Dataset Bias Research Articles

Evaluation methodology for deep learning imputation models.

There is growing interest in imputing missing data in tabular datasets using deep learning. Existing deep learning-based imputation models have been commonly evaluated using root mean square error (RMSE) as the predictive accuracy metric. In this article, we investigate the limitations of assessing deep learning-based imputation models by conducting a comparative analysis between RMSE and alternative metrics in the statistical literature including qualitative, predictive accuracy, statistical distance, and descriptive statistics. We design a new aggregated metric, called reconstruction loss (RL), to evaluate deep learning-based imputation models. We also develop and evaluate a novel imputation evaluation methodology based on RL. To minimize model and dataset biases, we use a regression imputation model and two different deep learning imputation models: denoising autoencoders and generative adversarial nets. We also use two tabular datasets from different industry sectors: health care and financial. Our results show that the proposed methodology is effective in evaluating multiple properties of the deep learning-based imputation model's reconstruction performance.

A survey on bias in visual datasets

Computer Vision (CV) has achieved remarkable results, outperforming humans in several tasks. Nonetheless, it may result in significant discrimination if not handled properly. Indeed, CV systems highly depend on training datasets and can learn and amplify biases that such datasets may carry. Thus, the problem of understanding and discovering bias in visual datasets is of utmost importance; yet, it has not been studied in a systematic way to date. Hence, this work aims to: (i) describe the different kinds of bias that may manifest in visual datasets; (ii) review the literature on methods for bias discovery and quantification in visual datasets; (iii) discuss existing attempts to collect visual datasets in a bias-aware manner. A key conclusion of our study is that the problem of bias discovery and quantification in visual datasets is still open, and there is room for improvement in terms of both methods and the range of biases that can be addressed. Moreover, there is no such thing as a bias-free dataset, so scientists and practitioners must become aware of the biases in their datasets and make them explicit. To this end, we propose a checklist to spot different types of bias during visual dataset collection.

P07.11.A Impact of awake craniotomy on extent-of-resection and performance in glioma: a retrospective, propensity-score matched cohort study

Abstract Background The goal of awake craniotomy is to safely optimize the extent-of-resection in patients with glioma. Lower postoperative tumor volume is associated with better overall survival, emphasizing the importance of optimal surgical resection. Awake craniotomy causes less postoperative neurological deficits than surgery under general anesthesia, however, it is unclear whether awake craniotomy leads to lower postoperative tumor volume compared with general anesthesia in presumed lower grade glioma. Methods Retrospective, matched cohort study in patients with astrocytoma, IDH-mutant grade 2 and 3, oligodendroglioma, IDH-mutant, and 1p/19q-codeleted grade 2 and 3, and low-grade glioma, IDH-wildtype (now designated glioblastoma, IDH-wildtype), who underwent resection with awake craniotomy or under general anesthesia between 2003 and 2021 at Erasmus MC Brain Tumor Center. Pre- and postoperative tumor volumes were measured by semi-automatic 3D MRI-segmentation. First, we performed a multivariate logistic regression to assess which factors predicted selection for awake craniotomy. Thereafter, matching based on propensity score was attempted. Outcome variables were postoperative tumor volume, resection percentage and Karnofsky Performance Status (KPS) 3 months after surgery. Results We identified 181 awake craniotomy-patients and 135 general anesthesia-patients. Awake craniotomy-patients were younger, in better condition, more often male, with tumors more often in eloquent areas, in the left side of the brain and non-contrast enhancing. When performing matching without replacement, only 68 awake craniotomy-patients could be matched with 68 general anesthesia-patients, underscoring the imbalance in the dataset. Matching with replacement yielded a matched dataset of 181 awake craniotomy-patients with 60 general anesthesia-patients with adequate matching on most baseline variables, except for eloquent area (47% for awake craniotomy and 21.7% for general anesthesia, p < 0.001). In this matched dataset, median postoperative volume in awake craniotomy was 5.8 mL (IQR 0 - 92.8) vs 12.2 mL (IQR 0 - 90.1) in general anesthesia (p-value = 0.114). Resection percentages did not differ between the awake craniotomy- and general anesthesia-groups. KPS scores at 3 months after surgery did not differ between awake craniotomy and general anesthesia (p = 0.15). Conclusion Postoperative tumor volume was not lower in awake craniotomy-patients than in general anesthesia-patients. Neither were resection percentage and KPS scores at 3 months after surgery significantly different between the groups. Adequate matching was only obtainable using replacement, underscoring the risk of bias in unmatched datasets. These data should be interpreted with care, given the retrospective nature, potential residual confounding and potential lack of generalizability due to unmatched normal resection patients.

Product- and Hydro-Validation of Satellite-Based Precipitation Data Sets for a Poorly Gauged Snow-Fed Basin in Turkey

Satellite-based Precipitation (SBP) products are receiving growing attention, and their utilization in hydrological applications is essential for better water resource management. However, their assessment is still lacking for data-sparse mountainous regions. This study reveals the performances of four available PERSIANN family products of low resolution near real-time (PERSIANN), low resolution bias-corrected (PERSIANN-CDR), and high resolution real-time (PERSIANN-CCS and PERSIANN-PDIR-Now). The study aims to apply Product-Validation Experiments (PVEs) and Hydro-Validation Experiments (HVEs) in a mountainous test catchment of the upper Euphrates Basin. The PVEs are conducted on different temporal scales (annual, monthly, and daily) within four seasonal time periods from 2003 to 2015. HVEs are accomplished via a multi-layer perceptron (MLP)-based rainfall-runoff model. The Gauge-based Precipitation (GBP) and SBP are trained and tested to simulate daily streamflows for the periods of 2003–2008 and 2009–2011 water years, respectively. PVEs indicate that PERSIANN-PDIR-Now comprises the least mean annual bias, and PERSIANN-CDR gives the highest monthly correlation with the GBP data. According to daily HVEs, MLP provides a compromising alternative for biased data sets; all SBP models show reasonably high Nash–Sutcliffe Efficiency for the training (above 0.80) and testing (0.62) periods, while the PERSIANN-CDR-based MLP (0.88 and 0.79) gives the highest performance.

Analysis of biases in automatic white balance datasets and methods

AbstractAnnotated datasets for automatic white balance (AWB) are used for the evaluation and, when necessary, the training, of AWB methods. Relying on such datasets requires awareness of the potential bias in their content and characteristics: some methods are designed to rely on the presence of particular elements, such as human skin, while other methods learn implicit relationships between image content and light properties from training data. The dependency on these relationships makes it fundamental to understand whether the available datasets are actually representative of common application scenarios, such as the presence of human subjects, the diversity of composition, or the illumination conditions. In this paper we overview the most common datasets for Automatic White Balance, including those for single as well as multiple illuminant estimation, providing a critical analysis on their characteristics. Furthermore, we identify a number of existing methods for single illuminant estimation, as a representative pool of approaches to the problem with various levels of complexity. We investigate how the performance of these correlate to the image content of common datasets.

Predictive modeling of microbiological seawater quality in karst region using cascade model

This paper presents an in-depth analysis of seawater quality measurements during the bathing seasons from year 2009 to 2020 in the city of Rijeka, Croatia. Due to rare occurrences of measurements with less than excellent water quality, considered dataset is deeply imbalanced. Additionally, it incorporates measurements under the influence of submerged groundwater discharges (SGD), which were observed in some bathing locations. These discharges were previously thought to dry up during the summer season and are now suspected to be one of the causes of increased Escherichia coli values. Consequently, and in view of the fact that the accuracy of prediction models can be significantly influenced by temporal and spatial variation of the input data, a novel cascade prediction modeling strategy was proposed. It consists of a sequence of prediction models which tend to identify general environmental conditions which confidently lead to excellent bathing water quality. The proposed model uses environmental features which can rather easily be estimated or obtained from the weather forecast. The model was trained on a highly biased dataset, consisting of data from locations with and without SGD influence, and for the time period spanning extremely dry and warm seasons, extremely wet seasons, as well as normal seasons. To simulate realistic application, the model was tested using temporal and spatial stratification of data. The cascade strategy was shown to be a good approach for reliably detecting environmental parameters which produce excellent water quality. Proposed model is designed as a filter method, where instances classified as less-than-excellent water quality require further analysis. The cascade model provides great flexibility as it can be customized to the particular needs of the investigated area and dataset specifics.

Ethical, Legal, and Financial Considerations of Artificial Intelligence in Surgery.

Machine learning systems have become integrated into some of the most vital decision-making aspects of humanity, including hiring decisions, loan applications, and automobile safety, to name just a few. As applications increase in both gravity and complexity, the data quality and algorithmic interpretability of the systems must rise to meet those challenges. This is especially vital for navigating the nuances of health care, particularly among the high stakes of surgical operations. In addition to inherent ethical challenges of enabling a "black box" system to influence decision-making in patient care, the creation of biased datasets leads to biased algorithms with the power to perpetuate discrimination and reinforce disparities. Transparency and responsibility are paramount to the implementation of artificial intelligence in surgical decision-making and autonomous robotic surgery. Machine learning has been permeating health care across diverse clinical and surgical contexts but continues to face sizable obstacles, including apprehension from patients and providers alike. To integrate the technology fully while upholding standard of care and patient-provider trust, one must acknowledge and address the ethical, financial, and legal implications of using artificial intelligence for patient care.

The silent trial - the bridge between bench-to-bedside clinical AI applications.

As more artificial intelligence (AI) applications are integrated into healthcare, there is an urgent need for standardization and quality-control measures to ensure a safe and successful transition of these novel tools into clinical practice. We describe the role of the silent trial, which evaluates an AI model on prospective patients in real-time, while the end-users (i.e., clinicians) are blinded to predictions such that they do not influence clinical decision-making. We present our experience in evaluating a previously developed AI model to predict obstructive hydronephrosis in infants using the silent trial. Although the initial model performed poorly on the silent trial dataset (AUC 0.90 to 0.50), the model was refined by exploring issues related to dataset drift, bias, feasibility, and stakeholder attitudes. Specifically, we found a shift in distribution of age, laterality of obstructed kidneys, and change in imaging format. After correction of these issues, model performance improved and remained robust across two independent silent trial datasets (AUC 0.85–0.91). Furthermore, a gap in patient knowledge on how the AI model would be used to augment their care was identified. These concerns helped inform the patient-centered design for the user-interface of the final AI model. Overall, the silent trial serves as an essential bridge between initial model development and clinical trials assessment to evaluate the safety, reliability, and feasibility of the AI model in a minimal risk environment. Future clinical AI applications should make efforts to incorporate this important step prior to embarking on a full-scale clinical trial.

Asian hate speech detection on Twitter during COVID-19.

Coronavirus disease 2019 (COVID-19) started in Wuhan, China, in late 2019, and after being utterly contagious in Asian countries, it rapidly spread to other countries. This disease caused governments worldwide to declare a public health crisis with severe measures taken to reduce the speed of the spread of the disease. This pandemic affected the lives of millions of people. Many citizens that lost their loved ones and jobs experienced a wide range of emotions, such as disbelief, shock, concerns about health, fear about food supplies, anxiety, and panic. All of the aforementioned phenomena led to the spread of racism and hate against Asians in western countries, especially in the United States. An analysis of official preliminary police data by the Center for the Study of Hate & Extremism at California State University shows that Anti-Asian hate crime in 16 of America's largest cities increased by 149% in 2020. In this study, we first chose a baseline of Americans' hate crimes against Asians on Twitter. Then we present an approach to balance the biased dataset and consequently improve the performance of tweet classification. We also have downloaded 10 million tweets through the Twitter API V-2. In this study, we have used a small portion of that, and we will use the entire dataset in the future study. In this article, three thousand tweets from our collected corpus are annotated by four annotators, including three Asian and one Asian-American. Using this data, we built predictive models of hate speech using various machine learning and deep learning methods. Our machine learning methods include Random Forest, K-nearest neighbors (KNN), Support Vector Machine (SVM), Extreme Gradient Boosting (XGBoost), Logistic Regression, Decision Tree, and Naive Bayes. Our Deep Learning models include Basic Long-Term Short-Term Memory (LSTM), Bidirectional LSTM, Bidirectional LSTM with Drop out, Convolution, and Bidirectional Encoder Representations from Transformers (BERT). We also adjusted our dataset by filtering tweets that were ambiguous to the annotators based on low Fleiss Kappa agreement between annotators. Our final result showed that Logistic Regression achieved the best statistical machine learning performance with an F1 score of 0.72, while BERT achieved the best performance of the deep learning models, with an F1-Score of 0.85.

Covert Network Construction, Disruption, and Resilience: A Survey

Covert networks refer to criminal organizations that operate outside the boundaries of the law; they can be mainly classified as terrorist networks and criminal networks. We consider how Social Network Analysis (SNA) is used to analyze such networks in order to attain a greater knowledge of criminal behavior. In fact, SNA allows examining the network structure and functioning by computing relevant metrics and parameters to identify roles, positions, features, and other network functioning that are not otherwise easily discovered at first glance. This is why Law Enforcement Agencies (LEAs) are showing growing interest in SNA, which is also used to identify weak spots and disrupt criminal groups. This paper provides a literature review and a classification of methods and real-case applications of disruption techniques. It considers covert network adaptability to such dismantling attempts, herein referred to as resilience. Critical problems of SNA in criminal studies are discussed, including data collection techniques and the inevitable incompleteness and biases of real-world datasets, with the aim of promoting a new research stream for both dismantling techniques and data collection issues.

Detection of Fake News Based on Typical Machine Learning Models

With the rapid expansion of the network, the glut of news spread everywhere. Because of the obscurity of news sources and the unrestricted types of viewers, the harmful impact of false news is more pervasive than ever before. The goal of this study is to evaluate the efficacy of five machine learning models, namely Decision Tree, Logistic Regression, Random Forest, Multilayer Perceptron (MLP) and Naive Bayes to detect false news using a dataset obtained from Kaggle. Following the application of five models for predicting false news based on the news' title and comparison of the training and testing accuracies of each model, the results indicate that Random Forest is the best model, with Decision Tree and MLP models also having very high testing accuracies. Surprisingly, the Naive Bayes model, widely recognized as the optimal classifier for text data, had the lowest testing accuracy in this study, implying that more research is required to explain this outcome. Finally, the limits of current machine learning algorithms, as well as the possibility of bias in datasets, provide a good direction for future studies.

Kernel dependence regularizers and Gaussian processes with applications to algorithmic fairness

Current adoption of machine learning in industrial, societal and economical activities has raised concerns about the fairness, equity and ethics of automated decisions. Predictive models are often developed using biased datasets and thus retain or even exacerbate biases in their decisions and recommendations. Removing the sensitive covariates, such as gender or race, is insufficient to remedy this issue since the biases may be retained due to other related covariates. We present a regularization approach to this problem that trades off predictive accuracy of the learned models (with respect to biased labels) for the fairness in terms of statistical parity, i.e. independence of the decisions from the sensitive covariates. In particular, we consider a general framework of regularized empirical risk minimization over reproducing kernel Hilbert spaces and impose an additional regularizer of dependence between predictors and sensitive covariates using kernel-based measures of dependence, namely the Hilbert-Schmidt Independence Criterion (HSIC) and its normalized version. This approach leads to a closed-form solution in the case of squared loss, i.e. ridge regression. We also provide statistical consistency results for both risk and fairness bound for our approach. Moreover, we show that the dependence regularizer has an interpretation as modifying the corresponding Gaussian process (GP) prior. As a consequence, a GP model with a prior that encourages fairness to sensitive variables can be derived, allowing principled hyperparameter selection and studying of the relative relevance of covariates under fairness constraints. Experimental results in synthetic examples and in real problems of income and crime prediction illustrate the potential of the approach to improve fairness of automated decisions.

Toward a perceptive pretraining framework for Audio-Visual Video Parsing

Audio-Visual Video Parsing (AVVP) is a new multi-modal weakly supervised task which aims to detect and localize events leveraging the partial alignment of audio and visual streams and weak labels. We identified two significant challenges in the AVVP: Cross-mode semantic misalignment and Contextual audio-visual dataset bias. For challenge 1, the existing methods tend to leverage the temporal similarity of the features. However, it is inappropriate for our AVVP task because multi-modal features with the same label do not always have the same semantics. Thus, we propose an instance-adaptive multi-modal time series max-margin loss (MTSM) which uses the temporal information to align features adaptively. Furthermore, to restrict the inescapable noise introduced during the feature fusion, we reuse the expression of MTSM in the single-mode. For the second challenge, we argue that bias mitigation should seek help from model generalization. Thus, we propose collocating pre-trained models: either” traverse” or based on domain-adaptation. First, we prove a hypothesis and then propose a method based on the Alternating Direction Method of Multipliers(ADMM) to decouple the optimal pre-trained model collocation solution, which reduces the time consumption. Experiments show that our method outperforms the contrastive methods.

SCORCH: Improving structure-based virtual screening with machine learning classifiers, data augmentation, and uncertainty estimation

IntroductionThe discovery of a new drug is a costly and lengthy endeavour. The computational prediction of which small molecules can bind to a protein target can accelerate this process if the predictions are fast and accurate enough. Recent machine-learning scoring functions re-evaluate the output of molecular docking to achieve more accurate predictions. However, previous scoring functions were trained on crystalised protein-ligand complexes and datasets of decoys. The limited availability of crystal structures and biases in the decoy datasets can lower the performance of scoring functions. ObjectivesTo address key limitations of previous scoring functions and thus improve the predictive performance of structure-based virtual screening. MethodsA novel machine-learning scoring function was created, named SCORCH (Scoring COnsensus for RMSD-based Classification of Hits). To develop SCORCH, training data is augmented by considering multiple ligand poses and labelling poses based on their RMSD from the native pose. Decoy bias is addressed by generating property-matched decoys for each ligand and using the same methodology for preparing and docking decoys and ligands. A consensus of 3 different machine learning approaches is also used to improve performance. ResultsWe find that multi-pose augmentation in SCORCH improves its docking power and screening power on independent benchmark datasets. SCORCH outperforms an equivalent scoring function trained on single poses, with a 1 % enrichment factor (EF) of 13.78 vs. 10.86 on 18 DEKOIS 2.0 targets and a mean native pose rank of 5.9 vs 30.4 on CSAR 2014. Additionally, SCORCH outperforms widely used scoring functions in virtual screening and pose prediction on independent benchmark datasets. ConclusionBy rationally addressing key limitations of previous scoring functions, SCORCH improves the performance of virtual screening. SCORCH also provides an estimate of its uncertainty, which can help reduce the cost and time required for drug discovery.

Limited generalizability of single deep neural network for surgical instrument segmentation in different surgical environments

Clarifying the generalizability of deep-learning-based surgical-instrument segmentation networks in diverse surgical environments is important in recognizing the challenges of overfitting in surgical-device development. This study comprehensively evaluated deep neural network generalizability for surgical instrument segmentation using 5238 images randomly extracted from 128 intraoperative videos. The video dataset contained 112 laparoscopic colorectal resection, 5 laparoscopic distal gastrectomy, 5 laparoscopic cholecystectomy, and 6 laparoscopic partial hepatectomy cases. Deep-learning-based surgical-instrument segmentation was performed for test sets with (1) the same conditions as the training set; (2) the same recognition target surgical instrument and surgery type but different laparoscopic recording systems; (3) the same laparoscopic recording system and surgery type but slightly different recognition target laparoscopic surgical forceps; (4) the same laparoscopic recording system and recognition target surgical instrument but different surgery types. The mean average precision and mean intersection over union for test sets 1, 2, 3, and 4 were 0.941 and 0.887, 0.866 and 0.671, 0.772 and 0.676, and 0.588 and 0.395, respectively. Therefore, the recognition accuracy decreased even under slightly different conditions. The results of this study reveal the limited generalizability of deep neural networks in the field of surgical artificial intelligence and caution against deep-learning-based biased datasets and models.Trial Registration Number: 2020-315, date of registration: October 5, 2020.

Cost-Sensitive Metaheuristic Optimization-Based Neural Network with Ensemble Learning for Financial Distress Prediction

Financial distress prediction is crucial in the financial domain because of its implications for banks, businesses, and corporations. Serious financial losses may occur because of poor financial distress prediction. As a result, significant efforts have been made to develop prediction models that can assist decision-makers to anticipate events before they occur and avoid bankruptcy, thereby helping to improve the quality of such tasks. Because of the usual highly imbalanced distribution of data, financial distress prediction is a challenging task. Hence, a wide range of methods and algorithms have been developed over recent decades to address the classification of imbalanced datasets. Metaheuristic optimization-based artificial neural networks have shown exciting results in a variety of applications, as well as classification problems. However, less consideration has been paid to using a cost sensitivity fitness function in metaheuristic optimization-based artificial neural networks to solve the financial distress prediction problem. In this work, we propose ENS_PSONNcost and ENS_CSONNcost: metaheuristic optimization-based artificial neural networks that utilize a particle swarm optimizer and a competitive swarm optimizer and five cost sensitivity fitness functions as the base learners in a majority voting ensemble learning paradigm. Three extremely imbalanced datasets from Spanish, Taiwanese, and Polish companies were considered to avoid dataset bias. The results showed significant improvements in the g-mean (the geometric mean of sensitivity and specificity) metric and the F1 score (the harmonic mean of precision and sensitivity) while maintaining adequately high accuracy.

Robust Detection of Fake News Using LSTM and GloVe Embeddings

The pervasive issue of fake news on digital platforms poses a significant threat to public opinion and trust in media. This paper addresses the problem of fake news detection by leveraging advanced natural language processing (NLP) techniques and deep learning models. Utilizing the ISOT Fake News Dataset, which comprises balanced samples of verified and fake news articles, we develop and evaluate two primary models: a Long Short-Term Memory (LSTM) network and a Convolutional Neural Network (CNN). The LSTM model employs pre-trained GloVe embeddings, followed by LSTM layers and a fully connected layer for classification, while the CNN model incorporates convolutional layers, max-pooling, and dropout for comparative analysis. Extensive pre-processing and exploratory data analysis (EDA) were conducted to clean the data and understand its characteristics. Our results demonstrate that the LSTM model outperforms the CNN model, achieving an accuracy of 99.58% on the test set. However, the high performance raises concerns about dataset biases, suggesting the need for more diverse and challenging datasets to ensure model robustness. Future work will focus on adversarial training and explainability techniques to enhance the model’s resilience and interpretability.

Obtaining Calibrated Probabilities with Personalized Ranking Models

For personalized ranking models, the well-calibrated probability of an item being preferred by a user has great practical value. While existing work shows promising results in image classification, probability calibration has not been much explored for personalized ranking. In this paper, we aim to estimate the calibrated probability of how likely a user will prefer an item. We investigate various parametric distributions and propose two parametric calibration methods, namely Gaussian calibration and Gamma calibration. Each proposed method can be seen as a post-processing function that maps the ranking scores of pre-trained models to well-calibrated preference probabilities, without affecting the recommendation performance. We also design the unbiased empirical risk minimization framework that guides the calibration methods to learning of true preference probability from the biased user-item interaction dataset. Extensive evaluations with various personalized ranking models on real-world datasets show that both the proposed calibration methods and the unbiased empirical risk minimization significantly improve the calibration performance.

Intelligent robust malware detection by implementing deep learning

Malicious software (ransom ware) cyber attacks in frequency and severity, posing an increasingly serious threat to computer systems everywhere. Malware detection is a hot study area as several multiple computers, organisations, and governments have been affected by an exponential rise in malware attacks. Dynamic and static assessment of malicious characteristics and behaviour patterns is time expensive and useless in real-time malware detection, according to current technologies. It is becoming increasingly common for malicious apps to use polymorphic and adaptive techniques to rapidly modify their behaviour and develop a number of new malicious apps. In order to undertake an effective malware analysis, machine learning techniques (MLAs) are increasingly being used to create new malware varieties. This approach is time-consuming since it requires considerable feature engineering, learning and representation of features. Moreover the feature extraction process could be effectively eliminated by using advanced MLAs like deep learning. These methods have been shown to perform better with a biased training dataset, which restricts their practical application in real-time scenarios. A new improved approach for successful zero-day malware detection must be developed in order to eliminate biases and analyze these approaches autonomously.

Scene graph generation with award-punishment strategy

The scene graph can provide a structured representation for downstream tasks given an image. To generate a fine-grained one, many researchers have attempted to alleviate the long-tailed dataset bias by energy-based learning, causal reasoning or other mechanisms. Nevertheless, they are still restricted to the biased dataset because of the redundancy of head information and the lack of tail information. In this work, we propose a comprehensive and ground-breaking model called Balanced Award-Punishment Model (BAPM) to tackle the problem. The BAPM consists of the stochastic strategy module (SSM), the knowledge transfer module (KTM) and the lateral inhibition loss (LIL). Concretely, the SSM takes dropout to form two different domain spaces for KTM, enhancing the tail information of the object level by constantly learning from another domain. The KTM aims to acquire abundant knowledge of tail predicates by transferring the fine-grained information from one domain to another. The SSM and KTM can be regarded as knowledge awards (KA) due to the incentive for tail data. The LIL mimics the competitive mechanism of neurons to smoothly adjust the weight of each object or each predicate by focal strategy. We regard the LIL as redundancy punishment (RP) owing to taking the restriction for head data into account. Under the joint award-punishment scheme, our approach has achieved state-of-the-art performance on two complementary metrics. The quantitative and qualitative experimental results on Visual Genome dataset show that our BAPM further reduce the dataset bias and generates more diverse scene graphs.