What was mapped

What is ignored

author_profile controls the Minimal Mistakes sidebar, which DataLog does not have; DataLog shows the author card at the end of the post instead. seo_type: article is redundant, because DataLog already marks posts as articles for social cards and structured data.

Where the rules live

The mapping is a single build hook in _plugins/front_matter_compat.rb. It only fills DataLog fields that are absent, so a post that sets image or description explicitly keeps those values. The full field table and the URL-preservation settings are in the migration guide.

Download options

• retail-demand-benchmark.csv via the dataset page
• Parquet export via _datasets/retail-demand-benchmark.md
• Companion notebook: _notebooks/batch-anomaly-detection.ipynb

Schema

Usage guidelines

1
2
3
4
5
6
7
8
9
10
11
data_quality:
  freshness: P1D
  source_system: retail_warehouse
  owners:
    - name: Diogo Ribeiro
      email: diogo.debastos.ribeiro@gmail.com
    - name: Analytics Platform Team
      email: platform@example.com
  sla:
    anomalies: notify-analytics-oncall@example.com
    completeness: >= 0.98

Governance tip: Store ownership metadata in _data/catalog.yml so the theme can surface it across dataset and notebook pages.

Next steps

1. Publish baseline forecasts and accuracy benchmarks in the notebooks collection.
2. Embed interactive Plotly charts showing store-level variability.
3. Solicit community contributions via GitHub Discussions and credit accepted pull requests in the showcase gallery.

With schema documentation and ownership metadata in place, teams can trust the dataset before wiring it into production pipelines.

Project overview

• Client: Streaming media startup scaling personalization
• Role: Lead machine learning engineer
• Timeline: Q1 2024 (8 weeks)

Repository snapshot

Why it matters: Badges refresh automatically so stakeholders always see the latest adoption metrics.

Architecture

graph TD
  A[User events] --> B[Feature store]
  B --> C[Candidate generation]
  C --> D[Ranking model]
  D --> E[Realtime API]
  E --> F[Personalized UI]

Outcomes

Supporting assets

• _notebooks/recsys-evaluation.ipynb – Offline evaluation and fairness analysis
• _datasets/recsys-segment-metrics.csv – Segment-level KPI tracking
• _portfolio/market-segmentation-case-study.md – Extended client background

Wrap up by inviting readers to clone the repo, run the notebooks, and explore the dashboards embedded elsewhere on the site.

Hypotheses

Sample-size planning

1
2
3
4
5
6
7
8
9
10
11
12
13
from statsmodels.stats.power import NormalIndPower

effect = 0.04  # minimum detectable effect (absolute)
alpha = 0.05
power = 0.8
baseline = 0.32

analysis = NormalIndPower()
n_per_group = analysis.solve_power(effect_size=effect / (baseline * (1 - baseline)) ** 0.5,
                                  power=power,
                                  alpha=alpha,
                                  ratio=1.0)
print(round(n_per_group))

Reminder: Adjust for multiple comparisons if you expect to peek at intermediate checkpoints.

Test plan

Decision framework

1. Pre-register hypotheses and guardrails in _datasets/experiment-hypotheses.csv.
2. Automate metric extraction via notebooks stored in _notebooks/.
3. Attach Tableau or Looker dashboards with the viz-block include for executive readouts.

By combining Markdown tables, statistical code, and callouts, the post becomes a reusable experimentation template for every squad.

Theorem

Theorem 1. Let $f : [a, b] \to \mathbb{R}$ be twice continuously differentiable with $f(a) = f(b) = 0$. Then there exists $c \in (a, b)$ such that $f''(c) + \frac{\pi^2}{(b-a)^2} f(c) = 0$.

Proof

We adapt the standard Wirtinger inequality. Consider the sine basis function $g(x) = \sin\left(\frac{\pi(x-a)}{b-a}\right)$ and define

Integration by parts shows that

Combining Equations \eqref{eq:inner-product} and \eqref{eq:ibp} yields the desired critical point when $f$ is not identically zero. $\square$

Notes for authors

• MathJax automatically numbers equation environments so you can reference them with \eqref{}.
• Inline math, such as $\int_a^b f(x)\,\mathrm{d}x$, remains crisp across light and dark modes.
• Use definition, lemma, and corollary blocks as needed—Markdown blockquotes keep the typography consistent.

Add proof sketches, exercises, or downloadable solution PDFs to round out your mathematical articles.

Revenue trends

Why it matters

• Keyboard users gain export controls via the toolbar.
• Metadata automatically exposes the chart title and version in screen reader-friendly text.
• Embedding raw JSON keeps version diffs tight during code review.

Pair this Plotly block with D3, Observable, and Shiny embeds to demonstrate the full visualization gallery offered by the theme.

Abstract

We evaluate adaptive experimentation for recommendation systems, focusing on policy regret minimization across cold-start cohorts. Empirical results indicate a 12% lift in engagement relative to static baselines while maintaining fairness constraints.

Introduction

Personalized experiences need to balance accuracy and fairness. Prior work on contextual bandits¹ and constrained optimization² lays the foundation for our framework.

Methodology

We define policy regret as

where $r_t$ is the reward and $a_t^\star$ is the action chosen by an oracle. Algorithm 1 summarizes the constrained Thompson sampling procedure.

Initialize posterior priors for all arms
for each round t = 1..T:
  sample reward estimates from posterior
  project samples to satisfy fairness constraints
  choose arm with highest adjusted draw
  update posterior with observed reward

Results

Discussion

Equation \eqref{eq:regret} highlights how regret decomposes into reward differences. Future work will incorporate causal constraints to prevent drift.

Cite this work

1
2
3
4
5
6
7
8
9
10
@article{ribeiro2024adaptive,
  title = {Adaptive Recommendation Under Fairness Constraints},
  author = {Ribeiro, Diogo and Smith, Ada},
  journal = {Journal of Responsible AI},
  year = {2024},
  volume = {12},
  number = {2},
  pages = {45--63},
  doi = {10.1234/jrai.2024.5678}
}

Add this BibTeX block to your citation manager or to the CITATION.cff file when you release accompanying code. The DataLog theme handles footnotes, equations, and code blocks seamlessly in a single article.

Baseline query

1
2
3
4
5
6
7
8
SELECT
  customer_id,
  SUM(spend) AS total_spend,
  AVG(spend) AS avg_spend,
  ROW_NUMBER() OVER (PARTITION BY customer_id ORDER BY purchase_ts DESC) AS purchase_rank
FROM analytics.fact_orders
WHERE purchase_ts >= DATEADD('day', -90, CURRENT_DATE)
GROUP BY 1

Use the built-in explain tools to gather diagnostics:

1
2
3
4
5
6
7
8
9
10
11
12
13
EXPLAIN USING JSON
SELECT *
FROM (
  -- baseline aggregation
  SELECT
    customer_id,
    SUM(spend) AS total_spend,
    AVG(spend) AS avg_spend
  FROM analytics.fact_orders
  WHERE purchase_ts >= DATEADD('day', -90, CURRENT_DATE)
  GROUP BY 1
) src
JOIN analytics.dim_customer dc USING (customer_id);

Optimization checklist

1. Materialize the 90-day window as an incremental model.
2. Cluster the fact table on purchase_ts to prune partitions.
3. Replace repeated JSON parsing with persisted staged columns.
4. Cache high-cardinality dimension joins using search optimization services.

Note: Store the JSON explain output in _datasets/ so reviewers can diff the query plan over time.

Annotate performance wins

Wrap up by linking to dbt models, scheduling notes, and alert thresholds so stakeholders can keep the warehouse humming.

Notebook overview

The project notebook notebooks/churn-segmentation.ipynb contains:

• Feature engineering with pandas and scikit-learn ColumnTransformer
• Model training using xgboost.XGBClassifier
• MLflow logging for parameters, metrics, and artifacts

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
from xgboost import XGBClassifier
from sklearn.compose import ColumnTransformer
from sklearn.pipeline import Pipeline
from sklearn.preprocessing import OneHotEncoder, StandardScaler

numeric = ["monthly_charges", "tenure", "support_tickets"]
categorical = ["contract", "region"]

preprocess = ColumnTransformer(
    [
        ("num", StandardScaler(), numeric),
        ("cat", OneHotEncoder(handle_unknown="ignore"), categorical),
    ]
)

model = Pipeline(
    steps=[
        ("preprocess", preprocess),
        (
            "classifier",
            XGBClassifier(
                max_depth=4,
                n_estimators=200,
                subsample=0.8,
                colsample_bytree=0.9,
                eval_metric="auc",
            ),
        ),
    ]
)

Launch options

Readers can open the notebook in the environment of their choice, while the theme preserves accessibility labels for screen readers.

Track experiments

1
2
3
4
5
6
7
8
import mlflow

mlflow.set_experiment("churn-segmentation")
with mlflow.start_run(run_name="xgboost-baseline"):
    model.fit(train_features, train_labels)
    auc = model.score(test_features, test_labels)
    mlflow.log_metric("test_auc", auc)
    mlflow.xgboost.log_model(model.named_steps["classifier"], "model")

Embed evaluation tables and charts produced by MLflow in the post so stakeholders understand progress:

Checklist before deployment

• Notebook executed from top to bottom without errors
• Model registered with reproducible environment metadata
• Alert thresholds documented for precision/recall trade-offs

DataLog’s notebook integration keeps workflows transparent—link to runnable notebooks, surface experiment logs, and capture decisions alongside the code that produced them.