Github user myui commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
This kind of behavior could often be happen and Libffm's early stopping
strategy is too aggressive.
```
7 0.43239 0.46952
8 0.42362 0.46999
Github user takuti commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
Make sense as a compromise in terms of memory consumption.
I'll note on documentation to clarify the fact that our `-early_stopping`
option does not return the best of the best
Github user myui commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
```
iter tr_logloss va_logloss
1 0.49738 0.48776
2 0.47383 0.47995
3 0.46366 0.47480
4 0.45561
Github user myui commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
It might be better to reconsider `eta0` when enabling `l2norm` by the
default and by enlarging`max_init_size`. In my experience for FM, init random
size should be small when the avg feature
Github user myui commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
@takuti so then, better to enable l2_norm by the default and
`-disable_l2norm` to disable l2 normalization. My concern is that L2
normalization performed worse for small datasets with
Github user takuti commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
I'll change default options and consider to implement early stopping option
as you suggested.
> What happens without `-l2norm` ?
Once we drop instance-wise L2
Github user myui commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
Also, it's better to revise default `-iters` from 1 to 10 (at least 10
iterations with early stopping).
---
Github user myui commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
BTW, it might be better to implement `early stopping` using validation data.
https://github.com/guestwalk/libffm
We can use a similar approaches to `_validationRatio` used in
Github user myui commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
@takuti Thank you for detailed verification. Let's disable linear term by
the default.
Remove `-disable_wi` and `-enable_wi` (alias `-linear_term` ) to enable
linear term.
Github user takuti commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
### With linear terms
Hivemall
```sql
INSERT OVERWRITE TABLE criteo.ffm_model
SELECT
train_ffm(features, label, '-init_v random -max_init_value 0.5
Github user myui commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
@takuti I advice to check 2-3 updates to investigate how gradient updates
differ.
---
Github user takuti commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
Note: I've extended LIBFFM code so it uses linear terms:
https://github.com/takuti/criteo-ffm/commit/9aca61d93ed8f583025729206ed0dbfd54806a44
However, I cannot observe significant
Github user takuti commented on the issue:
https://github.com/apache/incubator-hivemall/pull/149
Evaluation has been conducted at:
[takuti/criteo-ffm](https://github.com/takuti/criteo-ffm). See the repository
for detail.
As an example, I have used tiny data provided at
13 matches
Mail list logo
