2. BUSINESSCASE(-S)
• Question A: how big is a probability many products of a given type fail (help to
estimate the product cost including service cost)
2
3. BUSINESSCASE(-S)
• Question A: how big is a probability many products of a given type fail (help to
estimate the product cost including service cost)
• Question B: how to compare products/groups/brands (help to provide the right
assortment)
3
4. BUSINESSCASE(-S)
• Question A: how big is a probability many products of a given type fail (help to
estimate the product cost including service cost)
• Question B: how to compare products/groups/brands (help to provide the right
assortment)
• Question C: how many service cases we get every month/week/day (stuffing
for our service center)
4
5. BUSINESSCASE(-S)
• Question A: how big is a probability many products of a given type fail (help to
estimate the product cost including service cost)
• Question B: how to compare products/groups/brands (help to provide the right
assortment)
• Question C: how many service cases we get every month/week/day (stuffing
for our service center)
5
Observations previously
reported:
orders and services
within a given period
6. BUSINESSCASE(-S)
WE CONCENTRATED ON:
• Question A: how big is a probability many products of a given type fail (help to
estimate the product cost including service cost)
• Question B: how to compare products/groups/brands (help to provide the right
assortment)
• Question C: how many service cases we get every month/week/day (stuffing
for our service center)
6
7. BUSINESSCASE(-S)
WE CONCENTRATED ON:
• Question A: how big is a probability many products of a given type fail (help to
estimate the product cost including service cost)
• Question B: how to compare products/groups/brands (help to provide the right
assortment)
7
Might need a
product/group
characteristics
8. BUSINESSCASE(-S)
WE CONCENTRATED ON:
• Question A: how big is a probability many products of a given type fail (help to
estimate the product cost including service cost)
• Question B: how to compare products/groups/brands (help to provide the right
assortment)
8
Might need a
product/group
characteristics
9. BUSINESSCASE(-S)
WE CONCENTRATED ON:
• Question A: how big is a probability many products of a given type fail (help to
estimate the product cost including service cost)
• Question B: how to compare products/groups/brands (help to provide the right
assortment)
9
Might need a
product/group
characteristics
19. • Failure rate by period
• Cumulativa failure rate
• Other transformations (creating failure rate categories, .log-transform, other)
• Important milestones: service ratio after 2, 3, 5 years
• Level of granularity (ask stackholders more)
19
WHAT TOPREDICT
21. PRODUCTDATA
21
Categorical data
- Encode
- Embed
Continuous data
- Use as is
- Regroup to create categories
- Create extra features
(e.g., min/max/avg of
guarantee per chain, to overcome
bad data quality)
Text data:
- Use as is in CNNs
- Can create new text data
(example: ‘AenergiklasseVa+++’)
- Can create continuous variables: tf-idf + svd/nmf; ...
- Can create categorical variables: topics; ...
23. TIMESERIESPREDICTION
OUR PRACTICE
Something that worked nicely for us:
- Floating ‘time’ between the observations and what we predict
-> grows sample size
- Use possible positional and date/time features
-> differentiate between different samples
-> allows to use temporal CNNs if wanted
-> if sequence observations have a logical ‘start’,
like zero-age for Failure Rate, more positional
features can be created
23
24. SIMPLEMESSAGES
ESTABLISH THE BASELINE – TIMESERIES WITHOUT PRODUCT DATA IS STILL USEFUL
Example after half a year of observations:
await a significant drop of services after the first half-a year
24
25. SIMPLEMESSAGES
PREDICT WITH CONFIDENCE 😎
- Quantile estimations (quantile estimates also have a mean error 😱)
- BNNs (Monte Carlo dropout, VAE, inheretent noise) - Uber recommends 👍
- Ensembeling by retraining – the simplest way
- ...
25
28. SIMPLEMESSAGES
CHOOSE ROBUST MODELS
Try simple versions of everything
- RNN, CNN, Temporal CNN with/without Gating
- Choose good validation loss
- Plan for future ensemble predictions
28
29. SIMPLEMESSAGES
CHOOSE ROBUST MODELS
Try simple versions of everything
- RNN, CNN, Temporal CNN with/without Gating
- Choose good validation loss
- Plan for future ensemble predictions if possible
29
30. SIMPLEMESSAGES
CHOOSE ROBUST MODELS
Try simple versions of everything
- RNN, CNN, Temporal CNN with/without Gating
- Choose good validation loss
- Plan for future ensemble predictions if possible
30
31. SIMPLEMESSAGES
What simplifies our life (IN ANY PROJECT) :
- Clear priorities from the departments, their support
- Snowflake as DW + connector to PowerBI + connector to Databricks
- Common practice (in DEV)
31