TAPES
A task-aligned joint forecasting pretraining framework for transaction sequences. It selects forecasting objectives by evaluating their downstream utility through short-range pretraining trials, and models conditional co-occurrence relationships among objectives using a shared discrete latent variable in a joint likelihood.
Timeline (last 90 days)
TAPES is a task-aligned joint predictive pretraining framework for transaction sequences.
TAPES evaluates the downstream utility of candidate target sets through short-range joint pretraining trials, rather than using target predictability as the selection criterion.
TAPES pretrains with a joint likelihood parameterized by shared discrete latent variables to explicitly model the conditional co-occurrence relationships among prediction targets.
In TAPES's discrete latent variable model, given the value of the latent variable, the prediction targets are conditionally independent; after marginalizing out the latent variable, a conditional co-occurrence structure emerges among the targets.
TAPES's complete pipeline is: construct candidate target set → evaluate downstream utility with short-range trials → select target set → reinitialize the encoder → joint pretraining → downstream fine-tuning.
TAPES was evaluated on real-world transaction risk scenarios and multiple public financial benchmarks, covering diverse downstream tasks.
Experimental results show that TAPES outperforms representative baseline methods.
Ablation studies verify the contribution of the two core components, the prediction target selection mechanism and the joint modeling strategy, to TAPES's performance.
The effectiveness of TAPES's method is currently mainly demonstrated within the experimental scope reported in the paper.
All Facts (9)
TAPES is a task-aligned joint predictive pretraining framework for transaction sequences.
50%UnverifiedTAPES evaluates the downstream utility of candidate target sets through short-range joint pretraining trials, rather than using target predictability as the selection criterion.
50%UnverifiedTAPES pretrains with a joint likelihood parameterized by shared discrete latent variables to explicitly model the conditional co-occurrence relationships among prediction targets.
50%UnverifiedIn TAPES's discrete latent variable model, given the value of the latent variable, the prediction targets are conditionally independent; after marginalizing out the latent variable, a conditional co-occurrence structure emerges among the targets.
50%UnverifiedTAPES's complete pipeline is: construct candidate target set → evaluate downstream utility with short-range trials → select target set → reinitialize the encoder → joint pretraining → downstream fine-tuning.
50%UnverifiedTAPES was evaluated on real-world transaction risk scenarios and multiple public financial benchmarks, covering diverse downstream tasks.
50%UnverifiedExperimental results show that TAPES outperforms representative baseline methods.
50%UnverifiedAblation studies verify the contribution of the two core components, the prediction target selection mechanism and the joint modeling strategy, to TAPES's performance.
50%UnverifiedThe effectiveness of TAPES's method is currently mainly demonstrated within the experimental scope reported in the paper.
50%