From Rice Dependence to Ninamote | NU Manila Social Media Analytics Presentation

National Context • Staple Diversification in October 2026

Why Staple Diversification Matters Beyond Our Classroom

Before discussing our NLP algorithms and machine learning models, let's ground this research in what happened earlier this October 2026: private rice-import agreements, national resource scrutiny, and a simple thought experiment on what happens when Filipinos make room for a locally grown companion staple.

Manila Standard • Oct 8, 2026 O. Campos Article

"Philippines Wants US Rice": 100K MT (2027) to 200K MT/Year (2028)

The Philippine Rice Industry Stakeholders Movement (PRISM) signed an MOU with the USA Rice Federation targeting 100,000 MT of American rice in 2027 and 200,000 MT annually beginning in 2028. Why does a country with such a strong agricultural identity remain so dependent on imported rice?

2027 Initial Target 100,000 MT ~₱3.6B – ₱4.5B Scenario

2028 Annual Target 200,000 MT ~₱7.2B – ₱9.0B / yr

Broader Stakes FX & Farms Supply & farmer future

Analytical Precision Note: This MOU is a private industry agreement—not a completed purchase or direct government spending program—and rice imports help stabilize supply when domestic output is short. The ₱7.2B–₱9.0B figure is an illustrative price-based scenario (at ₱36,000–₱45,000/MT), not a disclosed contract value.

Public Resource Scrutiny COA, Senate & DOF Reviews

The Public Spending Comparison: National Resources & Resilience

During the government's review of flood-control projects, around ₱545 billion in projects came under scrutiny—raising a vital public-policy question: How effectively are we using resources to strengthen our country's long-term agricultural and community resilience?

Under Scrutiny ~₱545B Flood-control review

Unspecified Details ₱350B+ Lacked structure detail

DOF Est. Ghost Loss ~₱118.5B Estimated economic loss

Analytical Caveat: These figures represent distinct audit and economic-loss benchmarks—they are not all proven losses and cannot simply be added together. Yet they highlight why diversifying into resilient local root crops deserves serious attention alongside infrastructure.

Opportunity Cost Calculator Interactive: Click all 3

What Could Those Billions Mean in Concrete Terms?

A hypothetical scale comparison (not a government costing study or automatic budget transfer) to help visualize the magnitude of economic resources in food supply and public investment:

Selected Comparison Benchmark:

200K MT US Rice Midpoint Scenario (~₱7.6B/yr)

Hospitals / Rural Health

Assumed ~₱100M per Rural Health Facility

Public Classrooms

3,040

Assumed ~₱2.5M per standard classroom

Livelihood Support

38,000

Assumed ~₱200,000 annual agri-livelihood

Kamote Processing Hubs

380

Assumed ~₱20M per local processing hub

Key Takeaway: Stronger local root-crop value chains can create opportunities for farmers, processors, and small businesses—provided we also address production capacity, logistics, and consumer readiness.

"Kanin + Kamote" Thought Experiment Drag: 5% → 10% → 5%

What If Filipinos Made Just a Small Dietary Change?

Earlier PSA baseline data showed 109.87 kg of plain rice per person/year vs. only 4.53 kg of kamote. What if total rice consumption were reduced by just 5%, with some food demand shifting toward kamote?

Hypothetical Rice Reduction / Companion Shift (%): 5%

Proportional Rice Volume

631,775 MT

~632,000 MT annually at 5%

vs. 200K MT US Rice MOU

3.16x

Multiple of proposed 2028 US volume

Illustrative Rice Value

₱24.0 Billion

Arithmetic scenario at ₱38k/MT

Per Capita Rice Shift

-5.49 kg/yr

From 109.87 kg/yr baseline

The Catch That Starts Our New Paper: "A mathematical scenario is easy to build. Convincing real people to change their daily food habits is much harder. Before we talk about rice-demand reduction, will people even consider incorporating kamote alongside rice?"

Empirical Foundation • Survey-Based Social Sensing (Sections I–III)

Survey-Based Social Sensing: Listening to 357 Filipino Consumers

Originally, we planned to scrape public Facebook comments. Instead, we started with a more direct, high-signal empirical foundation: 357 digitally collected responses via Google Forms and Meta Lead Forms—framing our study as survey-based social sensing anchored in Diffusion of Innovations (DOI).

Study Figure 2 N = 357 Total

Survey Responses by Digital Collection Channel

How social media expanded our recruitment reach far beyond personal networks:

Google Forms (n = 35 | 9.8%): Circulated to friends, relatives, former coworkers, and senior farmers reached with the help of Pablo A. Gumpal, Jr.

Meta Lead Forms (n = 322 | 90.2%): Included 320 Facebook leads and 2 Instagram leads—proving social media's power for rapid digital sensing recruitment.

Study Table I

Dataset Overview, Q5 Audit & Analytic Subsets

Zero Duplicate IDs

Analytic Subset / Item	Count	Methodological & Audit Notes
Total harmonized digital responses	357	35 Google Forms; 322 Meta Lead Forms (320 FB, 2 IG)
Eligible binary ML modeling cases	296	223 Willing (positive class, 75.3%) vs. 73 Conditional or Unwilling (negative class, 24.7%)
Excluded from main binary model	61	16 already practicing + 45 unclear Q5 responses
Distinct normalized text response groups	517	Grouped from free-text answers across Q3 (cost), Q4 (culture/habit), and Q9 (channels)
↳ Rule-Development Groups	85	Used for human-guided codebook refinement & rule tuning
↳ Reserved Blind Validation Groups	60	Strictly held out to test frozen NLP rules without data leakage

What is "Ninamote"? Coined from Kanin + Kamote—making kamote a regular companion to rice, not a total replacement.

Theoretical Architecture (Section II)

Why Diffusion of Innovations (DOI) Leads & Social Cognitive Theory (SCT) Contextualizes

Our questionnaire was designed around 5 practical aspects (Price & Affordability, Availability & Logistics, Cultural Factors, Health & Nutrition, and Environmental Awareness). However, our paper rigorously separates broad survey intent from validated computational measurement:

Primary Lens: Rogers' DOI (2003)

• Relative Advantage (Strongest Measure): Captured via Q3 comparative economic perceptions (Acc: 0.850).
• Compatibility (Partial Measure): Captured via Q4 cultural stigma & daily rice habits (weaker psychometrics).
• Complexity, Trialability, Observability: Treated transparently as underdeveloped for future waves.

Secondary Lens: Bandura's SCT (2001)

• Explains why reporting a social-media channel (Q9) ≠ behavioral persuasion.
• True social learning requires measuring self-efficacy, observational learning, and reinforcement.
• Guides the next phase of Project TIWALA experimental messaging.

Filipino farmers harvesting kamote connected with barangay consumers on smartphones

Mula Bukid Hanggang Digital Hapag

Bridging Senior Farmers & Social Media Leads

Combining ground-level rural perspectives (via Google Forms) with broad Facebook & Instagram recruitment (322 Meta leads) under Project TIWALA.

Methodology • Human-Guided Filipino & Taglish NLP (Answering RQ#1 & RQ#2)

Human-Validated NLP: Turning Taglish Nuance Into Auditable Signals

Filipino and Taglish survey responses are short, informal, and culturally nuanced—some say "Mas mura", others say "Mahal din ang kamote" or "Sanay na kami sa kanin." Generic off-the-shelf sentiment tools miss these implicit comparisons, so we built a human-validated, frozen-rule NLP workflow.

Study Figure 1

Implemented Survey-Based Social Sensing Workflow

Human Checkpoints Before Freezing Rules

Stage 1

Digital Survey Collection

357 responses (35 Google Forms + 322 Meta Lead Forms)

Stage 2

Data Harmonization

Unified schema, de-identification & text-quality audit

Stage 3

Q5 Outcome Locking

5 classes audited → 296 binary eligible cases locked

Stage 4

Human-Guided NLP Coding

517 distinct groups; 85 dev groups for rule design

Stage 5 ★

Blind Rule Validation

Frozen rules tested on 60 reserved blind groups

Stage 6

Interpretable ML Modeling

Regularized Logistic Regression (5×5 Repeated CV)

Stage 7

Ablation & TIWALA Action

Coefficients, sensitivity & streaming roadmap

Interactive Taglish NLP Inspector Q3, Q4 & Q9 Features

How Our Frozen Rules Decode Real Pinoy Survey Phrases

Click a real Filipino/Taglish survey expression below to inspect how the human-validated codebook transforms raw text into structured indicators:

Question Q3 • Economic Perception Held-Out Acc: 0.850 ★

Raw Expression: "Mas mura ang kamote sa palengke kumpara sa bigas" / "Madaling itanim sa bakuran"

→ Extracted Label: economic_direction = ADVANTAGE

Human-in-the-Loop Rule Note: Captures explicit comparative savings, backyard planting ("walang gastos"), or budget relief relative to rice. Strongest validated predictor of adoption readiness (78.7% willing).

Scientific Honesty: "In research, knowing what your model cannot measure reliably is just as valuable as knowing what it can measure."

Study Table II 60 Blind Validation Groups

Held-Out NLP Feature-Extraction Performance

Evaluating our frozen rules on 60 unseen response groups explains why Economic Direction (0.850 accuracy) earned its place in the primary readiness model while weaker constructs were documented with caution:

Extracted Feature	Held-Out Metric	Value	Paper Interpretation
Economic direction (Q3)	Accuracy / Macro-F1	0.850 / 0.707	Strongest validated feature ★
Information channels (Q9)	Exact-set / Micro-F1	0.650 / 0.847	Good practical multi-label extraction
Stigma stance (Q4)	Accuracy / Macro-F1	0.600 / 0.444	Useful but weak & error-prone
Rice-habit stance (Q4)	Accuracy	1.000	Sparse-category result; interpret cautiously

Main Empirical Discovery • Economic Relative Advantage (Section IV & Figure 4)

"Makakatipid Ba Kami?": Economic Perception Drives Adoption Readiness

Among the 296 eligible respondents (223 willing; 73 conditional or unwilling), one dramatic pattern stood out: respondents who see an economic advantage in kamote report 78.7% willingness, whereas those who perceive an economic disadvantage drop to 50.9%—nearly a 28-percentage-point gap!

Study Figure 4

Unqualified Stated Willingness by Economic Perception

Across all 5 Q3 economic-direction categories (Figure 4), expressed economic disadvantage (50.9%) and mixed cost views (57.1%) show markedly lower unqualified willingness than economic advantage (78.7%).

Economic Advantage 78.7% 118 of 150 willing

Economic Disadvantage 50.9% 28 of 55 willing

Logistic Odds Multiplier ~3.66× Higher odds of willingness

Ang Tanong sa Bawat Kusina

What Filipino Families Actually Ask Before Changing Habits

Maybe one of the biggest questions Filipinos ask is not simply whether kamote is nutritious—it's whether it makes practical economic sense for the family budget:

"Makakatipid ba kami?"

"Mas mahal ba ito kaysa kanin?"

"Sulit ba itong isama sa araw-araw?"

DOI Relative Advantage & Robustness Check

In Rogers' Diffusion of Innovations, people are more willing to adopt an innovation when they perceive a clear Relative Advantage over current practice.

Holds in Meta-Only Subset (n=180 explicit): Willingness is 81.1% for economic advantage vs. 50.9% for disadvantage—proving the gap isn't an artifact of mixing Google Forms and Facebook leads.
Association, Not Causation: We carefully state that perceived economic disadvantage is strongly associated with lower stated readiness in our survey.

Model Evaluation • Interpretable Machine Learning & Ablation (Answering RQ#3)

The Machine Learning Result That Surprised Me

Using Regularized Logistic Regression with 5×5 repeated stratified cross-validation (n=296), we tested whether survey source, economic perceptions, and reported social-media channels predict adoption readiness—and ran an ablation experiment that revealed a crucial lesson for Social Media Analytics.

Model M0 AUC: 0.500

Majority Baseline

Predicts everyone as "Willing" (75.3% accuracy, but learns zero signal about the negative class).

• Accuracy: 0.753

• Balanced Acc: 0.500

• Minority F1: 0.000

Model M1 AUC: 0.521

Survey Source Only

Uses only collection channel (Google Forms vs. Meta Lead Forms)—very limited discrimination.

• Accuracy: 0.701

• Balanced Acc: 0.521

• Minority F1: 0.207

Model M2 AUC: 0.659

Theory-Guided Features

Economic perception + reported social media channel without survey source indicator.

• Accuracy: 0.654

• Balanced Acc: 0.584

• Minority F1: 0.379

Model M3 ★ AUC: 0.656

Combined Model

Combines survey source + theory-guided features; achieves highest Balanced Accuracy & Minority F1.

• Accuracy: 0.708

• Balanced Acc: 0.632 ★

• Minority F1: 0.441 ★

Table III & Figure 3

Primary Model Comparison (5×5 Repeated CV)

Adding theory-guided features jumps ROC-AUC from 0.521 (M1 Source Only) to 0.659 (M2) and 0.656 (M3). Why isn't 0.67 higher? Because our goal is interpretable signal discovery, not black-box overfitting!

Why Accuracy Alone Misleads: Because 75.3% of the 296 respondents are willing, M0 gets 0.753 accuracy with 0.000 Minority F1. Balanced Accuracy (`0.632`) and Minority F1 (`0.441`) prove M3 actually learns who is hesitant!

Table IV • The Ablation Surprise

Economic Signal vs. Reported Social Media Channel

Look at what happens when we add predictors one by one to Survey Source (`0.521` ROC-AUC): adding Economic Perception jumps ROC-AUC to 0.671 (+0.150!), whereas adding Reported Social Media yields 0.520 (zero gain!).

Analysis Stage	Model Specification	Bal Acc	ROC-AUC	Min F1
Ablation	Source only	0.521	0.521	0.207
Ablation ★	Source + Economic Perception	0.631	0.671	0.441
Ablation	Source + Reported Social Media	0.498	0.520	0.270
Ablation	Source + Economic + Social Media	0.632	0.656	0.441
Sensitivity (n=205)	Pooled Explicit (Source + Econ)	0.628	0.660	0.473
Sensitivity (n=180)	Meta-Only Explicit (Econ Only)	0.656	0.656	0.505

Why This Surprised Me in a Social Media Analytics Class

"Magkaiba ang Source of Information at Actual Influence."

Just because somebody says they saw kamote information on Facebook, YouTube, or TikTok doesn't mean social media persuaded them. Reporting an information channel is not the same as demonstrating social learning or behavior change—which requires measuring what message they saw, how they interpreted it, whether they trusted it, and whether they acted (Bandura's Social Cognitive Theory).

Channel ≠ Persuasion

Synthesis & Roadmap • Discussion, Streaming Analytics & Closing (Sections V–VIII)

From Personal Food Choices to Project TIWALA Streaming Analytics

Connecting our October 2026 rice-import story, the 5% dietary shift simulator, and our 357-respondent social sensing study across Three Connected Levels—and charting how today's batch NLP workflow evolves into real-time Streaming Analytics.

Macro Level (National Policy & Supply)

The National Picture

Rice import dependence (100K–200K MT US rice MOU), agricultural competitiveness, flood-control resource scrutiny, and building resilient local root-crop value chains for Filipino farmers.

Micro Level (Household Kusina)

The Household Kusina

Price, convenience, daily rice habits, and the practical decision of what to serve. Even if a food innovation makes sense on paper, it won't succeed if families find it too expensive (50.9% vs. 78.7% willingness).

Computational Level (Digital Sensing)

The Digital Sensing Loop

Using human-guided Filipino/Taglish NLP, interpretable ML, and future Streaming Analytics to listen to communities, detect emerging barriers early, and test what works.

Future Direction • Real-Time Sensing (Section VI)

From Batch Survey Sensing to Project TIWALA Streaming Analytics

Today, we built an auditable batch analytical workflow. Tomorrow, under Project TIWALA, consented survey feeds and platform-permitted comments can flow into a continuous Streaming Analytics architecture (Bifet et al., 2010; Gama et al., 2014):

Early Detection of Community Concerns (Concept Drift)

Instead of waiting months for another survey, detect right away when people start talking more frequently about kamote price spikes, preparation time, or local supply shortages.

Applications in Business & Public Governance

In business, faster learning helps organizations respond to shifting customer needs. In governance, timely community feedback identifies pain points that deserve faster public-service action.

Guardrail: "Faster Analytics Without Rigor = Faster Mistakes"

Speed alone isn't enough. Streaming pipelines still require RA 10173 Data Privacy compliance, consented data, validated Taglish rules, and ongoing human-in-the-loop review.

Section VII • Methodological Rigor 7 Scope Boundaries

What Our Paper Claims—And What It Does Not Claim

1. Self-Selected Sample: 357 digital responses (322 Meta vs 35 Google); not nationally representative.

2. Stated Willingness: Measures stated readiness in Q5, not observed kitchen consumption or import reduction.

3. Q5 Health Wording: Q5 mentioned low-glycemic diabetes benefits, which may inflate baseline willingness.

4. Binary Simplification: Pooled conditional + unwilling (n=73) vs willing (n=223); excluded 61 cases.

5. Single-Researcher Audit: Structured human-guided adjudication, not a formal multi-coder kappa study.

6. Internal CV Only: 5×5 repeated CV checks stability; external validation & full DOI/SCT scales are future work.

Appendices A, B, C & D • Open Science

Reproducible Colab Workflow, Data Dictionary & AI Disclosure

Appendix A • Google Colab 10-Stage Python Notebook

Appendix D • Google Drive Data Dictionary & Files

Appendix B (AI Assistance Disclosure): Generative AI assisted with academic phrasing and debugging Python/Colab scripts; the researcher retained full responsibility for research design, annotation judgments, and verified outputs.

Closing • Key Takeaway

"Hindi kailangang talikuran ang kanin para bigyan ng mas malaking papel ang kamote."

Before we launch campaigns, create recipes, or try to persuade people through social media, we first need to understand them. Using survey-based social sensing, human-validated NLP, and interpretable machine learning, we found that economic perceptions strongly shape adoption readiness—and that reporting a social-media channel is not the same as persuasion.

First, Listen. Survey & social sensing

Then, Understand. Human-validated Taglish NLP

Then, Test What Works. Project TIWALA streaming loop

Project TIWALA Survey Instrument

Kanin + Kamote: Pananaliksik sa Kamote na Katuwang ng Bigas (Q1–Q9)

Core Questionnaire Items Analyzed in the Paper:

Tanong 1 & 2 (Context): Lalawigan/probinsya at pangunahing karanasan sa kamote (Mamimili/kumakain vs. Magsasaka/nagtatanim).

Tanong 3 (Q3 → Economic Direction): "Bakit niyo po nasabing nakakatipid (o hindi) kapag isinasalit ang kamote sa pangaraw-araw na pagkain?" (Held-out Acc: 0.850).

Tanong 4 (Q4 → Stigma & Rice Habit): "May kasabihan na ang kamote ay 'pangkain ng mahirap.' Ano po ang pananaw ninyo rito?"

Tanong 5 (Q5 → Provisional Outcome Variable): "Dahil napatunayan na ang kamote ay mabisang panlaban sa diabetes (low-glycemic), handa po ba kayong gawin itong regular na katuwang ng kanin?" (223 Willing vs. 73 Conditional/Unwilling).

Tanong 6, 7 & 8: Preferred forms (harina/pandesal, pansit, chips, sariwa), farming resilience vs. palay, and adoption motivators.

Tanong 9 (Q9 → Reported Information Channels): "Saan ninyo madalas nakikita o naririnig ang mga impormasyon tungkol sa benepisyo sa kalusugan o tamang pagtatanim ng kamote?" (Social media, health center/DA/LGU, kapitbahay/kapwa magsasaka).