Reading roadmap

seed#moc#literature#reading-list

The tight path through everything collected here, sequenced to the 12-month plan. Read in service of doing (the next experiment), not completionism — but as a newcomer, read enough prior art to avoid reinventing solved wheels (Fast iteration beats parallelism). ✅ = open access.

Full lists: Foundations reading list · Clickstream foundational papers · Sequence modeling papers · Anomaly & drift papers · Uplift modeling papers.

▶ Start tomorrow (don't overthink it)

MacKay ch. 2 (entropy/KL) → then do the by-hand exercise in Cross-entropy and KL divergence.
Skim srivastava2000webusagemining (the WUM pipeline) — 30-min pass-1 only.
Read laub2015hawkes §1–2 — you'll use it in the generator.

Phase 0 · Foundations & the lab habit (Q1)

Goal: enough math to build the synthetic generator and measure recovery honestly.

Info theory & ML toolkit: mackay2003information (entropy/KL chapters) · hastie2009esl or murphy2022pml (skim).
The generator's math: rabiner1989hmm (HMM/Viterbi) · laub2015hawkes + rizoiu2017hawkes (Hawkes) · clauset2009powerlaw (heavy tails) · bacry2017tick (simulate it).
Tie-in: these back the by-hand exercises in 10_Foundations-MOC and the Track — Synthetic generator + 3-algorithm bake-off.

Phase 1 · Domain grounding (Q1→Q2)

Goal: don't reinvent sessionization / path modeling.

srivastava2000webusagemining (the field's vocabulary) → spiliopoulou2003sessionreconstruction (sessionization, the hard part) → montgomery2004pathanalysis (Markov path prediction).
Backs What is a clickstream · Sessionization · Funnels and journeys.

Phase 2 · First real track + skepticism reflex (Q2)

Goal: one shippable-or-killed result, with leakage/measurement caught. Read only the arm your track needs.

If prediction/discovery: rendle2010fpmc → hidasi2016gru4rec → kang2018sasrec (+ wang2019sequential to map the field). Prereq: vaswani2017attention.
If anomaly/drift: chandola2009anomaly → gama2014conceptdrift → bifet2007adwin / liu2012isolation.
Always: evaluation discipline — Evaluation theory, run the Leakage checklist and Offline→online checklist.

Phase 3 · The causal turn (Q3)

Goal: predict vs cause — uplift, not just propensity.

gutierrez2017uplift (survey) → kunzel2019metalearners (S/T/X-learners) → wager2018causal (causal forests) → anderl2016mapping (Markov attribution).
Backs Causal inference primer · Attribution / uplift · Predict vs cause checklist.

Phase 4 · Communicate & compound (Q4)

Re-read the surveys (wang2019sequential, gama2014conceptdrift) to place your results in the field.
One quote per source, paraphrase the rest — every paper gets a note from 80_Literature-Template.