Temporal Convolutional Networks (TCN) have showcased strong per- formance in various forecasting tasks. Their capacity to exploit parallelization enables quicker and more efficient training and inference for the model. Moreover, TCNs can be combined with certain encryption methods to support privacy-preserving time series prediction, making them an appealing solution for introducing as an effective forecasting model. In this research, we incorporated a 1D-TCN model into TIMEX, a forecasting tool that offers an automated pipeline to reduce human intervention in executing forecasting tasks. However, TCN models require architecture optimization tailored to each dataset for optimal forecasting performance. To address this, we introduced a more flexible approach to constructing and training the TCN model by implementing a Neural Architecture Search pipeline built upon a recently released AutoML solution from Microsoft Research. Our solution is named TIMESTRAMP: Time Series Temporal Masterpiece with AutoML Powered Prediction. Ultimately, we evaluated our proposed solution on four benchmarks, to fairly compare our approach with the solution proposed in previous corresponding research work.
Le reti temporali convoluzionali (TCN) hanno dimostrato ottime prestazioni in vari compiti di previsione. La loro capacità di sfruttare la parallelizzazione consente un apprendimento e un’inferenza più veloci ed efficienti per il modello. Inoltre, i TCN possono essere combinati con specifici metodi di crittografia per supportare la previsione delle serie temporali che preservano la privacy, rendendoli una soluzione interessante da introdurre come un efficace modello di previsione. In questa ricerca, abbiamo incorporato un modello 1D-TCN in TIMEX, uno strumento di previsione che offre una pipeline automatizzata per ridurre l’intervento umano nell’esecuzione dei compiti di previsione. Tuttavia, i modelli TCN richiedono un’ottimizzazione dell’architettura su misura per ciascun set di dati per una previsione ottimale. Per affrontare questo problema, abbiamo introdotto un approccio più flessibile alla costruzione e all’addestramento del modello TCN, implementando una pipeline di ricerca dell’architettura neurale basata su una soluzione AutoML recentemente rilasciata da Microsoft Research. La nostra soluzione si chiama TIMESTRAMP: Time Series Temporal Masterpiece with AutoML Powered Prediction. Infine, abbiamo valutato la nostra soluzione proposta su quattro benchmark, per confrontare in modo equo il nostro approccio con la soluzione proposta in precedenti lavori di ricerca correlati.
AutoML for privacy-preserving temporal convolutional neural network
Sondoqah, Mousa
2022/2023
Abstract
Temporal Convolutional Networks (TCN) have showcased strong per- formance in various forecasting tasks. Their capacity to exploit parallelization enables quicker and more efficient training and inference for the model. Moreover, TCNs can be combined with certain encryption methods to support privacy-preserving time series prediction, making them an appealing solution for introducing as an effective forecasting model. In this research, we incorporated a 1D-TCN model into TIMEX, a forecasting tool that offers an automated pipeline to reduce human intervention in executing forecasting tasks. However, TCN models require architecture optimization tailored to each dataset for optimal forecasting performance. To address this, we introduced a more flexible approach to constructing and training the TCN model by implementing a Neural Architecture Search pipeline built upon a recently released AutoML solution from Microsoft Research. Our solution is named TIMESTRAMP: Time Series Temporal Masterpiece with AutoML Powered Prediction. Ultimately, we evaluated our proposed solution on four benchmarks, to fairly compare our approach with the solution proposed in previous corresponding research work.File | Dimensione | Formato | |
---|---|---|---|
2023_05_Sondoqah.pdf
solo utenti autorizzati dal 09/04/2024
Descrizione: Thesis
Dimensione
1.86 MB
Formato
Adobe PDF
|
1.86 MB | Adobe PDF | Visualizza/Apri |
Executive_Summary___Mousa_Sondoqah.pdf
solo utenti autorizzati dal 09/04/2024
Descrizione: Executive Summary
Dimensione
822.53 kB
Formato
Adobe PDF
|
822.53 kB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/204214