Armaan update
Siddharth and Harsha
Conv Filter
Call Features
- call status
- duration
- date
- start time(hr)
- gest_stage (10 dims)
Gradient CAM
Assign importance to input features based on activations from Conv layer.
X: Features, Y: Call No
Engagement
Non Engagement
Shap
SHapley Additive exPlanations
Typical for non-engagement
Call Features
- call status
- duration
- date
- start time(hr)
X: Features, Y: Call No
Shap (Static)
Income >44
Attention Weights
Most attention on last call for >62.5% examples
Example Attentions
Label1
Label0
X: Call index, Y: Data index
Program Engagement
(Current Formulation)
- Input:
- Call patterns for the first 4 weeks
- Static Features
- Output
- 1 if >50% of rest of the calls received are engaged.
| Model | Accuracy |
|---|---|
| Random Forest | 73% |
| RNN-Model | 80% |
Next up...
- Better definitions for program engagement
- Defining re-engagement
- Interpreting discrete features with large ranges: ngo_hosp_id, gest_age
Features
Static Features
- enroll_gest_age
- enroll_delivery_status
- ngo_hosp_id
- age
- language
- education
- phone owner
- income bracket
- call slot
- calls so far recieved, engaged
Call Features
- call status
- duration
- date
- start time(hr)
- gest_stage
P1: Short-term engagement
- Input:
- Call Data for the past 2 weeks
- Static features
- Output:
- 1 if at least one call is engaged in the next 2 weeks(0 otherwise)
- Class imbalance -> 2:3
RNN-Model
Attn-Model
Conv-Model
Window size: 3
Results
| Model | F1_1 | F1_0 | Accuracy |
|---|---|---|---|
| Random Forest | 75.2 | 61.4 | 70.2 |
| RNN-Model | 88.3 | 77.8 | 84.6 |
| Attn-Model | 89.3 | 79.7 | 85.6 |
| Conv-Model | 94.2 | 76.8 | 87.2 |