Two papers from NTT Laboratories have been accepted for publication for KDD 2026

Two papers authored by NTT laboratories have been accepted at KDD 2026 (32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining), to be held in Jeju, Korea, from August 9 to 13, 2026. KDD is known as one of the most prestigious international conferences in the field of data mining, with an acceptance rate of about 19% (4,164 papers submitted).

Abbreviated names of the laboratories:
HI: NTT Human Informatics Laboratories
CD: NTT Computer and Data Science Laboratories
CS: NTT Communication Science Laboratories

■Fast Vector Quantization Algorithm for ScaNN

Yasuhiro Fujiwara(CS), Angel Lopez Garcia-Arias(CS), Yasutoshi Ida(CD), Atsutoshi Kumagai(CD), Masahiro Nakano(CS), Makoto Nakatsuji(HI), Akisato Kimura(CS)

Vector quantization is a method that replaces vectors with codewords and is widely used to achieve fast and accurate similarity search based on inner products over large-scale data. ScaNN is one of the representative vector quantization methods, which achieves high approximation accuracy by replacing a vector with the codeword that minimizes the quantization error. However, since ScaNN requires computing the error for all codewords during replacement, it incurs a high computational cost, making the quantization process slow for large-scale datasets. In this study, we accelerated vector quantization while maintaining the search accuracy of ScaNN by efficiently computing upper and lower bounds on the quantization error and pruning codewords that cannot yield the minimum error. The proposed method will contribute to performance improvements in various applications that handle large volumes of data, including image retrieval, deep learning, and natural language processing.

■Moment Matters: Mean and Variance Causal Graph Discovery from Heteroscedastic Observational Data

Yoichi Chikahara, Research Scientist (CS)

Causal graph estimation aims to infer “what influences what” from observational data obtained from complex and unknown phenomena. However, conventional methods cannot distinguish whether a cause changes the “mean” of an outcome or its “variance,” which limits their ability to support a deeper understanding of complex phenomena and the design of effective interventions. In this study, we propose a new causal graph estimation technique that separately identifies causal relationships related to the mean and those related to the variance from observational data with heteroscedasticity. The proposed technique not only establishes the theoretical identifiability of these graphs but also quantifies the uncertainty of the estimation results. This enables the discovery of highly reliable causal relationships even from limited data and is expected to contribute to the understanding of complex social and scientific phenomena, including drug discovery, cellular response analysis, and economic policy.