{"path":"research/corpus-graph/finding-unwritten-connections-as-a-helper-read-it.md","content":"# Finding connections nobody wrote down: what older fields learned\n\n*A helper's report, kept word for word (2026-10-10; Sonnet, read-only web research, 38 tool uses). Asked\nfor before the corpus graph was built, on the founder's word: \"Research gotchas and insights a bit online\nand locally to keep in mind here first.\" The helper was told to read sources themselves where it could and\nto say how deeply it read each. Nothing here was checked by Claude. How it bears on the build is in\n[README.md](README.md).*\n\n---\n\nResearch on finding unwritten connections: what older fields learned, by topic\n\nRead depth is tagged for each source: FULL, PARTIAL, or SNIPPET (search excerpt only). I paraphrased throughout. I fetched nothing locally.\n\n**1. Literature-based discovery (LBD)**\n(a) Swanson's unconnected literatures share no articles, authors or citations. His fish oil/Raynaud and magnesium/migraine cases were later backed by clinical studies. Open discovery starts from A and ranks candidate Cs. Closed discovery starts from A and C and ranks the bridging Bs.\n- Candidates explode. One PubMed study counted about 35,000 hidden pairs per term unfiltered and about 2,200 after the best filter, with a time-sliced F-score near 0.003.\n- Evaluation means rediscovering a handful of Swanson cases, or time-slicing, where any later co-occurrence counts as a discovery. A 2024 paper calls standard practice methodologically poor. Its improved discovery set still reaches only about 40% precision.\n- In a 2025 test, 19–28% of sampled candidates were already known.\n- Similarity-ranked systems return incremental findings. Adoption stalled for two reasons. Experts had to wade through raw bridge-term lists of about 600 terms. Researchers already had too many hypotheses. Smalheiser's answer was the two-node (closed) search, where the human brings both A and C and the system ranks bridges.\n(b) Trap: a flood of candidates, trivial or already-known links, and scoring a method by whether it rediscovers known results.\n(c) Sources:\n- https://pmc.ncbi.nlm.nih.gov/articles/PMC5771422/ (PARTIAL: all sections except the tail)\n- https://arrowsmith.psych.uic.edu/arrowsmith_uic/tutorial/smalheiser_2012.pdf (SNIPPET)\n- https://exa.ai/library/publication/bxzckb4gtsl (Preiss & Stevenson, SNIPPET)\n- https://link.springer.com/article/10.1186/s12859-024-05881-9 (SNIPPET)\n- https://www.mdpi.com/2076-3417/15/16/8785 (SNIPPET)\n- http://abel.lis.illinois.edu/tutorial/scientist_1998.html (1998 magazine piece, SNIPPET)\n- Swanson 1990, \"The absence of co-citation as a clue to undiscovered causal connections\": title only, not read.\n(d) Your definition matches Swanson's. Propose candidates per new piece, closed-style, with a capped list. Never run all pairs. Keep rejections.\n\n**2. Bibliographic coupling and co-citation**\n(a) Rankings flip with the gold standard.\n- Boyack & Klavans 2010 (2.15M biomedical articles): coupling slightly beat co-citation, direct citation was worst, and a citation+text hybrid was best.\n- Klavans & Boyack 2017, using articles with 100 or more references as the standard: direct citation won.\n- Sjögårde & Ahlgren 2024 (3,000 systematic reviews, 5 seeds each): co-citation beat coupling and direct citation. A combination of all three beat co-citation alone, and adding text helped further.\n- Waltman et al., judging by text relatedness (BM25): coupling beat co-citation and direct citation; co-citation was worst.\n- Co-citation cannot see uncited or new items. Coupling is fixed when the author writes.\n(b) Trap: circular evaluation, where citation signals are judged by citation signals, plus the cold start for new items.\n(c) Sources:\n- https://onlinelibrary.wiley.com/doi/10.1002/asi.21419 (abstract, SNIPPET)\n- https://arxiv.org/abs/1511.05078 (abstract)\n- https://arxiv.org/pdf/2403.09295 (PARTIAL: intro, data and methods; results from excerpts)\n- https://arxiv.org/abs/1901.06815 (abstract and excerpts, SNIPPET)\n- https://kops.uni-konstanz.de/server/api/core/bitstreams/df0f0c23-c406-4088-92ed-61521a16a8f2/content (Gipp et al.: studies contradict because of datasets and gold standards, SNIPPET)\n(d) A shared source or reference is precise but only fires for ties someone already wrote. Co-citation-style signals need later notes. Text is needed for new notes.\n\n**3. Link prediction**\n(a)\n- Liben-Nowell & Kleinberg: common neighbours (CN) and Adamic-Adar beat random by roughly 17–55 times, but the random hit rate was only 0.15–0.48%. No method won everywhere.\n- Zhou et al. 2009: in one network, 99.59% of over 10 million pairs scored zero under CN. Of the rest, 91% scored 1, so there are many ties.\n- Ghasemian 2020: across 203 algorithms and 550 networks, none was best overall. Individual topological or model-based predictors averaged AUC 0.64; stacking reached 0.89. Social networks were easiest and technological ones hardest.\n- Kovács 2019: in protein networks, higher neighbour overlap meant lower link probability, and CN did worse than a degree-preserving random baseline.\n- Li et al. (NeurIPS 2023): random negatives are trivial (almost none share a neighbour). Heuristics were under-reported; on ogbl-citation2 they scored about 75% MRR against about 50% previously reported.\n- Gelato (KDD 2025): GNN results on balanced samples do not carry over to the full imbalanced setting, where simple topology heuristics often win. CN is weaker in very sparse graphs. Local heuristics did fine on the very dense ogbl-ddi graph. I found no evidence of heuristics collapsing on dense graphs.\n- Hub bias:\n  - A graph neural network (GCN) scores links roughly in proportion to the geometric mean of the endpoint degrees (Subramonian 2024).\n  - Precision, F1 and NDCG favour high-degree nodes; Recall does not (Wang & Derr 2022).\n  - Link-prediction GNN recommendations amplify within-cluster links over cross-cluster ones (NeurIPS 2023).\n  - Adamic-Adar and resource allocation downweight high-degree shared neighbours; Chakrabarti's COLT theory paper supports this.\n- Document graphs: on Wikipedia, text (LSA) and graph embeddings performed about equally and hybrids added nothing. West et al. got precision@k of 0.4–0.5, but manual review found most suggested links were truly missing, never written.\n(b) Trap: held-out edges count valid-but-unwritten links as errors, which is your exact class. There are also trivial random negatives, degree bias, and CN's reach of only two hops.\n(c) Sources:\n- https://arxiv.org/pdf/1505.04094 (Yang et al., PARTIAL: first sections)\n- https://people.csail.mit.edu/dln/papers/link/paper.pdf (SNIPPET)\n- https://arxiv.org/html/0901.0553 (SNIPPET)\n- https://pmc.ncbi.nlm.nih.gov/articles/PMC7519231/ (SNIPPET)\n- https://www.nature.com/articles/s41467-019-09177-y (SNIPPET)\n- https://arxiv.org/html/2306.10453v3 (SNIPPET)\n- https://arxiv.org/pdf/2412.00261 (SNIPPET)\n- https://arxiv.org/html/2309.17417v2 (SNIPPET)\n- https://tylersnetwork.github.io/papers/icdmw22-degreebiaslinkpred.pdf (SNIPPET)\n- https://ar5iv.labs.arxiv.org/html/2105.11734 (SNIPPET)\n- https://ashwinparanjape.github.io/assets/pdf/West-Paranjape-Leskovec_WWW-15.pdf (SNIPPET)\n(d) This is my inference, not a finding. For pairs with no link, no shared reference and no co-editing, neighbour-overlap scores are near zero, so they only re-rank pairs that are already close. Use text similarity to generate candidates and graph distance as the filter. Treat your found pairs as positives.\n\n**4. Serendipity in recommenders**\n(a)\n- Kotkov's survey: serendipity means relevant, novel and unexpected. Unexpected is measured as distance from the user's profile. There is no consensus on definition or metric.\n- Ge et al.: a purely content-based recommender cannot surprise.\n- Auralist user study: serendipity rose, some accuracy and per-item enjoyment fell, and overall satisfaction rose. Users tolerated some accuracy loss for serendipity. Most of the gain came at small mixing weights.\n- Spinneret (mind-mapping): non-obvious suggestions were accepted 45% of the time in an open task versus 30% in a constrained one.\n- Offline proxy metrics correlate poorly with user judgement. The best LLM judge reached Pearson about 0.22 (RecSys 2025).\n(b) Trap: surprise is relative to a particular person at a particular time, and offline ground truth is weak.\n(c) Sources:\n- https://denis.kotkov.me/publications/A%20survey%20of%20serendipity%20in%20recommender%20systems.pdf (PARTIAL: sections 2 and 5)\n- https://dl.acm.org/doi/10.1145/2926720 (abstract)\n- https://dmice.ohsu.edu/bedricks/courses/cs606-ir/papers/zhang_2012.pdf (PARTIAL)\n- https://web-ainf.aau.at/pub/jannach/files/Conference_RecSys2010.pdf (PARTIAL)\n- https://ar5iv.labs.arxiv.org/html/2001.02746 (PARTIAL)\n- https://dl.acm.org/doi/full/10.1145/3705328.3748167 (SNIPPET)\n- https://arxiv.org/html/2508.17571 (SNIPPET)\n(d) Define unexpected relative to the source notes of the piece. Record who judged the surprise and when. Keep rejected pairs as negatives.\n\n**5. Note-taking graphs**\n(a) Only a little is evidence.\n- Luhmann's card index (Schmidt 2018): about 90,000 cards; roughly 20,000 references on about 23,000 cards in the first collection and about 30,000 on about 67,000 in the second. That is about one link per card early and one per two later. His keyword index listed only 1–4 entry points per term, and weakly linked sections faded (his own phrase was black holes). He kept adding references whenever new cards were integrated.\n- Zhu et al. (UIST 2024, guided tours with 23 scholars): users build hub, index and incubator notes. Indices carry a description of what each link is.\n- Interview and autoethnography studies exist (Ferreira 2025; Hypertext 2026, one person). The autoethnography found that AI-proposed links needing human curation worked well for its builder.\n- Opinion, not evidence:\n  - The global graph view is mostly decoration, while the local graph, backlinks, and orphan checks help.\n  - Past a few hundred notes the global graph becomes a dense ball, and an edge does not say why.\n  - Unlinked mentions behave like a standing text query that depends on titles and aliases.\n  - Eighty backlinks is noisier, not better connected.\n  - A vendor doc says computed semantic links are temporary, and conclusions should be kept as authored links.\n(b) Trap: links made by habit and carrying no reason, orphans, and upkeep. I found no controlled study.\n(c) Sources:\n- https://sociologica.unibo.it/article/view/8350 (via web.archive.org/web/20200324112915/sociologica.unibo.it/article/view/8350/8270, PARTIAL)\n- https://joelchan.me/assets/pdf/Zhu%20et%20al_2024_Patterns%20of%20Hypertext-Augmented%20Sensemaking.pdf (PARTIAL)\n- https://arxiv.org/pdf/2509.20187 (SNIPPET)\n- https://dl.acm.org/doi/10.1145/3800935.3830837 (SNIPPET)\n- https://www.aloha.place/blog/obsidian-graph-view-and-alternatives (blog, SNIPPET)\n- https://danholloran.me/posts/obsidian-unlinked-mentions-the-half-of-the-backlinks-pane-you-ignore (blog, SNIPPET)\n- https://smartconnections.app/smart-graph/faq/ (vendor doc, SNIPPET)\n(d) Give each kept edge a why, a type, an asserter and a date. Bare lines carry none of these.\n\n**6. Agent memory with typed, dated edges**\n(a)\n- Zep/Graphiti: edges carry four timestamps (valid and invalid, created and expired). An LLM judges contradictions and always favours the newer information. Each edge links back to its source episodes, so provenance is the text, not the asserting agent. On LongMemEval (gpt-4o-mini, full-context 55.4% vs Zep 63.8%; gpt-4o 60.2% vs 71.2%), the knowledge-update category was mixed: 74.4% vs 76.9% with gpt-4o-mini, 83.3% vs 78.2% with gpt-4o.\n- Graphiti GitHub issues (user-reported, small samples):\n  - With a non-reasoning small judge, contradiction detection scored 7/15, or 14/15 with a reasoning-first schema.\n  - Invalidation can retire unrelated facts that share an entity.\n  - A negation (\"released from\") left the old assigned edge valid.\n  - Shared-endpoint edges cannot be told apart as replacement versus additive without knowing the relation's cardinality.\n  - A reported 56% rate of invented dates was retracted by its reporter.\n- A-Mem: untyped links picked by an LLM among embedding neighbours. Its memory-evolution step overwrites old notes and keeps no history. It was evaluated on LoCoMo only.\n- Mem0: an LLM picks add, update, delete or no-op. The base version deletes contradicted memories; the graph variant marks edges invalid instead and scored only about 2% higher overall.\n- MemoryAgentBench: every method failed multi-hop conflict resolution, with a ceiling of 28% in the ICLR 2026 version (6% in an earlier version).\n- Vendor benchmark claims are disputed: Zep's own LoCoMo score was revised from 84% to 75.14% by Zep, versus 58.44% in a rerun by Mem0's side.\n- A 2026 memory-poisoning study says current systems have no robust provenance tracking for stored entries.\n(b) Trap: last-writer-wins by an unaudited LLM judge, wrong in both directions, and confusing additive facts with replacements.\n(c) Sources:\n- https://arxiv.org/pdf/2501.13956 (FULL body, appendix cut off)\n- https://arxiv.org/pdf/2502.12110 (PARTIAL: through ablation)\n- https://arxiv.org/pdf/2504.19413 (PARTIAL: through baselines)\n- https://arxiv.org/abs/2507.05257 (SNIPPET)\n- https://github.com/getzep/graphiti/issues/1666 (SNIPPET)\n- https://github.com/getzep/graphiti/issues/1728 (SNIPPET)\n- https://github.com/getzep/graphiti/issues/1841 (SNIPPET)\n- https://github.com/getzep/graphiti/issues/1492 (SNIPPET)\n- https://github.com/getzep/zep-papers/issues/5 (SNIPPET)\n- https://arxiv.org/abs/2606.04329 (SNIPPET)\n- https://getzep-graphiti.mintlify.app/concepts/temporal-model (unofficial docs mirror, SNIPPET)\n(d) Your edges are assertions, not facts about the world. Store each assertion with its asserter, date and basis. Keep model proposals in a candidate state until accepted. Supersede only by explicit, attributed action. Do not auto-invalidate.\n\n**Not established**\n- Any controlled study of whether graph view, backlinks, unlinked mentions or embedding link suggestions change discovery in personal notes.\n- Link-prediction benchmarks on small, hand-curated note graphs of a few hundred nodes.\n- LBD applied within one author's notes, rather than across literatures by different authors.\n- Heuristics failing on very dense graphs. I found the opposite.\n- Staleness rates for human-asserted edges in memory systems. I found only single-user GitHub measurements.\n- A serendipity trade-off curve for document or knowledge items. The evidence is on music and movies.\n- Kessler 1963, Small 1973, the full text of Boyack & Klavans 2010 (paywalled), and the statistics in the Thilakaratne LBD review: not read.\n"}