NutriRecIndia19M

A large-scale synthetic dataset for food recommendation, dietary compliance, and health-outcome modeling grounded in Indian population statistics.

NutriRecIndia19M is a multi-table synthetic dataset (~19M rows across 10 linked tables) simulating one year of eating behavior, food recommendations, and health outcomes for 5,000 synthetic Indian users. It was built to support research and prototyping in recommender systems, health-behavior modeling, and nutrition-focused ML — in a domain where real user-level eating and health data is sensitive and rarely shareable.

Every distribution used to generate this dataset (demographics, meal timing, cuisine preference, fasting behavior, compliance rates, app-interaction patterns) is grounded in a cited, published source — census data, national health surveys, and public industry benchmarks — rather than arbitrary parameters. Full source list is in Data Source & Grounding.

Dataset Summary

Table	Approx. Rows	Description
users	5,000	Demographic, cultural, and health profile per user
meal_logs	~4.6M	Every logged meal: timing, dish, nutrition, occasion
interactions	~7.3M	Recommendation impressions: skip / click / order events
reorder_events	~2.2M	Repeat-order behavior per user–dish pair
social_eating_context	~4.6M	Social setting (alone / family / friends / restaurant) per meal
user_weekly_context	~41,600	Weekly nutritional gaps, budget state, compliance rate
fast_days	~7,900	Religious fasting events (Ramadan, Ekadashi, Paryushan, etc.)
skip_events	~11,000	Meal-skip events with reason and compensatory behavior
health_outcomes	~4,000	Quarterly BMI change and dietary-compliance trend per user
life_events	~640	Job change, marriage, pregnancy, diagnosis, relocation, etc.

Total: ~19 million rows, all joinable via a shared user_id key.

Intended Uses

• Training and benchmarking recommendation models (ranking, cold-start, session-based) in a food/health context
• Dietary-compliance and health-outcome prediction research
• Studying position bias, reorder/habit modeling, and social-context effects on eating behavior
• Teaching/coursework on recommender systems where real health data cannot be used
• Prototyping India-specific nutrition or food-delivery ML products before real user data is available

Out-of-scope uses: This dataset should not be used to make real clinical, medical, or individual health decisions. It is synthetic and does not represent any real person's data.

Dataset Structure

users.csv

Demographic and health profile: age, gender, religion, region/state, BMI, health conditions (diabetes, hypertension, PCOS, obesity), dietary restrictions (vegetarian, halal, Jain, gluten-free, dairy-free, low-sodium), occupation, income tier, health literacy, cooking skill, observance level, sleep hours, persona type. 1 row per user (5,000 rows, ~40 columns).

meal_logs.csv

One row per logged meal: user_id, timestamp, occasion (breakfast/lunch/dinner/snack/late-night), dish name, cuisine, calories, macros (protein/carbs/fat/fiber), glycemic index, portion multiplier, cooked-at-home vs. delivery, social context flag, health-compliance flag.

interactions.csv

Recommendation-impression log: user_id, session ID, dish shown, recommendation rank, action (skip/click/order), session duration, user_health_match, price_match_score, cuisine_affinity, final order flag.

reorder_events.csv

Per user–dish repeat-order history: days between orders, cumulative order count, reorder flag, rating proxy.

social_eating_context.csv

Per-meal social setting: solo / family / friends / colleagues / spouse / restaurant, group size, location type, budget multiplier, variety score.

user_weekly_context.csv

Weekly rollup per user: average calories, macro/fiber gaps vs. RDA targets, budget state (month-start vs. month-end), weekly health-compliance rate, stress level.

fast_days.csv

Religious fasting-day records: fast type (Ramadan, Ekadashi, Monday fast, Navratri, Paryushan, Jain monthly), observance level, calorie impact, complete vs. partial fast.

skip_events.csv

Meal-skip records: occasion skipped, reason (not hungry, running late, meeting, fasting, forgot), whether a compensatory meal followed and its calorie delta.

health_outcomes.csv

Quarterly per-user health trend: BMI at quarter start/end, BMI change, compliance-rate change, condition-severity change, overall health trend (improving/stable/declining).

life_events.csv

Life events per user: type (job change, marriage, pregnancy, health diagnosis, city relocation, started gym, financial stress), event date, and (where implemented — see Known Limitations) downstream behavioral effects.

Data Source & Grounding

Rather than choosing generator parameters arbitrarily, every major distribution in this dataset was calibrated against a published source, for example:

• Census of India (2011) — age, gender, religion, and regional population distributions
• NFHS-5 (National Family Health Survey, 2021) — BMI-by-age/gender, chronic condition prevalence, vegetarian rate, dietary compliance rates
• ICMR "What India Eats" (2021) — meal-occasion timing, regional meal frequency, dish/cuisine pool
• MOSPI Household Consumer Expenditure Survey (2022) and RBI Household Finance Survey (2022) — income tiers and monthly budget cycles
• Pew Research Center (2021) — religiosity and fasting-frequency correlation
• Zomato Annual Report (2023) and Swiggy Engineering Blog (2022–2023) — recommendation click-through/order-conversion rates, reorder rates, session sizes
• ADA Standards of Medical Care in Diabetes (2023) and Appetite journal (2020, India study) — glycemic-index relevance and stress as a dietary-compliance predictor
• Academic recommender-systems literature (Covington et al. 2016; Joachims et al. 2007; Jannach et al. 2015; Lam et al. 2008) — used only to sanity-check that resulting model metrics (NDCG, AUC, cold-start accuracy) fall in realistic, published ranges, not to shape the raw data itself

A full bibliography of the ~20 sources used is available on request / in the accompanying methodology notes.

Why synthetic

Real, individually identifiable Indian eating and health-condition data cannot be ethically or legally shared at this granularity. Synthetic data calibrated against public aggregate statistics is an established substitute used elsewhere in industry (e.g., synthetic patient records for early-stage medical AI, synthetic listening/purchase sessions for recommender cold-start testing). This dataset follows the same pattern for the Indian food/nutrition domain.

Validation & Data Quality

The dataset was validated with an automated suite (150+ rule-based checks) covering biological plausibility, cultural/religious correctness, temporal consistency, and statistical alignment with the source distributions above. As of the latest generation pass, 151 checks pass.

Passing checks include (non-exhaustive):

• Vegetarian %, religion mix, and BMI distributions matching NFHS-5/Census targets within a few percentage points
• BMI↔diabetes and BMI↔hypertension odds ratios in the expected clinical range
• Meal-occasion timing and regional breakfast/lunch/dinner timing offsets matching ICMR patterns
• Fasting-frequency correlation with religious observance (r ≈ 0.73, consistent with Pew Research findings)
• Recommendation position-bias lift, skip/click/order ratios, and reorder rates within published food-delivery benchmarks
• No cross-religious dietary violations (e.g., no pork for Muslim users, no non-veg on Hindu fast days for the vast majority of records)

Known Limitations

This dataset is shipped with its validation gaps disclosed rather than hidden:

• Life-event effects are only partially propagated. Events such as starting a gym routine, financial stress, or a new health diagnosis do not yet consistently shift downstream protein intake, budget, or compliance in meal_logs / user_weekly_context. Treat life_events.csv as largely independent of the meal-behavior tables for now.
• A small number of dietary-restriction violations remain, e.g., a minority of vegetarian users have logged meals (~0.2–0.3% of meal_logs rows) containing a non-vegetarian dish, and small counts of no-dairy/no-gluten violations. These are being tightened in the restriction-filtering logic.
• Minor cross-cuisine leakage in the Northeast, mainly Tripura, where Bengali-cuisine dishes are somewhat over-represented (~37% of meals) versus the intended Assamese/Odia/Northeast cuisine pool. Historically not entirely implausible (large Bengali-origin population in Tripura) but flagged as higher than intended.
• Some meal-ordering-within-day inconsistencies exist (a small number of days where dinner precedes lunch in timestamp order).
• Health-outcome BMI-change values were highly uniform in earlier generator versions (near-constant per quarter); this has been substantially improved in later versions (std increased from 0.00 to 0.24) but should still be treated as a simplified physiological model, not a clinical-grade trajectory simulator.
• compliance_score and compliance_improvement were derived from a simple deterministic formula in earlier versions; later versions add real variance, but users doing rigorous modeling should inspect these fields before treating them as fully independent signals.
• A location-assignment bug causes a subset of non-hostel-living users to be tagged with location_type = hostel in social_eating_context.csv.
• user_health_match and cuisine_affinity in interactions.csv are close to random-uniform rather than fully computed from actual user–dish nutritional/preference matching, which caps ranking-model performance (NDCG@10 ≈ 0.48) below what a fully-featured version would achieve.
• Family medical history is not yet correlated with a user's own condition prevalence in users.csv — this remains close to independent random sampling.

None of these are hidden — they're listed so downstream users can decide which tables/fields are safe for their use case and which need caution or their own re-derivation.

Considerations for Using the Data

Synthetic, not real people. All 5,000 users, their conditions, meals, and outcomes are generated, not observed. No real individual's health or eating data is present anywhere in this dataset.

Not for clinical use. Health-outcome and compliance fields are simplified statistical models, not validated medical predictions. Do not use this dataset to inform real dietary or medical decisions for real individuals.

Representativeness. The dataset is grounded in national-level Indian statistics but reflects a health-app-using, largely urban population sampling rather than the full diversity of the Indian population (e.g., rural populations, very low-income groups, and some smaller regional/religious communities are likely under-represented).

Cuisine/region caveats. As noted above, some Northeast Indian states currently show cuisine-assignment skew that does not fully reflect real regional cuisine diversity; users doing region-specific cuisine research should treat those subsets with extra caution or filter them out.

Licensing & Citation

This dataset is released under the CC BY 4.0 license (confirm/update before publishing if a different license is intended).

If you use this dataset, please cite it as:

NutriRecIndia19M: A Synthetic Multi-Table Dataset for Indian Food Recommendation
and Health-Behavior Modeling. [Darsh Vithlani /2026].