arXiv subjects

Haojun Zhang

Publications and source records attributed to Haojun Zhang.

3 recordsLinked to original sources

arXiv2026

X-AuT: Progressive Audio-Encoder Compression for Speech LLMs with Cross-Scale Distillation

Reducing audio-encoder depth lowers the inference cost of speech large language models, but removing complete blocks perturbs the embeddings consumed by the decoder and can cause deletion and premature end-of-sequence errors. We introduce X-AuT, a progressive framework that selects layer combinations through short behavioral probes and restores the pruned model through representation alignment, cross-scale distillation, scheduled student-policy supervision, and LoRA finetuning. The language-model backbone remains frozen, while attention LoRA adapters and the tied output embedding adapt during distillation. Training uses the highest-agreement tier from a transcript-consistency pipeline, followed by source reweighting during finetuning. On ten public Chinese--English benchmarks, compressing Qwen3-ASR-0.6B from 18 to 16 audio-encoder layers reduces macro-average error from 5.61% to 5.27%. The 14-layer model reaches 5.75% with 20.7% fewer audio-tower parameters. Under the matched recipe, the 1.7B teacher yields 5.55% mean error, compared with 8.45% for self-distillation, and progressive 18$\rightarrow$14 pruning outperforms direct pruning (5.75% vs. 6.73%). These single-run results establish two practical operating points and show that the accuracy effects vary across benchmarks. Project website: https://xpeng-ai.github.io/x-aut

cs.SD↗

arXiv2017

Toward a System Building Agenda for Data Integration

In this paper we argue that the data management community should devote far more effort to building data integration (DI) systems, in order to truly advance the field. Toward this goal, we make three contributions. First, we draw on our recent industrial experience to discuss the limitations of current DI systems. Second, we propose an agenda to build a new kind of DI systems to address these limitations. These systems guide users through the DI workflow, step by step. They provide tools to address the "pain points" of the steps, and tools are built on top of the Python data science and Big Data ecosystem (PyData). We discuss how to foster an ecosystem of such tools within PyData, then use it to build DI systems for collaborative/cloud/crowd/lay user settings. Finally, we discuss ongoing work at Wisconsin, which suggests that these DI systems are highly promising and building them raises many interesting research challenges.

cs.DB↗

arXiv2011

Controlling False Positives in Association Rule Mining

Association rule mining is an important problem in the data mining area. It enumerates and tests a large number of rules on a dataset and outputs rules that satisfy user-specified constraints. Due to the large number of rules being tested, rules that do not represent real systematic effect in the data can satisfy the given constraints purely by random chance. Hence association rule mining often suffers from a high risk of false positive errors. There is a lack of comprehensive study on controlling false positives in association rule mining. In this paper, we adopt three multiple testing correction approaches---the direct adjustment approach, the permutation-based approach and the holdout approach---to control false positives in association rule mining, and conduct extensive experiments to study their performance. Our results show that (1) Numerous spurious rules are generated if no correction is made. (2) The three approaches can control false positives effectively. Among the three approaches, the permutation-based approach has the highest power of detecting real association rules, but it is very computationally expensive. We employ several techniques to reduce its cost effectively.

cs.DB↗