Understanding Pinoy Agricultural-Commodity Consumption Patterns for Digital Food-Diversification Initiatives
Main Research Question
Rice & Sweeetpotato Companion Staple Paradigm"What can agricultural-commodity consumption patterns tell us about households that report kamote alongside rice, and how can those findings guide practical, responsible social-computing initiatives?"
About
Dr. Armida P. Salazar
Professor • apsalazar@national-u.edu.ph
Peter-Paul "Kuya Pete" Gumpal, MM, CLSSGB
Learner/Researcher • gumpalpd@students.national-u.edu.ph
Course & Department
DITSPEC7 — Social Computing Algorithms & Techniques
College of Computing & Information Technologies
Empirical Foundation
PSA 2015–2016 Survey of Food Demand (SFD)
Before diving into machine learning metrics, we must anchor our research in the urgent Philippine macroeconomic reality of October 2026: escalating foreign rice dependency, massive public-fund opportunity costs, and a population deriving over 50%+ of its daily energy from a single climate-vulnerable grain.
On October 7, 2026, the Philippine Rice Industry Stakeholders Movement (PRISM) and the USA Rice Federation signed an MOU in Manila targeting 100,000 MT of US-origin rice in 2027 and 200,000 MT annually from 2028—adding to the country's ~4.7M MT total annual rice import dependency since the 1990s.
*Valuation note: At landed/wholesale import parity (~$600–$650/MT or ₱36,000–₱45,000/MT retail equivalent), 200,000 MT represents ₱7.2B–₱9.0B flowing to foreign farmers annually.
While rice paddies remain acutely vulnerable to El Niño, La Niña, and typhoons, state reviews of the ₱545.64 Billion Flood Control Budget (2022–2025) exposed massive governance gaps that could have transformed Philippine food security, healthcare, and education:
The Connection: We pour hundreds of billions into reactive infrastructure and foreign grain imports while neglecting low-cost, typhoon-resilient root crops (Ipomoea batatas) and evidence-based community nutrition systems.
Select a benchmark figure below to see how redirecting wasted or import-drained billions could strengthen our Health System, Education System, and Local Jobs:
Selected Fund Benchmark:
200K MT US Rice Import Deal (~₱7.6B/yr)
76
At ~₱100M per equipped Rural Hospital / Center
3,040
At ~₱2.5M per standard DepEd classroom
38,000
At ~₱200,000 annual rural/agri-tech livelihood
380
At ~₱20M per municipal curing/flour facility
Instead, our research shows how low-cost Social Computing & Digital Information Systems can catalyze demand-side dietary resilience without waiting for multi-billion-peso mega-projects.
Filipinos derive 50%+ of their daily dietary energy from Rice, consuming 109.87 kg/person/year (2.11 kg/week) vs. only 4.53 kg/year of Kamote. Across ~115M Filipinos (~12.64M MT annual rice food demand), slide below to see what happens when households adopt Kanin + Kamote for a small fraction of their weekly meals:
631,775 MT
Metric tons of milled rice per year
3.16x
More than offsets the entire 2028 US MOU!
₱24.0 Billion
Redirected to Pinoy root-crop farmers
-5.49 kg/yr
Low-GI fiber boost for blood-sugar control
Key Takeaway: We don't need Pinoys to abandon rice. Even a 2% to 5% "Kanin + Kamote" companion habit wipes out the 200,000 MT US rice import quota while improving glycemic health!
In Social Computing, jumping straight into Sentiment Analysis and Opinion Mining without knowing who already consumes kamote, where, and alongside what foods is like designing a campaign in the dark. Going back to the drawing board yielded a far stronger scientific foundation.
Connecting PSA historical consumption evidence to testable digital services and future opt-in sentiment/behavior feedback (informed by Michie et al.'s Behavior Change Wheel: Capability, Opportunity, Motivation).
Public-use microdata from the PSA 2015–2016 Survey of Food Demand (SFD):
Auditable data engineering & scientific triangulation:
Evidence-to-action social computing blueprint:
Instead of quietly dropping difficult records, our pipeline maps raw PSA public-use microdata into 8 linked canonical tables and a leak-free 84-predictor analytical feature mart.
91.88%
Households Citing Rice as Staple
93.39% consumed plain rice (109.87 kg/yr per capita; 113.0 kg rural vs 102.3 kg urban; 88.24% bought).
24.30%
Households Consuming Kamote
Ahead of White Potato (15.49%), Cassava (13.19%), and Gabi (11.48%). Avg 4.53 kg/yr (17.33 kg/yr in Zamboanga Pen.).
49.76%
Cite "Preference" Over Rice
Per PSA Table 12.2: 49.76% nationally (and 75.21% in NCR!) substitute kamote because it is preferred, vs only 8.73% due to rice unaffordability.
18.32 kg
Non-Human / Wasted Rice per HH/Yr
PSA Tables 13.1–13.3: 1.68 kg spoiled/wasted + 11.26 kg leftovers fed to pets + 5.38 kg purposely cooked for pets/animals yearly.
Following Appendix C (Steps C1–C5), we engineered a reproducible pipeline that preserves source lineage and prevents target leakage:
Mapped two-level questionnaire headers across 4 survey rounds (Aug 2015, Nov 2015, Feb 2016, May 2016) covering 80 provinces + NCR.
Normalized into 8 relational CSV tables: Households, Members, Food Commodities, Rice/Corn Waste, Rice Substitution, Substitution Reasons, Substitution Foods, and Meal Patterns—storing conformance flags instead of silently deleting rows.
Denormalized 84 approved predictors. All direct kamote fields (quantity, price, source, substitution status) were strictly excluded to prevent target leakage.
| Pipeline Stage / Subset | Records | Kamote + Rice | Methodological Note |
|---|---|---|---|
| Source household-round records | 52,522 | 11,542 (21.98%) | 12,487 total kamote-positive (92.4% also ate rice) |
| Rice-consuming records | 49,064 | 11,542 (23.52%) | Excludes 3,458 non-rice observations |
| ML Cohort (Valid member data) | 49,004 | 11,527 (23.52%) | 60 records excluded due to member-data check |
| ↳ Primary Development Split (80%) | 39,203 | 9,221 (23.52%) | Sub-split into 29,402 fit (3-fold CV) + 9,801 val |
| ↳ Internal Model-Validation Split | 9,801 | 2,306 (23.53%) | Matched benchmark for LR, RF & CatBoost |
| ↳ Historical Exploratory Test Split (20%) | 9,801 | 2,306 (23.53%) | Examined in Exp. 1; disclosed as non-untouched |
household_id is a unique record identifier per survey round, not a verified cross-round physical household link. We therefore analyze household-round observations and treat random splits as exploratory.
Our descriptive mining overturns the assumption that kamote competes against rice. Instead, 92.4% of all kamote-reporting households also consumed rice in the same week—varying significantly across survey rounds, regions, and root-crop familiarity.
Unweighted share of all 52,522 household-round observations during the 7-day recall window.
Share of eligible rice-consuming observations also reporting kamote in each quarterly wave.
Comparing unweighted vs. provisional survey-weighted kamote rates by non-kamote root crop count (cassava, gabi, white potato).
Provisional weighted kamote co-reporting rises from 18.69% (1–2 members) to 30.45% (7+ members). Per PSA Table 14, 12+ member households eat 9.31 kg/yr per capita.
Rural households consume more kamote overall (5.26 kg/yr vs 2.78 kg/yr urban), but 97.87% of urban kamote-eating households also eat rice (vs 91.26% rural).
Nationally, 60.46% buy kamote and 26.57% produce their own (PSA Table 6.05). Top consumers: Zamboanga Pen. (17.33 kg/yr), Bicol (8.93 kg/yr), and Eastern Visayas (6.86 kg/yr).
We evaluated Logistic Regression, Random Forest, and CatBoost on the exact same 84 approved predictors (29,402 fit records; 9,801 internal validation records) to test whether household patterns hold predictive signal and where redundancy lies.
Regularized linear baseline that combines mostly additive relationships across the 84 predictors.
Aggregates 150 randomized decision trees (max_depth=16, min_samples_leaf=10, max_features=0.40) selected from 6 CV configs.
Builds sequential trees (400 iterations, depth=6, lr=0.05, L2=5) to correct prior errors and handle categorical geography natively.
Comparing Train ROC-AUC, 3-Fold CV ROC-AUC (n=29,402), and Internal Validation ROC-AUC (n=9,801). Notice Random Forest's large train–validation overfitting gap (0.1628).
CatBoost leads in every round (highest in May 2016 at 0.7396 and Aug 2015 at 0.7328; weakest in Nov 0.7017 and Feb 0.7011), proving why a single national metric hides seasonal/round variation.
Normalized PredictionValuesChange across top 15 individual predictors: Province (13.12), Region (8.44), and Survey Round (7.95) lead, followed by Other Food Variety (4.44) and Cassava Consumed (3.13).
When we sum importance across 54 individual foods (29.36%), they look huge. Yet when we remove all 54 foods + 7 diversity summaries and retrain CatBoost (Step Q), ROC-AUC drops by only -0.0076 (from 0.7267 to 0.7191)! Removing Geography + Round drops ROC-AUC by -0.0577 (to 0.6690)—over 7.5x larger loss!
| Retrained CatBoost Configuration | Kept | ROC-AUC | AP | Δ AUC |
|---|---|---|---|---|
| Full Model (All 84 Approved Predictors) | 84 | 0.7267 | 0.4688 | 0.0000 |
| Exclude 7 dietary-diversity summaries | 77 | 0.7265 | 0.4679 | -0.0002 |
| Exclude 54 individual non-kamote foods | 30 | 0.7231 | 0.4595 | -0.0037 |
| Exclude 54 foods + 7 diversity (23-feat model*) | 23 | 0.7191 | 0.4525 | -0.0076 |
| Exclude survey round | 83 | 0.7103 | 0.4577 | -0.0164 |
| Exclude geography (province & region) | 82 | 0.6812 | 0.4122 | -0.0456 |
| Exclude geography + survey round | 81 | 0.6690 | 0.4043 | -0.0577 |
Because historical commodity records cannot tell us who decided what to cook, why, or which digital message persuades families, our paper translates these findings into a 5-part Information System Blueprint and concrete pilots under Project TIWALA (www.tiwala.org.ph).
Rather than labeling or targeting individual households from historical 2015–2016 predictions, Project TIWALA hosts voluntary, consent-based digital services protected under Republic Act No. 10173 (Data Privacy Act of 2012):
Alalay sa Nutrisyon at Kalusugan
Hosts DOST-FNRI Pinggang Pinoy-aligned, culturally familiar "Kanin + Kamote" recipe cards (e.g., breakfast pairings, cassava/root-crop combos) and non-clinical glycemic explanations.
Ulat Produkto at Pagkaing Pinoy
Connects local commodity availability, municipal market price signals, and community-supplied questions—bridging supply bottlenecks with household demand.
How our completed data-mining pipeline evolves into an interactive Social Computing system where Opinion Mining & Sentiment Analysis now operate with high precision:
Automated schema mapping, canonical table validation, and audit-flag preservation ready to ingest future PSA SFD waves (including a possible 2025–2026 release).
Visualizes regional co-consumption patterns, seasonal survey rounds, and root-crop synergies for LGUs and researchers without exposing individual households.
Culturally familiar "Kanin + Kamote" explainers tailored to local cooking habits (home-prepared vs. bought) and reviewed by DOST-FNRI / nutrition specialists.
Two-question micro-surveys and Taglish sentiment/opinion mining capturing taste, convenience, cost, stigma, and family roles directly from consenting participants.
Evaluates whether digital messages actually shift voluntary kitchen trials (comparing e.g. breakfast kamote recipes vs. affordable side-dish framing).
| Horizon | Proposed Lead & Partners | Suggested Digital or Community Action | What to Measure (Metrics) |
|---|---|---|---|
| 0–3 Months | Project TIWALA + LGU / University | Publish a simple area-level dashboard and expert-reviewed "Kanin + Kamote" explainers | Reach, understanding, opt-in interest |
| 0–6 Months | A.N.A.K. + DOST-FNRI Partner | Test accessible, culturally familiar recipe cards and non-clinical nutrition explanations | Comprehension, recipe trial, feedback |
| 3–12 Months | Rice UPPP! + DA / DTI / LGU | Explore community-supplied food availability and public commodity-market price data | Coverage, accuracy, user value |
| 6–18 Months | Project TIWALA + Academe | Run consent-based household interviews and randomized digital-message pilots (ABSA / NLP ready) | Willingness, observed trial, repeat use |
| Long Term | PSA + Research Partners | Map newer surveys (e.g., 2025–2026 SFD wave), recover household linkage, and test transferability | Schema audit, independent replication |
The target is reported household kamote consumption during the same 7-day recall window as rice—not proof that kamote was eaten at the same meal or intended to replace rice.
Unique derived household_id values identify 49,004 rows, not verified independent physical households across the 4 rounds. Record-level splits may therefore be slightly optimistic.
Full PSA two-stage stratification and PSU sampling weights were not embedded in ML loss functions; bootstrap intervals are conditional and exploratory.
Both the 9,801 historical test split and the 9,801 internal validation split were consulted across Steps D–Q. Confirmatory claims require a genuinely unseen wave or recovered household-grouped split.
Generative AI was used strictly as a supporting writing, terminology-learning (e.g., distinguishing gradient descent from gradient boosting), and Python/Colab troubleshooting assistant—never as the source of survey numbers. All statistical outputs were verified directly against the researcher's Google Colab notebooks and official PSA SFD Volume 1 tables.
Complete reproducibility package preserves: (1) PSA SFD 2015–2016 source citation & questionnaire; (2) Raw-to-Canonical Python contract & 8 canonical CSVs; (3) Shared-v2 84-feature mart; and (4) Dedicated Google Colab notebooks for Logistic Regression, Random Forest, and CatBoost (Steps A–Q).