brycewang-stanford ICDM Writing Style
Revise a manuscript into the register ICDM reviewers reward: a data-mining mechanism made legible to a data-centric IEEE audience, with claims a reviewer can check and a scale story that is measured rather than asserted. Two hard constraints shape every choice — the Research Track is triple-blind, so the prose cannot lean on author identity, and the whole paper (references and appendix included) must fit 10 IEEE two-column pages.
Open with the data regime, not the model family
ICDM work starts from the shape of the data: streams seen once, power-law-sparse graphs, implicit rather than explicit feedback, evolving edges, high-dimensional transactions. The strongest openings name that regime and make the mechanism its consequence.
Weak (model-first):
"We propose a novel graph neural network with a new attention module
for anomaly detection, achieving state-of-the-art results."
ICDM-shaped (regime-first):
"Interaction graphs arrive as an unbounded edge stream seen once and
cannot be stored. We give a one-pass detector whose memory is
sub-linear in the number of edges, with a stated error bound."
The second version names the regime, implies why existing methods fail, and scopes the mechanism before a reviewer asks.
Name the mechanism, and give it a reason
- Replace architecture nouns ("a transformer with X") with a named mining mechanism and the one property that makes it work (isolation, sparsity, restart probability, sketching).
- Attach every strong adjective to the mechanism: not "our method is scalable" but "the sketch updates in amortized constant time per edge."
- State the objective or the guarantee plainly enough that a reviewer sees why it should generalize before seeing the tables.
Measure scale; do not adjective it
ICDM referees read "efficient" and "scalable" as claims to be tested. Convert them:
| Adjective (avoid) | Measured claim (prefer) |
|---|---|
| "scalable" | "constant per-edge latency as the stream grows 100x (Fig. 2)" |
| "efficient" | "linear in |
| "outperforms baselines" | "matches the batch detector's top-k ranking at 1/10 the memory" |
| "robust" | "ranking quality within one std over 20 seeds across noise levels" |
Write discovery-validity sentences
The distinctive ICDM instinct is "is the finding real, or an artifact of the evaluation?" Pre-empt it: say why a flagged pattern, cluster, or anomaly means what you claim — tie it to a bound, a controlled injection with known truth, or a baseline that shares the setup. A result that could be an evaluation artifact reads as unfinished here.
Keep self-reference triple-blind-safe
- Cite your own prior work in the third person: "Prior work [7] introduced X" — never "in our earlier paper [7] we introduced X."
- Remove "our deployed system," internal names, and acknowledgements from the body; they
are anonymity violations, not style slips (see
icdm-submission).
Compress into 10 all-inclusive pages
- Every paragraph competes with the appendix and references for the same 10 pages. Cut signposting roadmaps and background everyone knows.
- Push protocol detail to an in-cap appendix or the cited anonymized repository, and keep the body's argument self-contained — because with no rebuttal, the body is your only chance to be understood.
Vignette: rescuing a buried contribution
A draft opens with two paragraphs on the importance of data mining and a model diagram. The revision deletes both, opens on "transactions arrive faster than they can be labeled," names the sparsity mechanism in sentence two, converts three "efficient/scalable" claims into measured latencies, and adds one discovery-validity sentence tying flagged transactions to a known-injection experiment. Same result, half a page shorter, and now legible to an ICDM reviewer on the first read.
Output format
[Register] ICDM-ready / drifting to ML-flagship / drifting to journal
[First-page arc] regime / task / mechanism / evidence / validity - present or missing
[Scale language] measured / adjectival
[Anonymity] third-person self-cites: yes / leaks found
[Cut list] <paragraphs to remove to fit the 10-page all-inclusive cap>