icdm-writing-style

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 |E|, sub-linear memory in the number of edges" |

| "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>