see results

TF-IDF

-

Title + abstract + authors + topic, sparse cosine

Pick a seed to see results

Citation ALS

-

Implicit ALS over citing → cited edges, 96 latent factors

Pick a seed to see results

Offline evaluation

How four recommenders and a popularity baseline score

Held-out citations, 2,000 seeds, top-10, bootstrap 95% CIs. Hybrid is the deployed model.

AlgorithmMAP@10NDCG@10Recall@10Hit-rateCoveragep95 ms
Popularity0.00980.01470.02930.0310.0000.2
TF-IDF0.14910.18720.30700.3120.38653.0
Neural0.12140.15600.26590.2710.3664.6
Citation ALS0.05130.06940.12750.1330.1810.7
Hybriddeployed0.19930.25410.42580.4340.27259.4
MAP@10
NDCG@10
Recall@10

Data: OpenAlex computer-science works on arXiv, 2019 onward, citation graph included. Models: popularity, TF-IDF, MiniLM sentence-transformer, implicit ALS over citations, and a held-out-tuned hybrid. Serving: FastAPI + FAISS. Frontend: Next.js on Cloudflare Pages. Source on GitHub.