Explainable Mutual Fund Recommendation
Data
Please see 'DATA_DESCRIPTION.md' for mode detail.
Recommender System Methods
Baseline
- Collabarative Fiiltering
- PersonFreq
- PersonVolume
Stable
- LightFM Meta
- LightFM PureCF
- LightFM Hybrid
Advanced
- DGL
- GCN
Part I: Fund Recommedation
Training
Supported models
- Heuristic
- LightFM (CF/Hybrid/Meta)
- SMORe
# Process 3 models in parallel
bash run_all.sh
Arugments
You can also tune the detail parameter settings of each method in training pipeline.
--use_heuristic
">
# Commonly used arguments
--model
--model_type
--model_hidden_dimension
--evaluation_metrics
--use_heuristic
For example, LightFM with pure-CF method
EPOCHS=10
EMBED_SIZE=64
DATE=20181231
python3 train.py \
--path_transaction data/${DATE}/transaction_train.csv \
--path_transaction_eval data/${DATE}/transaction_eval.csv \
--path_user data/${DATE}/customer.csv \
--path_item data/${DATE}/product.csv \
--model 'LightFM' \
--model_path 'models/lightfm' \
--model_type 'cf' \
--model_hidden_dimension ${EMBED_SIZE} \
--model_max_neg_sample 100 \
--model_loss 'warp' \
--training_do_evaluation \
--training_verbose \
--training_num_epochs ${EPOCHS} \
--training_eval_per_epochs 1 \
--evaluation_diff \
--evaluation_regular \
--evaluation_metrics '[email protected]' \
--evaluation_metrics '[email protected]' \
--evaluation_metrics '[email protected]' \
--evaluation_metrics '[email protected]' \
--use_heuristic 'frequency' \
--use_heuristic 'volume' \
--evaluation_results_csv results/lightfm_cf_evaluation_${DATE}.csv \
--evaluation_rec_detail_report results/lightfm_cf_rec_detail_${DATE}.tsv \
> logs/lightfm_cf_exp_${DATE}.log
For another example, SMORe
python3 train.py \
--path_transaction data/${DATE}/transaction_train.csv \
--path_transaction_eval data/${DATE}/transaction_eval.csv \
--path_user data/${DATE}/customer.csv \
--path_item data/${DATE}/product.csv \
--model 'SMORe' \
--model_path 'models/smore' \
--model_hidden_dimension ${EMBED_SIZE} \
--model_max_neg_sample 100 \
--model_loss 'warp' \
--training_do_ \
--training_verbose \
--training_num_epochs ${EPOCHS} \
--training_eval_per_epochs 1 \
--evaluation_diff \
--evaluation_regular \
--evaluation_metrics '[email protected]' \
--evaluation_metrics '[email protected]' \
--evaluation_metrics '[email protected]' \
--evaluation_metrics '[email protected]' \
--evaluation_results_csv results/smore_evaluation_${DATE}.csv \
--evaluation_rec_detail_report results/smore_rec_detail_${DATE}.tsv \
> logs/smore_exp_${DATE}.log
Evaluataion
To use the evaluation pipeline, you need a prediction rec file with the format like the example below:
# prediction rec file
\t
\t
\t
\t
\t
CFDAXWccjJPoVInuiF0mMg== AG25 EXPLOIT SOLO 0 2 CFDAXWccjJPoVInuiF0mMg== XXXX EXPLOIT SOLO 0 1 CFDAXWccjJPoVInuiF0mMg== JJ15 EXPLOIT REGULAR 0 2 CFDAXWccjJPoVInuiF0mMg== XXXX EXPLOIT REGULAR 0 1 CFDAwH4y/ssuYSedFy8UMw== CC89 EXPLOIT REGULAR 0 2 CFDAwH4y/ssuYSedFy8UMw== XXXX EXPLOIT REGULAR 0 1 CFDA9UDJnLAm4/0txbPMVQ== AP06 EXPLORE NA 0 2 CFDA9UDJnLAm4/0txbPMVQ== XXXX EXPLORE NA 0 1
Later you could directly use the evaluate pipeline
bash rec_convert_eval.sh
In the evaluation pipeline, you need to convert the ground truth interaction into '.rec' format. For xample.
# truth rec file
\t
\t
\t
\t
CFDAXWccjJPoVInuiF0mMg== AG25 EXPLOIT SOLO 1.0 CFDAXWccjJPoVInuiF0mMg== JJ15 EXPLOIT REGULAR 1.0 CFDAwH4y/ssuYSedFy8UMw== CC89 EXPLOIT REGULAR 1.0 CFDA9UDJnLAm4/0txbPMVQ== AP06 EXPLORE NA 1.0
Convert from the evaluation transaction (includes the preprocess pipeline) by the following code, which will save the corresponding rec file in the defined argument '--path_trainsaction_truth'
DATE=20181231
python3 convert_to_rec.py \
--path_transaction data/${DATE}/transaction_train.csv \
--path_transaction_eval data/${DATE}/transaction_eval.csv \
--path_user data/${DATE}/customer.csv \
--path_item data/${DATE}/product.csv \
--path_transaction_truth rec/${DATE}.eval.truth.rec
And evaluate by the code "rec_eval.py"
DATE=20181231
python3 rec_eval.py \
-truth rec/${DATE}.eval.truth.rec \
-pred rec/pred.rec \
-metric '[email protected]' \
-metric '[email protected]' \
-metric '[email protected]' \
-metric '[email protected]'
The results would be like
TRUTH REC FILE EXISTED: 'rec/20181231.eval.truth.rec'
EvalDict({
SUBSET USERS EXAMPLES
* EXPLORE 2305 2826
* EXPLOIT 33355 62403
* REGULAR 31763 59054
* SOLO 2747 3349
})
==============================
[email protected] on EXPLORE 0.0001
[email protected] on EXPLORE 0.0004
[email protected] on EXPLORE 0.0004
[email protected] on EXPLORE 0.0004
[email protected] on EXPLOIT 0.0000
[email protected] on EXPLOIT 0.0001
[email protected] on EXPLOIT 0.0001
[email protected] on EXPLOIT 0.0001
[email protected] on REGULAR 0.0000
[email protected] on REGULAR 0.0001
[email protected] on REGULAR 0.0001
[email protected] on REGULAR 0.0001
[email protected] on SOLO 0.0001
[email protected] on SOLO 0.0004
[email protected] on SOLO 0.0004
[email protected] on SOLO 0.0004
==============================
Results
| Methods | [email protected] | [email protected] | [email protected] | [email protected] |
|---|---|---|---|---|
| Collabarative Fiiltering | - | - | - | |
| PersonFreq | - | - | - | |
| PersonVolume | - | - | - | |
| LightFM Meta | - | - | - | |
| LightFM PureCF | - | - | - | |
| LightFM Hybrid | 0.000 | 0.000 | 0.000 | 0.000 |
| DGL | - | - | - | |
| GCN | - | - | - |
1 Jul 14, 2022
55 Dec 25, 2022
2.8k Dec 29, 2022
3.6k Dec 30, 2022
53 Dec 05, 2022
197 Dec 28, 2022
37 Nov 28, 2022
28 Dec 28, 2022
39 Oct 26, 2022
3 Feb 15, 2022
10.6k Jan 01, 2023
1 Nov 15, 2021
1 Dec 14, 2021
4.2k Jan 02, 2023
375 Dec 30, 2022
32 Nov 29, 2022
87 Nov 29, 2022
1.6k Dec 05, 2022
3 Apr 22, 2022