The tricky point in stock recommendation is that recommendations should give good investment performance but also should not ignore individual preferences. To tackle this challenge, we propose PfoTGNRec, which can handle time-varying collaborative signals and incorporates diversification-enhancing contrastive learning.
We provide source code for (1) Dynamic graph learning models (including our model), (2) Recommendation models, and (3) Stock recommendation models.
(1) Dynamic graph learning models
While most dynamic methods have not been utilized for recommendation tasks, we adapt their original architectures to recommendation task by incorporating negative sampling during training and applying BPR loss. The codes are implemented within the temporal graph network (TGN) framework by Rossi et al, 2020.
- PfoTGNRec: Our proposed model based on TGN that incorporates diversification-enhancing contrastive learning.
- DyRep (Trivedi et al., 2019): learns representation over dynamic graphs as latent hidden process bridging two observed processes of Topological Evolution of and Interactions on dynamic graphs.
- Jodie (Kumar et al., 2018): models the future trajectory of nodes in an embedding space to predict user-item interactions in dynamic scenarios.
- TGAT (Xu et al., 2020): incorporates a novel time encoding technique upon GAT which is a graph attention mechanism applied to static graphs.
(2) Recommendation models
Collaborative filtering (CF) algorithms that capture user preferences based on user behaviors. The codes are implemented within the Recbole framework by Zhao et al, 2021.
- BPR: a personalized ranking based recommender.
- WMF : a matrix factorization based recommender.
- LightGCN : a graph convolution network based collaborative filtering method.
- SGL : a self-supervised graph learning approach which incorporates a contrastive loss into the classical supervised task of recommendation.
You can check the code used here in baseline_rec.
(3) Stock recommendation models
CF based models that take into account both individual preferences and portfolio diversification. We implement the two most advanced methods with the code used in Chung et al, 2023.
- two-step method (Swezey and Charron, 2018): ranks items based on user preferences, and then re-ranks them based on the modern portfolio theory (MPT).
- MVECF (Chung et al, 2023): uses a ranking loss function based on the MPT into a matrix factorization model.
You can check the code used here in baseline_stock_rec.
In the paper, we used mock trading dataset obtained from a stock trading platform AlphaSquare. The datasets are available in Google Drive. The data is composed of seven sub-datasets, each consisting of seven periods.
The experiment was conducted on Python 3.10.11 and packages listed in requirements.txt.
You can execute the following scripts:
-
DyRep: Utilize run_dyrep.sh to run DyRep model.
-
Jodie: Utilize run_jodie.sh to run Jodie model.
-
TGAT: Utilize run_tgat.sh to run Jodie model.
-
PfoTGNRec: For our proposed method, PfoTGNRec, execute run_ours.sh
-
PfoTGNRec w. hyperparameters: If you wish to run our model with hyperparameter tuning, use run_ours_hyper.sh
These scripts offer a convenient way to execute the dynamic graph learning models and evaluate the output at once.
-
Recommendation models : Utilize run_baseline_rec.sh to run, and utilize run_baseline_rec_evaluation.sh to evaluate recommendation models
-
Stock recommendation models : Utilize run_baseline_stock_rec.sh to run, and utilize run_baseline_stock_rec_evaluation.sh to evaluate stock recommendation models
Furthermore, explore our data preprocessing that is available in the data_prepocess.ipynb.