Video and slides synchronized, mp3 and slide download available at URL http://bit.ly/2y3TyB6. Sally Langford talks about the use of ML within StitchFix’s inventory forecasting system, the architecture they have developed in-house, their use of Bayesian methods, and how this technique can be applied to similar examples of beta binomial regressions. Filmed at qconnewyork.com. Sally Langford is a data scientist on the Merch Algorithms team at Stitch Fix, which is responsible for developing tools and methods for inventory optimization and product development.

1. Automating Inventory at Stitch Fix
Using Beta Binomial Regression for Cold Start Problems
Sally Langford - Data Scientist

2. InfoQ.com: News & Community Site
• 750,000 unique visitors/month
• Published in 4 languages (English, Chinese, Japanese and Brazilian
Portuguese)
• Post content from our QCon conferences
• News 15-20 / week
• Articles 3-4 / week
• Presentations (videos) 12-15 / week
• Interviews 2-3 / week
• Books 1 / month
Watch the video with slide
synchronization on InfoQ.com!
https://www.infoq.com/presentations/
stitch-fix-ml

3. Presented at QCon New York
www.qconnewyork.com
Purpose of QCon
- to empower software development by facilitating the spread of
knowledge and innovation
Strategy
- practitioner-driven conference designed for YOU: influencers of
change and innovation in your teams
- speakers and topics driving the evolution and innovation
- connecting and catalyzing the influencers and innovators
Highlights
- attended by more than 12,000 delegates since 2007
- held in 9 cities worldwide

4. How Stitch Fix works:

5. - Tell us about your style, fit and price preferences.
How Stitch Fix works:

6. - Tell us about your style, fit and price preferences.
- A personal stylist will curate five pieces for you.
How Stitch Fix works:

7. - Tell us about your style, fit and price preferences.
- A personal stylist will curate five pieces for you.
- Try all the items on at home.
How Stitch Fix works:

8. - Tell us about your style, fit and price preferences.
- A personal stylist will curate five pieces for you.
- Try all the items on at home.
- Give your stylist feedback on all items, then only pay for what you keep.
How Stitch Fix works:

9. - Tell us about your style, fit and price preferences.
- A personal stylist will curate five pieces for you.
- Try all the items on at home.
- Give your stylist feedback on all items, then only pay for what you keep.
- Return the other items in envelope provided.
How Stitch Fix works:

10. Benefits of Machine Learning in Inventory Management:
- Scalable with business.
- Rapid reforecasting.
- Capture nonlinear relationships.
- Cold start problems.

11. time
number of
units
order arrives
in warehouse
shirt is sent to
clients and is sold
plaid shirt

12. time
order arrives
in warehouse
shirt is sent to
clients and is sold
plaid shirt
number of
units

13. Inventory consumption of a style is proportional to;
- daily demand,
- clients for which the style is recommended,
- whether there are units in the warehouse,
- probability a stylist chooses to send the client this style,
- if the client buys the style.

14. Inventory consumption of a style is proportional to;
- daily demand,
- clients for which the style is recommended,
- whether there are units in the warehouse,
- probability a stylist chooses to send the client this style,
- if the client buys the style.

15. Inventory consumption of a style is proportional to;
- daily demand,
- clients for which the style is recommended,
- whether there are units in the warehouse,
- probability a stylist chooses to send the client this style,
- if the client buys the style.

16. Ranked styles recommended for client - which will the stylist choose to send?

17. Ranked styles recommended for client - which will the stylist choose to send?

18. Ranked styles recommended for client - which will the stylist choose to send?

19. Ranked styles recommended for client - which will the stylist choose to send?

20. Ranked styles recommended for client - which will the stylist choose to send?

21. plaid long-sleeve shirt

22. ?
plaid long-sleeve shirt

23. plaid long-sleeve shirt

24. plaid long-sleeve shirt

25. ...
plaid long-sleeve shirt

26. ...
P(chosen) + P(not chosen) = 1
plaid long-sleeve shirt

27. ?
...
plaid long-sleeve shirt
blue long-sleeve shirt

29. Prior Beliefs

30. Prior Beliefs
Evidence

31. Prior Beliefs
Posterior Beliefs
Evidence

32. N ~ Binom(Nav
, p)

33. N ~ Binom(Nav
, p)

34. N ~ Binom(Nav
, p)
p ~ B(α, β)

35. B(α’, β’) = B(α0
+ k, β0
+ n - k)

36. B(α’, β’) = B(α0
+ k, β0
+ n - k)

37. Step 1: Use maximum likelihood to calculate α0
and β0
for the distribution of p in
groups of similar styles.
Step 2: After a period of time, update this prior for the number of times the new
style has been recommended for a client (n), and chosen to be sent (k).
Step 3: Calculate the mean and confidence interval of p from the resulting
distribution. This is used as the probability that the new style will be chosen to be
sent to a client.
Step 4: Repeat steps 2-3.

38. VGAM (in python):
import rpy2.robjects as robjects
robjects.r.library("VGAM")
robjects.r("fit = vglm(cbind(successData,trialData - successData) ~ 1,
betabinomialff, trace=TRUE)")
alpha, beta = robjects.r("Coef(fit)")
import scipy
fit = scipy.stats.beta.fit(data, floc=0, fscale=1)
alpha, beta = fit[0], fit[1]
---
result = scipy.optimize.minimize(loss_function, p0, jac=True, **kwargs)

39. Data Storage
Job Scheduler
SQL Engine
SQL EngineHive Metastore
Flotilla: Auto scaling cluster
Job server for
Spark cluster
Data Scientist Code

41. Top 5 recommendations for client D
Top 5 recommendations for client C
Top 5 recommendations for client B
Top 5 recommendations for client A
Top 5 recommendations for client E

42. B(α , β ) = B(μ/σ, (1 - μ)/σ)
μ
σ

43. μ = μ0
+ μn
log(1 + n)

44. number of
units
time
now planned orders
new styles

45. number of
units
time
now
forecasted units
planned orders
new styles

46. How do we use our inventory forecast model?
- When should we re-order inventory?
- How should we buy inventory by size?
- How should orders be separated into different warehouses?
- When should a style not be sent out anymore, in place of a new option?

47. Metrics of success:
- Fraction of inventory out with clients compared to in the warehouse?
- How many styles are available to send to a client?
- ∆ in the beginning of month projected units.
- Cumulative units sold over time.

48. Do you want to calculate the probability of success in a binomial process?
Not enough data?
Use Beta Binomial Regression for your cold start problem!

49. slangford@stitchfix.com
Stitch Fix
Algorithms Blog
Algorithms Tour
@stitchfix_algo

50. Watch the video with slide
synchronization on InfoQ.com!
https://www.infoq.com/presentations/
stitch-fix-ml