M4 · Lesson 2 — Reproducing Works

Spotting
Missing Details

Every paper omits something important.
Learning to find what's not written is a core research skill.

01
M4 · L2 — Why Papers Omit

It's not always dishonesty

Why details go missing

  • Page limits — 8-page venue forces brutal cuts
  • Assumed knowledge — authors assume readers know standard tricks
  • Competitive advantage — some details are deliberately vague
  • Tuned but not justified — hyperparameters found by grid search, not theory
  • Bugs or mistakes — occasionally, the omission hides an inconsistency
"The paper says 'we train with Adam'. What learning rate? What decay schedule? What batch size? What weight initialisation?"

These details can change results by 10–20%. They matter enormously for reproduction.

02
M4 · L2 — Six Categories

What's usually missing

Six categories of
omitted details

Category 01

Initialisation

How are embeddings initialised? Random normal? Xavier? Zero?

→ Ask: what is E^(0) at t=0?
Category 02

Negative Sampling

How is j sampled? Uniformly? Popularity-weighted? How many per positive?

→ Ask: how is I⁻_u constructed?
Category 03

Learning Rate Schedule

Is lr constant? Does it decay? When does early stopping trigger?

→ Ask: how does α change over time?
Category 04

Embedding Dimensions

What is d? Is it the same for users and items? Is it tuned?

→ Ask: what is the shape of P and Q?
Category 05

Data Preprocessing

How are ratings normalised? What threshold defines an implicit positive?

→ Ask: what does r_{ui} = 1 mean exactly?
Category 06

Hyperparameter Tuning

Which values were grid-searched? On which split — validation or test?

→ Ask: were baselines tuned equally?
03
M4 · L2 — Reading Between the Lines

The "implementation details" section

What the paper says vs
what you need to know

K-RagRec — Section 4.1.4 Parameter Settings
"We use the 3 layers Graph Transformer as the GNN_Indexing and GNN_Encoding for MovieLens-1M and 4 layers for MovieLens-20M and Amazon Book. The layer dimension is set to 1024, and the head number is set to 4. The popularity selective retrieval policy threshold p is set to 50%."
❌ Missing: initialisation of E^(0) · negative sampling strategy · learning rate schedule · how KG triples are filtered to match each dataset

Rule of thumb: If you can't reproduce the exact experiment from the paper's "implementation details" section alone — details are missing. Check the appendix, then the code repository if available.

04
M4 · L2 — K-RagRec Applied

Applied to our running example

5 missing details in
K-RagRec

GNN initialisation

How are node embeddings z_n initialised before GNN_Indexing runs? Random? Pre-trained PLM only?

Negative sampling for cross-entropy loss

Section 3.8 trains with cross-entropy on (Y, A). How are negative candidate items selected during training?

KG filtering process

Freebase has millions of triples. The paper says "filter out triples related to the three datasets" — what is the exact filtering criterion?

Re-ranking similarity metric

Eq. 6 uses sim(p, z_g) for re-ranking. Is this cosine similarity or dot product? This changes results.

⚠️

p=50% threshold justification

Stated in Table 4 but Figure 4 shows performance is sensitive to p. Why 50% specifically? Grid searched on validation set?

05
M4 · L2 — What To Do

Practical recovery strategies

When details are missing,
go in this order

1

Check the appendix — most papers move implementation details here under page pressure

2

Check the GitHub repo — if code is released, it is the ground truth implementation

3

Check cited papers — if they use another paper's method, the details may be in the original paper

4

Use field conventions — e.g., Xavier initialisation and Adam with lr=0.001 are defaults in most RS work

Email the authors — a polite, specific question usually gets a response. Attach your pseudocode.

Document your assumptions — if you fill in a gap yourself, write it down clearly. This is important for reproducibility of your own work.

06
M4 · L2 — Key Takeaways

What to remember

01

Six categories to check

Initialisation · Negative sampling · LR schedule · Embedding dims · Preprocessing · Hyperparameter tuning

02

Code > paper

When a repo exists, it is the true implementation. The paper is a description — often incomplete.

03

Document your assumptions

When you fill in a gap yourself, write it down. Your reproduction is only trustworthy if you know exactly what you assumed.

Next: M4 · L3 — Building an Implementation Plan

07
← → arrow keys to navigate