TL;DR
TimeInf is a novel, model-agnostic method for estimating data point contributions in time series, effectively detecting anomalies and providing interpretable attributions while preserving temporal dependencies.
Contribution
The paper introduces TimeInf, a new influence-based approach specifically designed for time series data, addressing the lack of principled attribution methods in this domain.
Findings
TimeInf outperforms existing attribution and anomaly detection methods.
It effectively detects anomalies and distinguishes diverse patterns.
TimeInf enables interpretation and visualization of data contributions.
Abstract
Evaluating the contribution of individual data points to a model's prediction is critical for interpreting model predictions and improving model performance. Existing data contribution methods have been applied to various data types, including tabular data, images, and text; however, their primary focus has been on i.i.d. settings. Despite the pressing need for principled approaches tailored to time series datasets, the problem of estimating data contribution in such settings remains under-explored, possibly due to challenges associated with handling inherent temporal dependencies. This paper introduces TimeInf, a model-agnostic data contribution estimation method for time-series datasets. By leveraging influence scores, TimeInf attributes model predictions to individual time points while preserving temporal structures between the time points. Our empirical results show that TimeInf…
Peer Reviews
Decision·ICLR 2025 Poster
The idea of inferring interpretability and anomaly analysis by measuring the contribution of time series data to the model's output is quite novel. The mathematical derivation is well-developed, and the details are thorough.
1. Compared to existing methods, the advantages of this approach are not fully emphasized, and the innovative aspects are not clearly defined. The main contribution of the paper can be seen as an extension of existing theory, but the method primarily involves deriving formulas based on Equation (3) and considering the Mixture of \( \delta z^{[m]} \). This operation has limited significance for theoretical extension and does not effectively demonstrate the contribution claimed. I do not believe i
1. The paper is well-written and easy to follow. 2. The presented solution is technically solid. 3. The experiments on anomaly detection can demonstrate the efficacy of the proposed method.
1. The relationship between the proposed model and unsupervised anomaly detection is unclear and should be further explained. 2. The model's generalization ability to other time series tasks (e.g., forecasting, imputation) is uncertain. Specifically, how would the contribution score of historical timestamps be measured for future time points? 3. There is no discussion on scalability. How does the method perform as the number of historical time points increases?
The primary technical strength of TimeInf lies in its novel and theoretically sound approach to data contribution estimation in time series. It successfully extends influence functions to handle temporal dependencies, addressing a significant gap in time series analysis that previous methods overlooked. The method is built on a strong mathematical foundation combining robust statistics with innovative use of overlapping blocks to preserve temporal structure. This theoretical rigor is balanced wi
1. Methodological Limitations: The paper’s core claim about their “distinctive integration” considering various temporal patterns seems to overlap significantly with existing attention mechanisms, without adequately differentiating itself. The theoretical justification for TimeInf lacks rigorous analysis of its properties, such as consistency or asymptotic behavior. Additionally, the paper doesn’t thoroughly explore the sensitivity of TimeInf to various hyperparameters, like block length or mode
Code & Models
Videos
Taxonomy
MethodsFocus
