Chenfeng LI

I'll briefly talk about the research methodologies and result in this session.

As mentioned above, the dataset contained images URLs (do not necessarily contain faces of politician) and corresponding texts. After cleaning the invalid and duplicate data, we obtain the source of each image from the second level domain of its url.

We made use of two pretrained open source model: FaceNet, an robust facial detection and recognition model, and RMN, a facial expression recognition model. The methodologies contains the following three parts:

1. We need to build a name recognition model, which detects if there are faces and who they are for an input image. Therefore, we need a labelled dataset, containing images of politicians' faces and their names. Wikimedia is a proper dataset for this, as it contains some metapages of "21st politicians in United States", where the page of each politician contains a portrait as well as some relevant images of them. For each politician, We passed the portrait through FaceNet to get the "controlled embedding vector", and collected more faces with embedding vector closed to the controlled one. After this process we get the label dataset, then trained a KNN model with \(k=1\) as the name recognition model.
2. Then we could work on the media dataset. In this research, we worked on some influencial politicians. Use "Donald Trump" as an example. To increase the efficiency, we specified some prompts, in this example "Trump", and filtered the observations whose texts part contains the prompts. We passed thess images to the FaceNet to get the face crops and embedding vectors, and then passed the embedding vectors through the name recognition model. If some vectors predicted "Donald Trump", we stored the face crops for the next process.
3. For the rectangular face crops, we extended it as square, modified the grey scale, and passed the them through the RMN model to get the emotion logits, which are 7-dimensional vectors representing the combination of seven basic emotions. Then we can analyze these logits.

We then analyzed the emotion logits by categorizing them by the eight news corporations and/or the political bias (progressive or conservative). There are two solutions for this problem. The first and intuitive one is to visualize the logits. We implemented the UMAP to reduce the dimension of the logits to 2, and for each news corporation we created a scatter plots with different colour for each politicians. The plots

• The first and intuitive one is to visualize the logits. We implemented the UMAP to reduce the dimension of the logits to 2, and for each politicians we created a scatter plots with different colour for different news corporations or media bias. The above are some examole plots. According to these plots, we could not distingished the emotions based on news corporations
• However, the result after reducing the dimension may not be generalized. Therefore, we built some KNN classifier on the emotion logits with the same idea, and evaluate the models with validation accuracies and confusion matrixes. We found that these models are not necessarily better than the baseline models, which derived a same conclusion that the relationship between politicians' emotion and news corporations/media bias is not significant.

In conclusion, our results indicated that there is no clear relationship between the facial expressions of specific politicians in media images and news corporations or their political bias. Some possible reasons are listed as follows:

• Lack of Deliberate Selection of Facial Expressions by News Corporations, as it may have negative affect on long-term interest.
• Limited Representation of Political Bias Across News Corporations. Eight news media may not be enough.
• Inadequate Representation of Facial Expressions by Emotion Logits. The seven 7-dimensional emotion logits may not represent complex facial emotion well.
• Same expression may have different meaning. An angry expression could be described as irritable and agitated, and in contrast passion and assertiveness.
• The Impact of Timing on Photo Selection. The time of each image is not given, as the rating of a politicans for each media may change over time.

For a supervised learning problem, with observed features \(X_1, \dots ,X_n\) and corresponding target \(Y\), we want to use the \(X\)'s that cause \(Y\) and avoid using those caused by \(Y\). We call the correlation between \(Y\) and \(X\)'s caused by \(Y\) as spurious correlations. The models with spurious correlation are possibly made wrong prediction.

The regular model (such as ERM) couldn't handle such problem, and we could consider DRO model. The idea of DRO is to grouped the data based on features, then minimize the risk of the wrost cases. We use colored MNIST dataset as an example.

MNIST is a dataset with images of hand-written digits and the corresponding numbers. For each image, label it as \(Y_i=0\) if the digits within 0-4 or \(Y_i=1\) if the digit within 5-9. Then colored the label 0 images with red and label 1 images green. For the training set, flip the color with probability 0.25, while for the testing set, flip the color with probability 0.8.

Under this setting, Image -> Digit -> Label \(Y\) -> Color, where "->" denotes causation, and Image and Color are observed features. The correlation between label and color are spurious. If applying ERM, the model would learn this relationship from training set, but as such causation in testing set are different, the testing performance is poor.

To apply the DRO model, we separate the dataset into four subgroups based on color (red or green) and label (0 or 1). Then train a model that minimize the risk of the group with highest risk. This effectively eliminate the influence of spurious correlation. Moreover, to avoid overfitting, we use DRO model regularization. This model have satisfying training and testing accuracy.

Slice sampler (SS): For a probability distribution \(f(x)\) with two or more peaks, and small function value in between, some sampling method (such as Metropolis–Hastings algorithm) may keep sampling values around one peak and ignore the others. Slice sampler is available for this situation. SS is an MCMC algorithm. For an initial \(x_0\), it first uniformly sample a value between 0 and \(f(x_0)\), then uniformly sample \(x_1\) from \(\{x|f(x)>f(x_0)\}\), and repeat this process.

Chinese Restaurant Franchise (CRF): This is a generalization of Chinese Restaurant Process (CRP). In CRP each step choose either one of existing parameters or sample a new parameter. In CRF, it assumes an amount of Chinese restaurants sharing same distribution of parameter.

Hierarchical Dirichlet Process (HDP): This is a expansion of Dirichlet Process (DP), which is an nonparametric Bayesian. In HDP, we set a prior distribution for the base distribution of DP to reflect its variation. Therefore, HDP is more flexible than DP but more computational expensive.

In this research, for the clustering problem from the same given dataset, the performance of CRF is better than the HDP with SS.

First client is Mr. Kwang from UChicago Medicine. This client wanted to study the effect of a medication for COVID patients on the requirement of ventilation. We advised to apply logistic regression, and provided suggestions on grouping the patients and strengthen the analysis result.

Second client is Ms. Wang from UChicago BSD. This client wanted to study the effect of HCBS (a kind of health-care service) on Post-Acute Care (PAC) based on a dataset with 700,000 observations, and she was interested in using fixed effects (FE) with propensity scores (PS) weighting in the model. We suggested to use logistic regression model without PS weighting.

Third client is Ms. Gan from UChicago Medicine. This client wanted to learn if the distance between the middle point of the chest and left entricle is significant accross different age group and other physical factors. We pointed out an issue of the covariate choice (BMI should not be included as it is a function of weight and height), advised linear regression model and provided the number of samples should be collected.

In general, we analyzed the medical, econimical and cultural value of Chinese medicine and concluded that it is worth developing. According to the data result we noticed that the public have positive view in Chinese medicine. However, the Chinese medicine faces many issues, such as non-scientific theory, treatment effectiveness, qualify of products and lack of talent. We raised suggestions such as pharmacological analysis, patent protection, commercialization.

In data analysis part, I wanted to study the effect of ages and education level on the idea of Chinese medicine (overall, theory, effectiveness and choice), and the perception of some specific symptoms, the advantages and shortages of Chinese medicine. Therefore, I designed a questionaire with 10 questions (required 1-2 minutes to finish), and distributed online and on-site with the help of teammates.

We collected 153 valid answers. I used linear regression on the first problem. As the ages and education level were ordinal categorical variables, and to analyze the difference between each two neighbor categories, I encoded with staircase dummy variables. The result showed that generally, the age had positive effect ,and education level had weak negative effect on perception. Besides, I found people thought Chinese medicine is more effective in treating minor and chronic diseases.

The limitation in this analysis is that the sample size is small, especially for the elderly group, and many people rejected to answer, which may cause bias on result. Besides, we can choose more covariates (such as income level).

This is my first research project, when I was new on machine learning.

I divided the review part into two, detection and recognition. In detection part I provided some personal thinking and reviewed Viola-Jones Algorithm, CNN and Multi-task Cascade Convolutional Network. In recognition part I reviewed Eigenface and CNN.

In the inplementation part I chose a dataset with 2,370 face images (of size 50*37) from 34 individuals from Sklearn (fetch_lfw_people with 30 minimum faces, resize = 0.4), and then constructed a CNN model through tensorflow. Split the data with training, validation and testing sets at a ratio 8:1:1. Achieve a testing accuracy of 80.6%.