| Title |
Laiko eilučių detekcija taikant autoenkoderiu grįstą požymių inžineriją / |
| Translation of Title |
Time series detection using autoencoder-based feature engineering. |
| Authors |
Leonavičius, Aurelijus |
| Full Text |
|
| Pages |
76 |
| Keywords [eng] |
time series analysis ; feature engineering ; autoencoder ; detection ; catch22 |
| Abstract [eng] |
Master’s final degree project topic - Time series detection using autoencoder-based feature engineering. Main task – explore the use of features extracted from autoencoder for timeseries detection. The need for features and methods of conventional methods of extraction are discussed in the literature review section. Theory behind techniques used is more thoroughly explored in methodology section. Two slightly different architecture types of autoencoders are explored and evaluated, some latent space visualizations are provided. Timeseries when transformed to their latent space representation are used in training the detection model. Detection model is based on gradient boosting technique. To build the model 42 different sets of timeseries was used, totaling a total of 45 807 unique time series. To evaluate the model’s performance the features extracted from latent space have been compared with catch22 feature set. A joint autoencoder-catch22 model has been created, which has an AUC score equal to 0.9416, which is slightly worse than catch22 model alone which has an AUC score of 0.9484, but due to the modification made computational speed of the detection task has been more than doubled. |
| Dissertation Institution |
Kauno technologijos universitetas. |
| Type |
Master thesis |
| Language |
Lithuanian |
| Publication date |
2022 |
