Healthy Recommendations - On the Role of Recommender Systems to Promote Health and Exercise
Invited keynote at the 26th International Conference on Case-Based Reasoning, Stockholm, Sweeden
2. After 25 years of recommending
movies, music, books, hotels, gadgets …
… is there more to my recommender-life?
3. Key Message
Exciting opportunities exist to take advantage of new sources
of real-world data (IoT, sensors, wearables, …)
Bringing novel opportunities to have a positive impact on how
we all live, work and play.
Especially health & well-being ⟹ improved quality of life.
4. Overview
1. From Small Sensors to Big Data
The power of the IoT to transform how we live, work, and play.
2. Why Exercise is the Best Medicine
From data-driven healthcare to personalised well-being.
3. A Case-Study in Marathon Running
Applications of recommender systems to support recreational
marathon/endurance runners.
10. how much information?
1 = 10bytes
18
exabyte
20,000 x all of the printed material in the
US Library of Congress.
Or all of the words spoken by humans. Ever!
11. how much information?
1 = 10bytes
18
exabyte
6 !
hours
But, we now create this
much information every
21. Sleep Tracking
Motion-based ‘sleep tracking’.
Duration vs Movement
Sleep Quality (≈ time/move)
Sleep Notes / Wakeup Moods
Comparative Analytics
22. Mood & Focus
The Melon Headband
Uses EEG to track brain
activity to assess ‘focus’.
Tagging, location, and
activity information helps
users to better assess
what impacts their focus.
23. Food & Diet
Meal logging and nutritional
analysis.
Manual vs Semi-Automatic.
Calorie goals and diet plans.
Integrated weight tracking.
24. Heart Rate Sensing
Using smartphone camera with your
finger. No external sensor required.
Detecting colour changes due to
capillary blood-flow.
Tagging, comparative analytics etc
25. Blood Glucose
External blood glucose sensor
automatically syncs readings with
app.
Readings tagged with mealtime,
exercise etc.
Analysis and visualization of
trends, logs, stats.
26. Mobile Spirometry
Using a mobile phone
microphone to evaluate lung
function.
FVC, FEV, PEF measures.
Audio Features Machine Learning.
Mean 5.1% error wrt clinical spirometry suitable for home-
based monitoring.
27. Mobile Spirometry
Using a mobile phone
microphone to evaluate lung
function.
FVC, FEV, PEF measures.
Audio Features Machine Learning.
Mean 5.1% error wrt clinical spirometry suitable for home-
based monitoring.
31. Born to Run
Humans evolved to be the world’s best
endurance runners.
… and quadrupeds cannot gallop long-
distances because they cant stay cool.
gallop ⟹ no panting ⟹ overheating
We can run long distances faster than
the trot-gallop transition speed of
quadrupeds!
32. Exercise as Medicine
Humans evolved to be adapted for
regular, moderate amounts of endurance
physical activity into late age.
⟹ significant disease consequences
associated with prolonged and chronic
inactivity.
33. Lieberman’s ‘Mismatch Diseases’
Obesity, type-2 diabetes, back pain, sleep disorders, …
Diseases primarily caused by lack of adaptation to novel gene-
environment contexts.
Cultural evolution speeding ahead of natural evolution
Diseases that are otherwise, previously rare, largely preventable, and
almost entirely absent from (well-adapted) hunter-gatherer populations.
The Paradox of the Modern World
Reduced mortality (living longer) comes with increased morbidity
(more years spent suffering from chronic preventable diseases).
34. Consequence of Inactivity
Type 2 Diabetes
Globally, 285m suffers growing by 50%+ by 2030 (>440m)
Heart Disease
150 mins exercise/week decreases risk by ~45%
Osteoporosis (~30% of elderly women and ~10% of men)
Physical activity promotes more peak bone-mass better able to cope with inevitable
decline in bone-mass in later life.
Some Types of Cancer
E.g. >2x reduction in breast cancer risk when comparing low to high activity groups
(adjsuted for age, smoking, BMI, alcohol etc.).
36. Benefits of Activity
Aerobics Center Longitudinal Study (>80k patients, 1970-2005)
All-cause mortality ~3x lower, improvers vs unfit, always-fit vs improvers.
Nurses’ Health Study (>111k women x 24 years)
Relative mortality (vs lean, active) increases by ~50% due to inactivity or obesity.
Stanford Runners Study (538 runners vs 423 control x 25 years)
Runners have 20% lower mortality and 50% lower disability scores .
37. Exercise = Best Medicine
Heart Disease
Stroke
Diabetes
High BP
Breast Cancer
Colon Cancer
Alzheimers
Depression
0 15 30 45 60
Consensus Reduction in Risk
Lower Costs
Fewer Side-Effects
As effective as drugs/therapy
+
38. Squaring the Geriatric CurveFunctionalCapacity
Age
High Risk (Inactive) Lifestyle
Low Risk (Active) Lifestyle
39. 0
20
40
60
1990 2000
%NoActivity
Males FemalesExercise Paradox
Humans evolved to be active …
Adapted for regular, moderate amounts
of endurance activity into late age.
While, prioritising rest/inactivity …
Limited food/energy ⟹ avoid unnecessary exertion ⟹ selection ignored
effects of chronic inactivity.
40.
41. A Personalisation Opportunity
Promoting exercise means altering our environments
Need to nudge people ⟹ relevant, personalised, fun, …
Much work to date by CBR/RecSys/UM communities …
Personalised/persuasive eHealth, health profiling for chronic disease
management, diagnosis, treatment planning, recommender systems
for food and diet, personalised exergaming, …
43. The Marathon
26.2 miles / 42.2 kms
Largest mass-participation sporting events in the world.
12-16 weeks of dedicated training (5-8 hrs/week, ~50km/week)
WRs 2.02.57 (Denis Kimetto, 2014), 2.15.25 (Paula Radcliffe, 2003)
~40 year old men (4:10 finish) vs ~36 year old women (4:36 finish)
45. Motive, Means, and Opportunity.
Motivated, yet often inexperienced, users/runners
need for advice/support/guidance.
Plentiful supply of rich data training, rest, races,
nutrition..
Many recommendation tasks/opportunities deliver
real-time recommendations (devices, audio, haptic etc.)
57. 12x400m Intervals 6x800m Intervals 3x1mi Intervals
Long Run (26km) Long Run (28km) Long Run (30km) Long Run (20km)
Tempo Hills Marathon Pace
Recovery Run
Rest Rest Rest Rest
Recovery Run Recovery Run
Easy Run
Rest Rest Rest Rest
Tempo
Recovery Run
Easy Run Easy Run Easy Run Easy Run
Week 9 Week 10 Week 11 Week 12
When and where to run and rest? How far, long, fast?
Type of session (easy, intervals, hills, long, tempo).
60. How to Run a Personal-Best?
Dual prediction/recommendation tasks:
1. Predicting challenging but achievable personal-best times.
2. Recommending a race-plan (pacing plan) to achieve this PB.
61. Race Records & Pacing PlansFasterSlower
RelativePace
10k5k 15k 20k 25k 30k 35k 40k Finish
Average Pace
6 mins/km
Finish
Time
4 hours
13 mins
Getting goal-time predictions or pacing wrong can have
disastrous consequences even for experienced runners …
70. non-PB PB
non- P
no P
Case Base
non-PB PB
Each case encodes a specific PB marathon progression for a given runner.
Large and varied case base captures progressions for runners of all abilities.
71. non-PB PB
non- P
no P
A Case-Based Approach
Target Runner
Recent Race
Case Base
79. Extended Cases
nPB (problem) PB (solution)
Now each case is reflects a complete runner history with a set of n-1 nPB races
paired with a single PB race. But …
How to match/compare runners with different numbers of races?
How should races be ordered in a case for matching?
80. Landmark Races
Certain types of races standout in a runner’s history as
examples of good/bad races:
MR(r)/LR(r) – Most/least recent race in a runner’s history.
MV(r)/LV(r) – Most/least varied (pacing) race.
PPB(r)– Previous PB race.
PW(r)– Personal worst race (slowest).
MnPB(r)– a pseudo-race based on the mean of the runner’s non-PB.
83. Landmark Races as Abstract Features
PBMRLV LRMVPW PPB
nPB (problem)
Each case has a fixed landmark-based representation and the same race can be
represented multiple times under different landmark labels.
Facilitates matching, normalises for different length race histories.
84. Richer Representations
Runner
Least Recent
PB (Solution)
Most Recent Previous PB
Personal Worst
Mean nPB
Many possible combinations of landmark races Multitude of representations.
Problem Description
85. Key Questions
Do richer representations lead to better predictions/
recommendations?
Are some landmark races better than others?
What is the impact of richer representations on runners of
differing abilities?
87. Datasets & Methodology
Berlin (34k), London (36k), New York (44k)
170k runners, >3 races, & 5km pacing (5k, 10k, … 40k, Finish)
Representations
64 x unique case representations (MR + [landmark races]* )
leading to 64 different test systems (prediction & recommendation)
Methodology
10-fold cross-validation (90/10 training/test x 10) using nPB’s as
queries and PBs as ground-truths.
ICCBR/RecSys 2017
Baseline
88. Evaluation Metrics
% Prediction Error
Based on the difference between the predicted PB times and the
actual/ground-truth PB times of the test-cases (that is, the actual PB
times the runners achieved).
Pace Profile Similarity
Relative difference between the recommended pacing-plan and the
actual pacing-plan (that is the actual/ground-truth pacing profiles the
runner achieved in the PB).
90. The Benefit of Richer Representations
In general, richer representations are associated with reduced prediction error, but …
ICCBR/RecSys 2017
Baseline
(Grouping representations by # landmark races)
92. Which Landmark Races Help
the Most?
Compare the average error of representations with/without a given landmark
race to compute relative benefit score for that race.
93. Which Landmark Races Help Most?
Including LV/PW races
disapproves prediction error.
(Benefit < 0)
Including
PPB greatly improves
prediction error
(Benefit ~ 30-40%).
95. Previously …
ICCBR/RecSys 2017, IJCAI 2018
Single-race representations (MR)
Prediction error disapproved with
decreasing runner ability…
E.g. ~6% error @ 4hr time, means
the PB might be off by almost 15
minutes.
97. Prediction Error x Finish-Time
Significant improvements in prediction error across all levels of ability (~3% vs 6-10%)…
Baseline
Best
98. Recommendation Quality x Finish-Time
… without any additional cost in pacing-plan recommendation quality.
99. PB Improvement x Finish-Time
Baseline
Best
Similarly, prediction accuracy is improved across all degrees of PBs.
100. Conclusions (1)
Richer representations better goal-times & pacing-plans.
But not all landmark races are created equally!
The need for >3 races limits applicability to more experienced
marathon runners consider other race types (5k, 10k, …)
Same-city cases mixed-city cases (transfer learning)
5km pacing 1km pacing, cadence, heart rate, power, etc.
103. Conclusions (2)
Unique opportunity for novel application domains,
enabled by the proliferation of data and devices.
Health-related applications offer huge potential because
recommender systems can help to nudge and guide
users towards healthier habits.
104. Conclusions (2)
Unique opportunity for novel application domains,
enabled by the proliferation of data and devices.
Health-related applications offer huge potential because
recommender systems can help to nudge and guide
users towards healthier habits.
Low-hanging fruit exists in a variety of early-adopter
niches (running, cycling, triathalons etc.)
105. More Information
An Analysis of Case Representations for Marathon Race Prediction & Planning,
ICCBR 2018
Marathon Race Planning: A Case-Based Reasoning Approach. IJCAI 2018
Fast Starts and Slow Finishes. Journal of Sports Analytics, 2018
Running with Cases: A CBR Approach to Running Your Best Marathon. ICCBR
2017: 360-374
A Novel Recommender System for Helping Marathoners to Achieve a New
Personal-Best. ACM Recommender Systems 2017: 116-120
Running with Data (medium.com/running-with-data)