GENERATIVE MODELS WITH TIME ENCODING FOR SEASONAL TIME SERIES SYNTHESIS

MANY DEEP LEARNING MODELS FOR TIME SERIES PROCESS INPUT USING SLIDING WINDOWS WITHOUT INCLUDING TIME AS AN EXPLICIT VARIABLE. IN HIGH-FREQUENCY CONTEXTS REQUIRING REAL-TIME ANALYSIS, WINDOW HYPERPARAMETERS DEMAND CAREFUL CALIBRATION AND ERRORS PROPAGATE RAPIDLY IN AUTOREGRESSIVE FORECASTS. IN PRACTICE, MODELS ARE TYPICALLY SUPPLIED WITH A CONTINUOUS STREAM OF NEW OBSERVATIONS TO IMPROVE ACCURACY, BUT THIS RENDERS THEM FRAGILE WHEN DATA ARE MISSING OR DELAYED. THIS DISSERTATION PROPOSES AN ALTERNATIVE APPROACH FOR SYNTHESIZING SEASONAL TIME SERIES THROUGH DEEP LEARNING MODELS THAT LEVERAGE GEOMETRIC REPRESENTATIONS OF TIME. THE FIRST CONTRIBUTION IS AN INVERTIBLE TRANSFORMATION, CALLED HELICAL TIME ENCODING (HTE), THAT MAPS TIME ONTO A HIGHER-DIMENSIONAL SPACE WHERE SEASONAL PERIODICITY AND TEMPORAL PROGRESSION ARE GEOMETRICALLY EMBEDDED. THE SECOND IS THE GENERATION WITH THE TIMESTAMP TRICK (GENTT) FRAMEWORK, WHICH EXPLOITS TIME REPRESENTATIONS WITH PROPERTIES ANALOGOUS TO HTE TO TRAIN MODELS CAPABLE OF GENERATING TIME SERIES BY CONSIDERING ONLY TIMESTAMPS, WITHOUT STORING HISTORICAL DATA OR PROCESSING ORDERED SEQUENCES. GENTT IMPLEMENTATIONS USE MULTILAYER PERCEPTRON NETWORKS, ENABLING DEPLOYMENT ON RESOURCE-CONSTRAINED DEVICES WITHOUT HARDWARE ACCELERATION AND REDUCING ENERGY CONSUMPTION BY OVER THREE ORDERS OF MAGNITUDE COMPARED TO RECURRENT MODELS. WE VALIDATED THE GENTT FRAMEWORK FOR FORECASTING AND ANOMALY DETECTION TASKS USING CONDITIONAL GENERATIVE ADVERSARIAL NETWORKS (GANS) AND VARIATIONAL AUTOENCODERS (VAES). IN SEASONAL TIME SERIES FORECASTING, GANS AVOIDED THE ERROR PROPAGATION TYPICAL OF LONG-TERM AUTOREGRESSIVE PREDICTIONS. FOR ANOMALY DETECTION, VAES GENERATED EXPECTED VALUES FROM LEARNED SEASONAL PATTERNS, ENSURING PREDICTIONS ROBUST TO BOTH RECENT OUTLIERS AND MISSING DATA. A VARIANT OF THESE MODELS, DEVELOPED FOR THE TARGET INSTANT PAYMENT SETTLEMENT (TIPS) SERVICE, ENABLES REAL-TIME ANOMALY DETECTION IN HIGH-FREQUENCY PAYMENT STREAMS WHILE SUPPORTING SELECTIVE UPDATES TO HANDLE MODEL DRIFT. INTEGRATING THIS APPROACH INTO A FAILURE DETECTION SYSTEM FOR INSTANT PAYMENT INFRASTRUCTURE FACILITATES EXPLAINABLE DIAGNOSTICS THAT REDUCE ANALYSIS TIME AND ALLOW IMMEDIATE INCIDENT RESPONSE.