Model performance at a glance
Synthetic comparison · low-rating class · values on a 0–1 scale
RecallF1 scorePR-AUC
Demo reading: a higher recall finds more low-rating orders, but precision and workload must also be considered.
Eligible orders
91,056
Reported cohort · RQ1 & RQ2
Low-rating share
9.45%
8,608 orders · 1–2 stars
Shopping baskets
95,146
723 multi-category baskets
Text reviews
32,644
Reported high / low review subset
RQ1 / CLASSIFICATION
THE CENTRAL QUESTION
Which orders deserve
attention first?
Rank delivered orders by predicted low-rating risk. Evaluate how many actual low-rating orders appear within the top review budget.
Top 10%
A practical review budget, not a model accuracy target.
Order profiles
Four illustrative segments · not fitted clusters
● Routine orders 42%
● High-value orders 24%
● Freight-heavy 19%
● Delayed orders 15%
Delivery experience & rating risk
Synthetic low-rating rates by delivery delay
Illustrative association only. This chart does not establish that delivery delay causes dissatisfaction.
Four questions. One customer journey.
Cumulative low-rating coverage
Generated from 500 synthetic orders · not the report cohort
Selected modelExpected random ranking
REVIEW BUDGET SIMULATOR
Make the trade-off visible.
Adjust the share of orders reviewed. All numbers are computed from the same synthetic cohort.
50Orders reviewed
—Low-rating coverage
—Review precision
Coverage = low-rating orders found ÷ all low-rating orders. Precision = low-rating orders found ÷ reviewed orders.
Highest-priority synthetic orders
Display identifiers and scores are generated for this prototype.
| Order | Risk score | Delay | Freight share | Observed synthetic label |
|---|
Order segmentation landscape
Synthetic 2D projection · 160 points · not a fitted model
RoutineHigh-valueFreight-heavyDelayed
INTERPRETATION PRINCIPLE
Group on behavior.
Compare on experience.
Use order value, item count, freight share and delivery delay to build clusters. Compare low-rating rates afterwards to avoid defining groups by their outcome.
PLANNED EXPERIMENT
Standardize numeric features · try 2–6 clusters · repeat random seeds · examine cluster size and stability.
Illustrative segment profiles
These names and values are placeholders; final profiles will follow the actual cluster centers.
| Segment | Share | Mean order value | Mean freight share | Mean delay | Low-rating rate |
|---|---|---|---|---|---|
| Routine | 42% | R$ 94 | 15% | −3.4 days | 4.8% |
| High-value | 24% | R$ 486 | 8% | −2.1 days | 7.2% |
| Freight-heavy | 19% | R$ 71 | 32% | −0.8 days | 10.4% |
| Delayed | 15% | R$ 138 | 18% | +8.2 days | 25.6% |
Category affinity network
Illustrative links only · line thickness represents example counts
An edge indicates co-occurrence, not causation or proven recommendation performance. This network contains no mined rules.
EVIDENCE BEFORE RECOMMENDATION
Sparse baskets. Careful claims.
723Reported multi-category baskets
0.76%Share of reported baskets
The proposed main threshold is 20 joint occurrences, confidence ≥ 10%, and lift > 1. All 95,146 baskets remain in the support denominator.
Compare thresholds of 10, 20 and 30. Reporting no qualifying real rules is a valid project outcome.
Example rule explorer
Count-based support uses 95,146 as denominator; confidence and lift use fabricated antecedent and consequent counts.
| Example rule | Joint count | Support | Confidence | Lift |
|---|
High-rating vocabulary
4–5 STARSDecorative term sizes · not measured TF–IDF weights
excelenterápidorecomendoqualidadeperfeitosatisfeitoentrega
“Produto excelente, chegou antes do prazo.”Constructed Portuguese example: Excellent product; it arrived before the deadline.
Low-rating vocabulary
1–2 STARSDecorative term sizes · not measured TF–IDF weights
atrasonãorecebiproblemadefeitodemoraerrado
“A entrega atrasou e o produto veio com defeito.”Constructed Portuguese example: Delivery was late and the product arrived defective.
Text classification snapshot
Fabricated holdout confusion matrix · 1,000 synthetic reviews
| Actual / predicted | Low rating | High rating |
|---|---|---|
| Low rating | 240 | 60 |
| High rating | 90 | 610 |
0.829Macro F1
0.800Low-class recall
0.871High-class recall
Scores are calculated from the displayed illustrative matrix; they are not Weka experiment results.
TEXT MINING PIPELINE
Words become features.
Context stays essential.
Normalize text → apply Portuguese stopwords → retain meaningful negation → StringToWordVector → TF–IDF → NaiveBayesMultinomial.
Learn the vocabulary and IDF weights from training data only. Star-rating groups are the classification labels; these are not manually annotated complaint topics.
Overall technical workflow
01Integrate & inspect
Review nine tables. Aggregate items and payments before order-level joins.
02Prepare task data
Apply eligibility rules. Build order, basket and review datasets.
03Run Weka experiments
Classification, clustering, association mining and text classification.
04Evaluate & communicate
Independent time tests, stability checks, sensitivity analysis and bounded conclusions.
Proposed RQ1 time cutoff: 11 June 2018, 00:00. Train only on orders delivered and reviewed before the cutoff. Orders delivered earlier but reviewed later are recorded separately.
Research tools
Versions shown below are supplied by the report; remaining versions need to be recorded.
| Tool | Version | Purpose |
|---|---|---|
| Python | 3.12.14 | Preparation and splitting |
| pandas | 3.0.1 | Aggregation and joins |
| Weka / Java | To record | Mining experiments |
| Matplotlib / Seaborn | Planned | Report figures |
| Streamlit / Plotly | Planned | Python-based dashboard |
This downloadable prototype uses standalone HTML, CSS, JavaScript and inline SVG. It does not run Weka or Streamlit.
Delivery checklist
Proposal-stage status based on the supplied project description.
Proposal & presentationIn preparation
Initial cleaning & cohort definitionReported complete
Independent time-test resultsPlanned
Final models & parameter logsPlanned
Final report & reproducible dashboardPlanned
Historical findings apply to eligible 2016–2018 orders. Review-text findings apply only to the written-feedback subset.