1. WOW World of Walkover-weight “My God, it’s full of cows!” (David Bowman, 2001)

2. Can walkover-weight suggest a cow needs attention?

3. Join with breeding information …

4. Position at the outset …  Obstacle: No health information!!!  Suggested: Milking order (i.e. where a cow is in the herd/line-up) is hierarchical and affected by health issues  Proposed goal: to predict a drop in milking order using WOW and other facts

5. Assumptions … deck of cards  Same cows come in for milking each time  Cows are well-behaved (e.g. arrive in a nice queue)  Data is in good shape (e.g. one reading per cow per milking)

6. Data problems  Multiple entries for cows (e.g. four entries for 22719193 in QBH2005)  Delete duplicate weights (SQL problem?)  Cow skipped and recycled back into order  Use average if more than one value

7. About a quarter of the data are zeroes … instances 0 weights λ weights BBYG2006182,93557,288815 BBYG2007206,54539,7261,193 JJVX20077,850073 QBH20057,850073 QBH2006324,36580,36272 QBH2007222,30067,1092,118 QBH200848,53410,535224

8. “zero” problems  Differentiate between a missing cow, a missing weight and a “zero” weight  Ignore missing cows  Cow skipped and recycled back into order  Time-based interpolation  Can be problematic if cow has been missing for a while  Add flag to indicate weight was “guessed”

9. other issues in data preparation  Change milking date to milk index  Change birthdate to age in months  Change parturition date to days since last calved  Additional derivatives  milking index - cow’s position in milk order  ∆-index – change in index for a cow over various time periods (1, 3 and 7 days)  mu-weight – average weight over varying-length periods (3, 7, 14, 21 and 28 milkings)  ∆-mu-weight – change in index for a cow (1, 3, and 7 days)

10. Does [change in] milk order correlate to WOW?

11. Correlation coefficients QBH2006 (dense)  WOW to index == 0.12  WOW to 14-day mu-weight == 0.93  Index to 10-day mu-weight == 0.14  3-day ∆-order to ∆-weight == 0.045

12. ∆∆ 3-day ∆-order and 3-day ∆-weight

13. Predict change in milking order  Use M5P to predict how the milking order will change for a cow at the next milking  Approx. 205,000 QBH2006 samples (with fewer than 5/25 missing attributes)  2/3 training 1/3 testing

14. Re-running took too long … but … you’ve all seen it before, where accuracy was 51.89% (discrimination 0.527) and the model tree was hugely ugly (65 nodes, 33 leaves). Also tried predicting cow’s index as decile and as ratio to herdsize.

15. Cow’s position (index) as ratio to herdsize

16. Cow index vs. herd size

17. Where to? ….  Data must still be scrubbed so that milking order makes sense (if milking order is going to be relevant)  Perhaps cow order needs to be described in completely different terms (e.g. cow buddies)  Easy visualization of herds/cows/breeds/dates/trends is needed this segued into another area of the project.. this segued into another area of the project..

18. Visualization tools (alpha and beta)

20. In the meantime … health data is obtained …

21. Can WOW predict onset of illness?  Combine original attributes and derivatives with health judgments  Cows with unknown health are considered healthy  Need equal number of positive and negative instances

22. Health data becomes available farmyear Qty > 50 BBYG200673 BBYG200795 BBYG2008220 QBH2005113 QBH2006282 QBH2007481 QBH2008253

23. Not so much health data  1613 recorded instances of health  913 different cows with health info  2540 cows with milking info  788 milked cows with health data  7 broad categories of illness:  Calving disorder  Metabolic disorder  Udder disorder (only one with >50 in herd)  Reproductive disorder  Lameness  Infectious diseases  Other ailments

24. Data sparseness QBH2006  75 instances out of 324,291 have health  63 udder disorder  10 metabolic disorder  2 lameness  Only.002% positives → will never be isolated → must subsample negatives  Random selection of 75 negatives → data sparseness → over-fitting likely

25. Data sparseness QBH2006  36 cows have illness at some time, so just learn those?  11,966 records for those cows, 76 of which have illness (still <1% positive)  Random selection of 1% as negatives (about 120)

27. Refinements to approach QBH2006  Restrict target objective to UDDER DISORDER  Randomly select equal number of negatives from cows who have health problem at some point goal: differentiate between healthy and unhealthy state

31. Detecting mastitis amidst random normal cows QBH2006  Restrict learning objective to UDDER DISORDER  Randomly select equal number of negatives from all cows that have been milked (63+,63-)

33. When is a cow sick?  So far, attempted to predict health label at point of milking, but..  … when was the health label attached? before, during or after the current milking?  Goal: predict whether cow needs attention at the next milking (i.e. time series)

36. === Summary === Correctly Classified Instances 90 70.3125 % Incorrectly Classified Instances 38 29.6875 % Kappa statistic 0.4026 Mean absolute error 0.3446 Root mean squared error 0.4532 Relative absolute error 68.8933 % Root relative squared error 90.5974 % Total Number of Instances 128 === Detailed Accuracy By Class === TP Rate FP Rate Precision Recall F-Measure ROC Area Class 0.508 0.108 0.821 0.508 0.627 0.707 UDDER DISORDER 0.892 0.492 0.652 0.892 0.753 0.707 NONE === Confusion Matrix === a b <-- classified as 32 31 | a = UDDER DISORDER 7 58 | b = NONE

38. Agenda  Replace quantified attributes with simpler (e.g. boolean, nominal) ones  Characterise exceptions  Below average weight for cow/herd/breed/age  Dropped decile/>50 in order  Broad statistical measures  How many std.devs. from mean  z-score (probability of variation)  Choose negative instances more carefully (select fewer interpolates)  Spend more time with people who know cows