We tested the following hyperparameter space:
Parameter | From | To | Steps |
---|---|---|---|
Loss Function | BPR-MAX | TOP1-MAX | - |
Final Activation Function | ELU-0.5 | Linear | - |
Learning Rate | 0.1 0.5 |
0.01 0.1 |
10 5 |
Momentum | 0.00 | 0.90 | 0.10 |
Drop-Out | 0.00 | 0.90 | 0.10 |
Constrained Embedding | True | False | - |
Dataset | Loss Function | Final Activation Function | Learning Rate | Momentum | Drop-Out | Constrained Embedding |
---|---|---|---|---|---|---|
RSC15/4 | BPR-MAX | Linear | 0.09 | 0.0 | 0.1 | True |
RSC15/64 | BPR-MAX | Linear | 0.05 | 0.3 | 0.1 | False |
DIGINETICA | BPR-MAX | ELU-0.5 | 0.02 | 0.5 | 0.1 | True |
DIGINETICA (STAMP) | BPR-MAX | ELU-0.5 | 0.05 | 0.1 | 0.2 | True |
We tested the following hyperparameter space:
Parameter | From | To | Steps |
---|---|---|---|
Number of Epochs | 10 | 30 | 10 |
Decay Rate | 0.0 | 0.9 | 10 |
Initial Learning Rate | 0.001 0.0001 |
0.01 0.001 |
10 10 |
Dataset | Number of Epochs | Decay Rate | Initial Learning Rate |
---|---|---|---|
RSC15/4 | 20 | 0.2 | 0.0002 |
RSC15/64 | 30 | 0.4 | 0.0004 |
DIGINETICA | 10 | 0.3 | 0.0007 |
DIGINETICA (STAMP) | 30 | 0.9 | 0.0004 |
We tested the following hyperparameter space:
Parameter | From | To | Steps |
---|---|---|---|
Learning Rate | 0.1 0.5 |
0.01 0.1 |
10 5 |
Dataset | Learning Rate |
---|---|
RSC15/4 | 0.002 |
RSC15/64 | 0.005 |
DIGINETICA | 0.0007 |
DIGINETICA (STAMP) | 0.002 |
We tested the following hyperparameter space:
Parameter | From | To | Steps |
---|---|---|---|
Learning Rate | 0.01 0.001 |
0.001 0.0001 |
10 5 |
Iterations | 10 | 30 | 10 |
Negative Sampling | True | False | - |
Dataset | Learning Rate | Iterations | Negative Sampling |
---|---|---|---|
RSC15/64 | 0.001 | 10 | False |
We tested the following hyperparameter space:
Parameter | From | To | Steps | Options |
---|---|---|---|---|
Steps | 1 | 20 | 1 | - |
Weighting | - | - | - | Div, Linear, Quadratic, Log, Same |
Dataset | Steps | Weighting |
---|---|---|
RSC15/4 | 2 | Log |
RSC15/64 | 4 | Quadratic |
DIGINETICA | 15 | Div |
DIGINETICA (STAMP) | 8 | Quadratic |
We tested the following hyperparameter space:
Parameter | Options |
---|---|
Number of Neighbors | 50, 100, 500, 1000, 1500 |
Sample Size | 500, 1000, 2500, 5000, 10000 |
Similarity | Cosine, Jaccard |
Dataset | Number of Neighbors | Sample Size | Similarity |
---|---|---|---|
RSC15/4 | 500 | 500 | Jaccard |
RSC15/64 | 500 | 1000 | Cosine |
DIGINETICA | 50 | 500 | Cosine |
DIGINETICA (STAMP) | 100 | 500 | Cosine |
We tested the following hyperparameter space:
Parameter | Options |
---|---|
Number of Neighbors | 50, 100, 500, 1000, 1500 |
Sample Size | 500, 1000, 2500, 5000, 10000 |
Weighting | Same, Div, Linear, Quadratic, Log |
Weighting Score | Same, Div, Linear, Quadratic, Log |
IDF Weighting | False, 1, 2, 5, 10 |
Dataset | Number of Neighbors | Sample Size | Weighting | Weighting Score | IDF_Weighting |
---|---|---|---|---|---|
RSC15/4 | 1000 | 1000 | Log | Quadratic | 5 |
RSC15/64 | 1000 | 5000 | Log | Quadratic | 2 |
DIGINETICA | 500 | 5000 | Quadratic | Div | 10 |
DIGINETICA (STAMP) | 500 | 10000 | Quadratic | Quadratic | 10 |
We tested the following hyperparameter space:
Parameter | Options |
---|---|
Expert | StdExpert, DirichletExpert |
Max Considered Context Length | 5,10,20,30,40,50,75 |
Number of Recent Candidates (Only for Adaptive Configuration) | 5,10,20,30,40,50,75 |
Dataset | Expert | Max Considered Context Length | Number of Recent Candidates |
---|---|---|---|
RSC15/4, RSC15/64, DIGINETICA, DIGINETICA(STAMP) | StdExpert | 50 | 1000 |
We tested the following hyperparameter space:
Parameter | From | To | Steps |
---|---|---|---|
Learning Rate | 0.01 0.001 |
0.001 0.0001 |
10 10 |
L2 Penalty | 0.0001 0.00001 |
0.00001 0.000001 |
10 10 |
Decay Rate | 0.1 | 0.9 | 10 |
Decay Rate Step | 3 | 7 | 2 |
Dataset | Learning Rate | L2 Penalty | Decay Rate | Decay Rate Step |
---|---|---|---|---|
RSC15/4 | ||||
RSC15/64 | 0.008 | 0.0001 | 0.45 | 3 |
DIGINETICA | 0.0002 | 0.00007 | 0.1 | 5 |
DIGINETICA (STAMP) |