pInstall nonEVC

Get Started. It's Free
or sign up with your email address
pInstall nonEVC by Mind Map: pInstall nonEVC

1. meta data

2. features

2.1. feature in flog

2.2. transform features

2.2.1. learnable_affected_slice

2.2.1.1. idfa trackable data on ios145+

2.2.2. has_ios_idfa_cutover

2.2.2.1. ios145- data + android data

2.2.3. overall_trainable

2.2.3.1. all idfa data, including android and ios

2.2.4. CLICK_ADJUSTED

2.2.4.1. 'time_since_click_days' - 1

2.2.5. DELAY_ADJUSTED

2.2.5.1. 'time_from_click_to_install_days' - 1

2.2.5.1.1. train_delay_tower

2.2.6. label_for_upper_window

2.2.6.1. label > 0 and 'time_from_click_to_install_days' < 7

2.2.7. DELAY_CHECK

2.2.7.1. same as label_for_window

2.2.8. train_conv_tower

2.2.8.1. all negative instance + positive instance within (1, 7)

2.3. label_for_window

2.3.1. label > 0 and 'time_from_click_to_install_days' within (1, 7)

3. towers

3.1. non_serving

3.1.1. use as pConv tower, defined in Namescope("nonserving_tower_logtis"), loss calculated in Namescope("pConv")

3.1.2. output: non_serving_probability

3.1.3. loss: use noserving_logits and output from log_rate to calculate

3.1.4. data: ios145- data + android data

3.2. log_rate

3.2.1. use as delay distribution tower defined in Namescope("log_rate") loss calculated in Namescope("Delay")

3.2.2. output: log_rate

3.2.3. loss: use formula from the theory derivation to calculate

3.2.4. data: ios145- data + android data

3.3. serving

3.3.1. use as serving tower, defined in Namescope("serving_tower_logits") loss calculated in Namescope("serving_tower")

3.3.2. output: serving_tower_pinstall

3.3.3. loss: use output from non_serving and log_rate tower to calculate loss

3.3.4. data: ios145- data + android data

3.4. reservation_calibration

3.4.1. output from serving_tower + user_agent_is_ios embedding

3.4.2. output: combined_reservation_logits

3.4.3. loss: logistic loss

3.4.4. data: is_not_walnut_ad

3.5. transfer

3.5.1. serving tower pctr loss

3.5.2. ios145+ idfa loss

4. utils