tda

Tor data (./tor)

• download SelectedFeatures-10s-TOR-NonTOR.csv dataset from https://www.unb.ca/cic/datasets/tor.html
• overall code was taken from : https://kepler-mapper.scikit-tda.org/en/latest/notebooks/TOR-XGB-TDA.html (HJ van Veen)
• main code is tor.py
• keplermapper implementation is in km.py
• output is a kind of hypergraph. nodes are clusters of covers. node color means the ratio or tor traffic.

Questions

• if a cover is empty (e.g. 3 records/ 1 hypercube when kmeans.n_cluster=5, 3 < 5 == True) then I just assigned all records to the same cluster, which is not logical.
• Assuming uniform prior looks not very logical. Maybe we need kernel density estimation & transform it into quantile, or some other method. Maybe noninformative prior and bayesian update can be helpful?

Financial data (./transaction)

• 2025-05-30 KOSPI mini future transaction data was obtained from the KRX.
• main code is fin.py
• label is "buy_investor" ; inestor type of the buyer(long). For simplicity, 0 means institution, 1 means individual.
• Currently time-series aggregation is not done. TODO: aggregate by sum, or takens embedding.

TODO

• Need to do inspections, tuning
• Current ".diff()" destroys topological informations in time series. Do takens embedding.
• Quote data contains more information, and the orderbook topology can be meaningful