Predicting HR Management Strategies and Employee Satisfaction in Enterprises Based on Deep Generative Models

Self-encoder is an unsupervised neural network model whose main purpose is to learn the implicit features of the input data and use the learned implicit features to reconstruct the original input data. These two steps become encoding and decoding, respectively, and the implicit features are called hidden vectors. Intuitively, auto-encoder is similar to feature dimensionality reduction, but it performs better than traditional dimensionality reduction methods such as principal component analysis because it can automatically extract valid features [25]. In addition to this, the new features learnt by the self-encoder can be fed into the downstream supervised model to perform subsequent tasks.

For an input sample x, the hidden vector z=E (x) is first obtained by an encoder, and in general the dimension of z is smaller than the dimension of x. Afterwards, the input sample x^ is reconstructed from z by means of a decoder. In order to make the reconstruction as good as possible, the mean square error between the two is generally minimised: 2LAE=‖ x−D(E(x)) ‖2

The variational autocoder is one of the most important generative models. Intuitively, VAE transforms real samples through an encoder to an ideal data distribution, such as the standard normal distribution, and then samples from this data distribution before passed to a decoder to get a bunch of generated samples. To measure the similarity of the generated samples to the original input samples, VAE uses KL scatter, as shown in Figure 2. Theoretically, VAE was originally implemented as a neural network for variational inference, which aims to model a complex unknown distribution from samples p (x).

Figure 2.

Variational A utoencoders

VAE does this by optimising a lower bound for p(x): 3logp(x)≥logp(x)−KL{ q(z|x)‖p(z|x) }Eq(z|x){ logp(x)+logp(z|x)−logq(z|x) }Eq(z|x){ logp(x,z)−logq(z|x) }Eq(z|x){ logp(x|z)+logp(z)−logq(z|x) }

The last line logp(x)≥E_q(z|x){logp(x|z)+logp(z)–logq(z|x)} of the above equation is also known as the Evidence Distribution Lower Bound (ELBO).

In the variational autocoder, the loss function of the variational autocoder is obtained by modelling the distribution q(z|x) by encoder E(·), modelling the distribution p(x|z) by decoder D(·) and replacing the expectation E with Monte Carlo sampling: 4LVAE=∑x∈X{ logp(x|z)+logp(z)−logq(z|x) }

Where the item log p(x|z) is equivalent to maximising the log-likelihood of the Gaussian distribution of the input samples at the output of the decoder, and the item log p(z)–log q(z|x) is equivalent to minimising the KL dispersion of the distribution q (z|x) of z the output of the encoder and the standard normal distribution. So compared to the ordinary two-self encoder, the variational self encoder differs in three main ways: (1) the outputs of the encoder and decoder are Gaussian distributions rather than a vector; (2) the constraints on the hidden variables have been added; and (3) the log-likelihood has been replaced from the reconstruction error.

Adversarial generative networks, like variational inference, are designed to estimate an unknown distribution from samples p(x). Adversarial generative networks do this by starting with a simple random distribution p(z) (typically a standard normal distribution) and then training a mapping from distribution p(z) to p(x), a decoder D(·). In order to make the decoder’s outputs as close as possible to the true distribution, adversarial generative networks also have a discriminator ϕ(·) [26]. The role of the discriminator is to input a sample and output the probability that it is a true sample and not generated by the decoder. The core idea of the adversarial generative network is to let the decoder and the discriminator “against each other”, that is, the decoder output samples to make the discriminator score it as high as possible, and the discriminator to make the decoder output score as low as possible, the real samples score as high as possible. The loss function is: 5 minDmaxϕV(D,ϕ)=Ex~p(x){ logϕ(x) }+Ez~p(z){ log(1−ϕ(D(z))) }

When sufficiently trained, the output of the decoder will be indistinguishable to the discriminator, i.e., the purpose of estimating the true distribution p(x) is achieved. However, in anomaly detection, it is necessary to get the corresponding hidden vector Z for a given input sample x, so if only GAN is used, a specific algorithm is needed to search the hidden vector z, which is very inefficient.

Adversarial Auto-Encoder (AAE) is a variant of an auto-encoder that uses adversarial training. For anomaly detection, the biggest benefit of AAE over GAN is that it can get the hidden vector corresponding to the input sample x directly from the encoder, while at the same time taking advantage of the powerful generative capabilities of GAN. The difference between AAE and a regular self-encoder is that a discriminator replaces the original reconstruction error.

The original image I can be represented as: 6I=λC+(1−λ)B

Where C represents the subject of the person in the image, B represents the background and λ controls the transparency of the pixels. The embedding target of the image can be represented as: 7I′=λC+(1−λ)B(B)

B represents the fuzzy algorithm. Here in this paper, a Gaussian blurring kernel of 99×99 is chosen. The average image information entropy of the test samples is reduced from 9.27 to 7.03 by the SOD priming and Gaussian blurring, which significantly reduces the stress of the embedding process.

The method proposed in this paper is applicable to any image embedding method. Here in this paper, a scheme for optimising potential coding is used as an example. Starting from an average potential encoding ω¯ , the iteration of ω and embedding loss L can be expressed as: 8L←L(G(ω),I′) 9ω←ω−ηF(∇ωL)

G(·) represents the generator network. The embedded latent codes ω and the embedded images G(ω) are optimised by the optimiser F.

Editing of face attributes can be achieved by moving the potential encoding ω in the potential space in the appropriate direction. The local scoring mechanism is an efficient way to adjust the existing orientation. To further optimise this method, this paper combines it with the residual attention MLP module. While the original local scoring mechanism has only been used to optimise linear orientations, the non-linear optimisation implemented in this paper further reduces the distortions generated during the editing process.

Given a pre-trained RA-MLP block, the target latent coding can be represented as: 10ω+=H(ω)=αF(ω)+(1−α)ω

Where F(·) represents the MLP module to be optimised for fine-tuning and α is used to control the magnitude of the attribute change.

Use r_i(ω) to denote the layer i activation of generator G(ω), and s_i(x,x⁺) to denote the average value of downsampling the segmentation mask of the semantic segmentation model of face attributes to the layer i resolution. Analogously, the objective function of this paper can be denoted as: 11LS(H(ω))=∑isi(x,x+)□| ri(ω)−ri(H(ω)) |2∑i| ri(ω)−ri(H(ω)) |2

The optimisation goal of this paper is not to obtain an edit direction for a potential code, but rather a nonlinear transformation network that can move the potential code more accurately.

Deep convolutional neural networks have been successfully utilized in image classification tasks to extract discriminative features. The basic idea is to perceive local regions in an image by a large number of neurons and increase the ability of the model to fit nonlinear problems by deepening the model depth [27]. Any face expression image x can be obtained with its corresponding different poses as well as the labelled face image G(x,y^p,y^p) under different expressions (for the sake of writing convenience G(x,y^e,y^p) will be abbreviated as G(x) next). Based on the original face image and the newly generated face image, a deep convolutional neural network can be used for feature extraction. The expression feature extraction network contains a total of 16 convolutional layers and 3 fully connected layers. In this network, each face expression image is propagated forward through the convolutional layers as well as the fully connected layers. The length of the output vector of the last fully connected layer is the same as the length of the face expression category in each database. After each convolutional layer, the emotional features of the original human face image x and the newborn human face image G(x) can be represented as: 12h(x)=Q(Wc*x+bc) 13h(G(x))=Q(Wc*G(x)+bc)

Where W_c and b_c denote the weight parameters as well as the bias terms of the expression feature extraction network, * denotes the convolution operation, and Q denotes the nonlinear activation function. With the deepening of the model layers, the model can learn emotion features with high level semantic information, which can effectively deal with the problem of low emotion recognition rate caused by external factors such as character differences, lighting changes, and posture changes.

The robust representation of emotion features can be obtained through the above steps, and the face expression classification task can be performed based on the extracted features. Specifically, a Softmax loss layer is used after the last fully connected layer of the expression feature extraction network to constrain the difference between the real expression categories and the predicted categories of the face image, so that its loss function can be expressed as: 14Lc(G,C)=Ex,y′[ −yelogC(h(G(x)),ye)−yelogC(h(x),ye) ]

This loss function contains two terms in total. Where the first term represents the cross its loss for the face expression images generated by the generative adversarial network, and the second term is the cross its loss for the original real face expression images. With the help of a large number of generated images, a good deep learning network can be adequately trained, which can be optimised using the Adam optimiser for the above loss function, and when the error loss of the loss function is no longer decreasing, the model training is finished, and an emotion classifier that can be used for multi-pose face expression recognition can be obtained.

According to equation (14) the total loss function of the model can be obtained as: 15 minG,CmaxD,DowαLcon(G)+βTV(G(f(x),y))+Lc(G,C)+Ex,y~pd(x,y)[ logDat(x,y) ]+Ex,y~pd(x,y) [ log( 1−Dat′(G(f(x)),y) ] +Ef*~prior(f)[ logDi(f*) ]+Ex~pd(x)[ log(1−Di(f(x))) ]

Where, α and β denote the coefficients of content consistency loss as well as total variation (TV) loss of the generated image, respectively, which are used to balance the smoothness as well as the resolution of the image. TV(·) denotes the total variation (TV) loss between neighbouring pixel points of the generated image, which can increase the smoothness of the generated image, which can be expressed as: 16LTV=∑ch=1CH∑w,h=1W,H‖ G(x)w+1,h,c−G(x)w,h,c ‖+‖ G(x)w,h+1,e−G(x)w,h,e ‖

Eq. (16) where ch denotes the number of channels of the image, and W and H denote the width and height of the image, respectively. During the model training process, forward propagation of the face image generation network and the expression feature extraction network will be performed first, followed by backpropagation. During the backpropagation process, the gradient of each network parameter will be computed by the Adam optimisation algorithm.

In order to verify the effectiveness of the algorithm in this chapter, facial expression recognition experiments were conducted using Cohn-Kanade, an international standard expression test database, which contains six types of expressions, namely, anger, happiness, sadness, surprise, disgust, and fear, each of which consists of a sequence of grey-scale, gradient expression images. Some of these data sequences are selected to be more precise for image sequence experiments. The main objectives of the experiment are (1) to test the degree of effectiveness of the algorithm proposed in this chapter in image sequence recognition. (2) To conduct comparative experiments between the static image recognition method and the sequence image recognition method using the Cohn-Kanade emoji library data and to conduct comparative analysis of the recognition results.

On the Cohn-Kanade emoji library selected, 650 pairs of complete image sequences for cross-validation design are shown in Figure 3. The experimental results indicate that this paper’s algorithm has an average recognition rate of 97.5%. The algorithms in this chapter have a better recognition rate. The error is mainly concentrated in the three expression images of sadness, fear, and disgust.

Figure 3.

Image array recognition rate

In order to compare the performance of the static image recognition and image sequence recognition modules, experiments are conducted on the Cohn-Kanade expression library on separate segmented image blocks and the overall image, respectively (1) A 16-layer convolutional neural network is designed for the recognition of static images, and the recognition results are shown in Table 1. (2) In this paper, the proposed algorithm is used to recognise dynamic sequence images and the recognition results are shown in Table 2. From the comparison results of Table 1 and Table 2, the main results are: (1) Compared with the average recognition rate of 93.52% for the static image algorithm, the average recognition rate of this paper’s algorithm is 98.72%, which is an improvement of 5 percentage points. (2) The two methods produce a large gap between the error results. The more obvious is the anger expression. Within the static recognition method, the recognition rate of 79.11% is low, while in the dynamic recognition of perceived changes in the state, the recognition rate is 100%.

Four groups of employees were set up, and each group of participants followed the same procedure: Baseline Measurement of Instructional Language→ Emotion→ Baseline Measurement of Advice→ Baseline Measurement of Job Satisfaction → Psychological Safety→ Situational Simulation 1→ Emotional Prediction Self-Report 1→ Suggestion Behavior Intention Posttest 1→ Situational Simulation 2→ Mood Prediction Self-report 2→ Suggestion Behavior Intention Posttest 2→ Situational Event Probability Judgment → Demographic Variable Survey.

Since each group of subjects had to go through two mental simulations of situational events, in order to avoid the effect of the simulation order of the constructed situation on emotion prediction, half of the subjects in each group were simulated in the order of AB from the constructed situation, and the other half were simulated in the order of BA. The test of the effect of the order of emotion prediction mainly used independent samples t-test and the test results are shown in Table 7. As can be seen from the table, there is no significant difference (p>0.05) in the subjects’ positive or negative emotion intensity prediction and duration prediction in each group of context combinations with different context order combinations, indicating that the order of the contexts did not affect the subjects’ prediction of the intensity and duration of the emotions in this investigation.