1 Introduction
2 Churn prediction in the literature
3 Proposed solutions to the SPA diffusion model
3.1 Basic reasoning toward the proposed model
3.2 Model for determining initial values using sociometric clique theory
3.2.1 Proposed model for determining initial values
3.3 SSA–SPA diffusion model
3.3.1 Influence function
3.3.2 Social status
3.3.3 Social status function
3.3.4 Number of iterations and energy threshold
3.4 Diffusion model prediction scheme
4 Materials and methods
4.1 Data
4.2 Evaluation procedure
4.3 Experiment
4.3.1 Phase 1: finding the optimal threshold and optimal number of iteration steps
\(K_{opt}\)
|
\(T_{opt}\)
| best F-m | |
---|---|---|---|
orig. init. + SPA | 2 | 0.16 | 0.182 |
clq init. + SPA | 2 | 0.02 | 0.293 |
orig. init. + SSA–SPA | 10 | 0.32 | 0.255 |
clq init. + SSA–SPA | 3 | 0.04 | 0.372 |
4.3.2 Sensitivity analysis of the results
4.3.3 Phase 2: evaluation of the diffusion model on neighbours of past churners
\(K\)
|
\(T\)
| tn | fn | fp | tp | prec. | recall | F-m | |
---|---|---|---|---|---|---|---|---|---|
orig. init. + SPA | 2 | 0.16 | 7381 | 61 | 2470 | 46 | 0.018 | 0.43 | 0.035 |
clq init. + SPA | 2 | 0.02 | 9541 | 92 | 310 | 15 | 0.046 | 0.14 | 0.069 |
orig. init. + SSA–SPA | 10 | 0.32 | 9098 | 87 | 753 | 20 | 0.026 | 0.19 | 0.045 |
clq init. + SSA–SPA | 3 | 0.04 | 9436 | 86 | 415 | 21 | 0.048 | 0.20 | 0.077 |
Used parameters | Best parameters |
\(\frac{\text {Used F-m}}{\text {Best F-m}}\,(\%)\)
| |||||
---|---|---|---|---|---|---|---|
\(K\)
|
\(T\)
| F-m |
\(K_{opt}\)
|
\(T_{opt}\)
| F-m | ||
orig. init. + SPA | 2 | 0.16 | 0.035 | 5 | 0.37 | 0.069 | 51 |
clq init. + SPA | 2 | 0.02 | 0.069 | 5 | 0.01 | 0.098 | 70 |
orig. init. + SSA–SPA | 10 | 0.32 | 0.045 | 3 | 0.46 | 0.056 | 80 |
clq init. + SSA–SPA | 3 | 0.04 | 0.077 | 9 | 0.13 | 0.087 | 89 |