Searcharxiv⌕ Search

arXiv subjects

Yijin Zeng

Publications and source records attributed to Yijin Zeng.

5 recordsLinked to original sources

DEFTS: Data-Efficient, Training-Free Few-Shot Learning for CLIP in the Text Subspace

Contrastive Language-Image Pre-training (CLIP) achieves impressive zero-shot image classification performance by comparing image embeddings against class-specific text embeddings. While recent few-shot methods leverage labeled training samples to improve upon this zero-shot baseline, they often predominantly rely on image embeddings and underutilize the semantic structure of the text embeddings. In this paper, we introduce DEFTS, a data-efficient, training-free few-shot classification method strictly in the text subspace. We first demonstrate empirically that while raw text embeddings can be suboptimal as direct classifiers, the linear subspace they span remains highly discriminative. DEFTS therefore proposes to reduce the image-text modality gap by projecting estimated image prototypes directly into this subspace. Importantly, by expressing these prototypes as linear combinations of text embeddings, DEFTS reduces a high-dimensional estimation problem to the derivation of just a few coefficients. These coefficients are then computed via a Bayesian posterior mean, using a prior naturally centered on the original text embedding coordinates. Extensive experiments across 12 image classification benchmarks demonstrate that DEFTS significantly improves upon existing training-free methods.

cs.CV↗

Scale two-sample testing with arbitrarily missing data

This work proposes a novel rank-based scale two-sample testing method for univariate, distinct data when a subset of the data may be missing. Our approach is based on mathematically tight bounds of the Ansari-Bradley test statistic in the presence of missing data, and rejects the null hypothesis if the test statistic is significant regardless of the missing values. This proposed scale testing method is then combined with the location testing method proposed by Zeng et al. (2024) using the Holm-Bonferroni correction for location-scale testing. We show that our methods control the Type I error regardless of the values of the missing data. Simulation results demonstrate that our methods have good statistical power, typically when less than 10% of the data are missing, while other missing data methods, such as case deletion or imputation methods, fail to control the Type I error when the data are missing not at random. We illustrate the proposed location-scale testing method on hepatitis C virus dataset where a subset of values is unobserved.

stat.ME↗

Exact Bounds of Spearman's footrule in the Presence of Missing Data with Applications to Independence Testing

This work studies exact bounds of Spearman's footrule between two partially observed $n$-dimensional distinct real-valued vectors $X$ and $Y$. The lower bound is obtained by sequentially constructing imputations of the partially observed vectors, each with a non-increasing value of Spearman's footrule. The upper bound is found by first considering the set of all possible values of Spearman's footrule for imputations of $X$ and $Y$, and then the size of this set is gradually reduced using several constraints. Algorithms with computational complexities $O(n^2)$ and $O(n^3)$ are provided for computing the lower and upper bound of Spearman's footrule for $X$ and $Y$, respectively. As an application of the bounds, we propose a novel two-sample independence testing method for data with missing values. Improving on all existing approaches, our method controls the Type I error under arbitrary missingness. Simulation results demonstrate our method has good power, typically when the proportion of pairs containing missing data is below $15\%$.

stat.ME↗

MMD Two-sample Testing in the Presence of Arbitrarily Missing Data

In many real-world applications, it is common that a proportion of the data may be missing or only partially observed. We develop a novel two-sample testing method based on the Maximum Mean Discrepancy (MMD) which accounts for missing data in both samples, without making assumptions about the missingness mechanism. Our approach is based on deriving the mathematically precise bounds of the MMD test statistic after accounting for all possible missing values. To the best of our knowledge, it is the only two-sample testing method that is guaranteed to control the Type I error for both univariate and multivariate data where data may be arbitrarily missing. Simulation results show that our method has good statistical power, typically for cases where 5% to 10% of the data are missing. We highlight the value of our approach when the data are missing not at random, a context in which either ignoring the missing values or using common imputation methods may not control the Type I error.

stat.ME↗

On two-sample testing for data with arbitrarily missing values

We develop a new rank-based approach for univariate two-sample testing in the presence of missing data which makes no assumptions about the missingness mechanism. This approach is a theoretical extension of the Wilcoxon-Mann-Whitney test that controls the Type I error by providing exact bounds for the test statistic after accounting for the number of missing values. Greater statistical power is shown when the method is extended to account for a bounded domain. Furthermore, exact bounds are provided on the proportions of data that can be missing in the two samples while yielding a significant result. Simulations demonstrate that our method has good power, typically for cases of $10\%$ to $20\%$ missing data, while standard imputation approaches fail to control the Type I error. We illustrate our method on complex clinical trial data in which patients' withdrawal from the trial lead to missing values.

stat.ME↗