Fan Cheng

Prof. Fan Cheng

orcid

Research Focus & Metrics

Academic Career Length

13Years (2014-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
2 April 2021 in IEEE Access
Verified
Journal Article
Open access
HLHaoran Luo
Haoran Luo
Fan ChengFan Cheng
Fan Cheng
HYHeng Yu
Heng Yu
YYYuqi Yi
Yuqi Yi
Deep neural networks have been proved a success in multiple fields. However, researchers still favor traditional approaches to obtain more interpretable models, such as Bayesian methods and decision trees when processing heterogeneous tabular data. Such models are hard to differentiate, thus inconvenient to be integrated into end-to-end settings. On the other hand, traditional neural networks are differentiable but perform poorly on tabular data. We propose a hierarchical differentiable neural regression model, Soft Decision Tree Regressor (SDTR). SDTR imitates a binary decision tree by a differentiable neural network and is plausible for ensemble schemes like bagging and boosting. The SDTR method was evaluated on multiple tabular-based regression tasks (YearPredictionMSD, MSLR-Web10K, Yahoo LETOR, SARCOS and Wine quality). Its performance is comparable with non-differentiable models (gradient boosting decision trees) and better than uninterpretable models (regular FCNN). On top of that, it can produce fair results with a restricted number of parameters, only using a small forest or even a single tree. We also propose an “average entropy” metric to evaluate the level of interpretability of a trained, soft decision tree neural network. This metric also helps to select proper structure and hyperparameters for such networks.

30 September 2024
Verified
Book Chapter
Closed access
PDF file
KZKecheng Zheng
Kecheng Zheng
Fan ChengFan Cheng
Fan Cheng
WWWei Wu
Wei Wu
FLFan Lu
Fan Lu
SMShuailei Ma
Shuailei Ma
YSYujun Shen
Yujun Shen

1 January 2022
Verified
Conference Paper
Open access
PDF file
ZCZhoujun Cheng
Zhoujun Cheng
HDHaoyu Dong
Haoyu Dong
RJRan Jia
Ran Jia
PWPengfei Wu
Pengfei Wu
SHShi Han
Shi Han
DZDongmei Zhang
Dongmei Zhang
Zhoujun Cheng, Haoyu Dong, Ran Jia, Pengfei Wu, Shi Han, Fan Cheng, Dongmei Zhang. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2022.

22 September 2015 in IEEE Transactions on Information Theory
Verified
Journal Article
Closed access
Fan ChengFan Cheng
Fan Cheng
YGYanlin Geng
Yanlin Geng
Let X be an arbitrary continuous random variable and Z be an independent Gaussian random variable with zero mean and unit variance. For t > 0, Costa proved that e 2h(X+√t Z) is concave in t, where the proof hinged on the first and second order derivatives of h(X + √t Z). In particular, these two derivatives are signed, i.e., (∂/∂t)h(X + √tZ) ≥ 0 and (∂ 2 /∂t 2 )h(X + √tZ) ≤ 0. In this paper, we show that the third order derivative of h(X + √tZ) is nonnegative, which implies that the Fisher information J(X + √tZ) is convex in t. We further show that the fourth order derivative of h(X +√tZ) is nonpositive. Following the first four derivatives, we make two conjectures on h(X +√tZ): the first is that (∂n/∂t n )h(X +√tZ) is nonnegative in t if n is odd, and nonpositive otherwise; the second is that log J(X + √tZ) is convex in t. The first conjecture can be rephrased in the context of completely monotone functions: J(X + √tZ) is completely monotone in t. The history of the first conjecture may date back to a problem in mathematical physics studied by McKean in 1966. Apart from these results, we provide a geometrical interpretation to the covariance-preserving transformation and study the concavity of h(√t X +√1 - t Z), revealing its connection with Costa's entropy power inequality.

31 May 2021 in IEEE Access
Verified
Journal Article
Open access
STSongkai Tang
Songkai Tang
LJLuhua Jin
Luhua Jin
Fan ChengFan Cheng
Fan Cheng
In online product review systems, users are allowed to submit reviews about their purchased items or services. However, fake reviews posted by fraudulent users often mislead consumers and bring losses to enterprises. Traditional fraud detection algorithm mainly utilizes rule-based methods, which is insufficient for the rich user interactions and graph-structured data. In recent years, graph-based methods have been proposed to handle this situation, but few prior works have noticed the camouflage fraudster’s behavior and inconsistency heterogeneous nature. Existing methods have either not addressed these two problems or only partially, which results in poor performance. Alternatively, we propose a new model named Fraud Aware Heterogeneous Graph Transformer (FAHGT), to address camouflages and inconsistency problems in a unified manner. FAHGT adopts a type-aware feature mapping mechanism to handle heterogeneous graph data, then implementing various relation scoring methods to alleviate inconsistency and discover camouflage. Finally, the neighbors’ features are aggregated together to build an informative representation. FAHGT shows a remarkable performance gain compared to several baselines on different datasets.

4 April 2014 in IEEE Transactions on Information Theory
Verified
Journal Article
Closed access
Fan ChengFan Cheng
Fan Cheng
RYRaymond W. Yeung
Raymond W. Yeung
Consider a communication network represented by a directed graph G = (V, ε), where V is the set of nodes and 8 is the set of point-to-point channels in the network. On the network, a secure message M is transmitted, and there may exist wiretappers who want to obtain information about the message. In secure network coding, we aim to find a network code, which can protect the message against the wiretapper whose power is constrained. Cai and Yeung studied the model in which the wiretapper can access any one but not more than one set of channels, called a wiretap set, out of a collection A of all possible wiretap sets. In order to protect the message, the message needs to be mixed with a random key K. They proved tight fundamental performance bounds when A consists of all subsets of ε of a fixed size r. However, beyond this special case, obtaining such bounds is much more difficult. In this paper, we investigate the problem when A consists of arbitrary subsets of ε and obtain the following results: 1) an upper bound on H(M) and 2) a lower bound on H(K) in terms of H(M). The upper bound on H(M) is explicit, while the lower bound on H(K) can be computed in polynomial time when |A| is fixed. The tightness of the lower bound for the point-to-point communication system is also proved.