Application of high performance computing and deep neural network learning in intelligent city mechanical product recommendation

Abstract

Intelligent city is a product of the deep integration of information technology, industrialization and urbanization, which has a large number of intelligent mechanical products. The users widely evaluate their application characteristics, and the selection of mechanical products based on user evaluation has become a trend. Nowadays, personalized mechanical product recommendation based on user evaluation is more and more widely used. However, due to the sparse evaluation data, the recommendation accuracy needs to be improved. In this paper, the principle of matrix decomposition is deeply analyzed in order to provide useful ideas for solving this problem. The bias weight hybrid recommendation model of user preference and rating object characteristics is proposed, and the corresponding hybrid recommendation algorithm is designed. First, estimated data obtained using the matrix decomposition principle is supplemented to the sparse data matrix. Secondly, according to the characteristics of users and ratings, initial positions were set based on the statistical distribution of high-performance computing data, and bias weights were set by incorporating each feature. Finally, the nonlinear learning ability of deep neural network learning is used to enhance the classification effectiveness. Practice has proved that the constructed model is reasonable, the designed algorithm converges fast, the recommendation accuracy is improved by about 10%, and the model better alleviates the problem of sparse scoring data. The practical application is simple and convenient, and has good application value.

Keywords

Deep neural networks high performance computing sparsity bias weight intelligent city fast setting hybrid recommendation

1. Introduction

The Internet and the application systems running on it have generated massive data online. In the design [1] and evaluation of mechanical products [2], the recommendation based on multidimensional evaluation has become the mainstream personalized recommendation method [3], and greatly improved the recommendation. Personalized recommendation is the ability to accurately identify the characteristics of the user’s needs for the product usage, and to select the suitable product for them in a targeted manner [4]. Users evaluate the subjective satisfaction of products to create a scoring matrix, and collaborative filtering is conducted based on the correlation between rows and columns. It has been applied to personalized recommendation in many industries. The richer the data, the more accurate the recommendation. Kabbur et al. [5] solved the problem of sparse similarity matrix between items by decomposing the similarity matrix into the product of two low-rank matrices. Sarwar et al. [6] analyzed the item-based prediction methods and proposed a new item similarity model for improving the effectiveness of recommendation. Common collaborative filtering has no absolute advantage. Different application scenarios have different recommendation effects [7]. Since the initial data will have a cold start problem, there will be a Matthew effect [8]. Due to various objective reasons, the score set always has sparsity, and the coverage rate is very low, which affects the recommendation accuracy. The ideal goal of recommendation quality and effect cannot be achieved.

The new generation of intelligent computing technology represented by deep learning will be further applied to the optimized design, quality evaluation, and use recommendation of mechanical products. By utilizing distributed computing, sparse computing, and GPU acceleration, deep neural networks can parallelly process large-scale data and computing tasks, effectively achieving high-performance computing. It can obtain hidden, valid and understandable knowledge from huge amount of data. By expressing, modeling and quantifying the features of various personalized needs, and then substituting these data into deep learning networks for training and searching, and finally outputting personalized results automatically, this is the process of deep learning networks in personalized recommendation [9]. Nonlinear learning ability of deep learning technology can bring opportunities to solve problems [10] such as sparse data. Better results have been achieved in complex model design [11], state identification [12], performance optimization [13], metrics evaluation [14], and personalized recommendation [15]. Deep learning can use infinite approximation of an arbitrary continuous function to learn the nonlinear feature representations between the user and the item. Deep neural networks are utilized instead of the dot product operation in traditional linear relationships to capture feature representations between users and items, and can also be combined with other recommendation models. Analyze the correlation of evaluation and fuse the feature preference [16], use implicit characteristics information as a supplement to explicit characteristics [17], and the effective combination of them alleviates the inaccurate information. The fusion of preference characteristics, the fast placement in the initial position to accelerate convergence and the deep learning technology will solve the problem of the sparse evaluation data and effectively improve the recommendation speed and accuracy, which will be an effective hybrid recommendation method.

The main contributions are as follows:

(1)
Aiming at the problems such as sparse rating data, lack of personalized factors, and random initial data in traditional mechanical product personalized recommendation, using matrix analysis principles and deep learning technology, a hybrid recommendation model is proposed based on user preferences and characteristics of the rating object.
(2)
Based on the statistical distribution of data obtained from high-performance computing, the initial position is set, and weights are set by merging each feature. An intelligent recommendation algorithm is designed using deep neural network technology.
(3)
Practice has proven that the constructed model is reasonable, the designed algorithm has a fast convergence speed, and the recommendation accuracy has been improved by about 10%, effectively alleviating the problem of sparse rating data.

2. Recommendation model

Traditional recommendation models mainly include content recommendation, collaborative filtering, matrix decomposition, logistic regression, deep learning, etc., which have their own characteristics, as shown in Table 1.

Table 1
Recommendation models’ characteristics

Model	Advantages	Disadvantages
Content recommendation	Personalized recommendations	Limitations of user interest
Collaborative filtering recommendation	The principle is simple, direct, and widely used, solving the problem of cold start	Poor generalization ability, poor ability to handle sparse matrices, and significant head effects in recommendation results
Matrix decomposition	Compared to collaborative filtering, the generalization ability has been enhanced, and the processing ability for sparse matrices has been enhanced	It is difficult to utilize other user, item, and contextual features in addition to historical user behavior data
Logistic regression	Capable of integrating multiple types of different features	The model lacks the ability to combine features and has poor expression ability
Deep learning	Capable of handling large and complex data	High requirements for model training and computation

According to the principle of nonnegative matrix factorization, for a given nonnegative matrix A, a nonnegative matrix U and a nonnegative matrix V can be found. Thus, a nonnegative matrix is decomposed into the product of two nonnegative matrices. Given any nonnegative matrix ${A}\in R_{+}^{n\times m}$ , find nonnegative matrix ${W}\in R_{+}^{n\times k}$ and ${V}\in R_{+}^{k\times m}$ , and make $A_{n\times m}\approx W_{n\times k}V_{k\times m}$ , in which ${k}\ll{n}$ , ${k}\ll{m}$ , and generally $({{n}+{m}}){k}<n\ast m$ [18].

In the construction of intelligent cities, intelligent mechanical products have different application characteristics, and users will evaluate each feature and generate an evaluation matrix A. The element $a_{ij}$ in matrix A represents the i-th user’s evaluation scores of the j-th application feature, $\mu$ is the existing average scores, $b_{i}$ is the users’ rating tendency, and $b_{j}$ is item rating tendency. The recommendation model is established as follows:

$\displaystyle\text{Min }\textit{SSE}=\mathop{\sum}\limits_{n,m}({A-\hat{A}})^{2}$ $\displaystyle\text{s.t. }\widehat{a_{ij}}=\mu+b_{i}+b_{j}+w_{i}v_{j}^{T},i=1,2% ,\ldots n,j=1,2,\ldots,m$

Because of its simplicity, interpretability, less storage and other obvious characteristics, it can be used to reduce the matrix dimension describing the problem, and can also compress and summarize a large amount of data. Nonnegative matrix factorization is an NP problem. Since the product of $W$ and $V$ is an approximate estimate of $A$ , the standard of decomposition is the difference between the two, which can be solved by construct the objective function ${E}\in R_{+}^{n\times m}$ , ${E}={A}-{WV}$ to find suitable $W$ and $V$ to minimize $||{E}||$ .

$W$ and $V$ are obtained by minimizing the sum of squared errors (SSE), and the score prediction is performed.

$\displaystyle\textit{SSE}=\mathop{\sum}\limits_{n,m}({A-\hat{A}})^{2}=\mathop{% \sum}\limits_{n,m}(A-W_{n,k}V_{m,k}^{T})^{2}$ (1)

The optimization objective is to find the minimum value of SSE, and transform it into a machine learning problem.

$\displaystyle\text{Let }\widehat{a_{ij}}=w_{i}v_{j}^{T}=\mathop{\sum}\limits_{% k=1}^{K}w_{ik}v_{kj}$ (2) $\displaystyle e_{ij}^{2}=({a_{ij}-\widehat{a_{ij}}})^{2}=\left({a_{ij}-\mathop% {\sum}\limits_{k=1}^{K}w_{ik}v_{kj}}\right)^{2}$

The gradient descent method can be used to solve $w_{i,k}$ and $v_{k,j}$ .

$\displaystyle\frac{\partial}{\partial w_{i,k}}e_{i,j}^{2}=-2\left({a_{ij}-% \mathop{\sum}\limits_{k=1}^{K}w_{i,k}v_{k,j}}\right)v_{k,j}=-2e_{i,j}v_{k,j}$ $\displaystyle\frac{\partial}{\partial v_{k,j}}e_{i,j}^{2}=-2\left({a_{ij}-% \mathop{\sum}\limits_{k=1}^{K}w_{i,k}v_{k,j}}\right)w_{i,k}=-2e_{i,j}w_{i,k}$

$w$ and $v$ are updated according to the direction of the negative gradient, i.e., updated from $w_{i,k}$ to $w^{\prime}_{i,k}$ and from $v_{k,j}$ to $v^{\prime}_{k,j}$ , $\alpha$ is the update coefficient, then:

$\displaystyle w^{\prime}_{i,k}=w_{i,k}-\alpha\frac{\partial}{\partial w_{i,k}}% e_{i,j}^{2}=w_{i,k}+2\alpha e_{i,j}v_{k,j}$ (3) $\displaystyle v^{\prime}_{k,j}=v_{k,j}-\alpha\frac{\partial}{\partial v_{k,j}}% e_{i,j}^{2}=v_{k,j}+2\alpha e_{i,j}w_{i,k}$ (4)

Iterate and update until $\mathop{\sum}\nolimits e_{ij}^{2}\leqslant$ threshold, and finally converge.

To prevent over fitting, regularization term can be added, then:

$\displaystyle E_{i,j}^{2}=\left({a_{ij}-\mathop{\sum}\limits_{k=1}^{K}w_{i,k}v% _{k,j}}\right)^{2}+\frac{\beta}{2}\mathop{\sum}\limits_{k=1}^{K}(w_{i,k}^{2}+v% _{k,j}^{2})$ (5)

Also use the gradient descent method to solve.

$\displaystyle\frac{\partial}{\partial w_{i,k}}E_{i,j}^{2}=-2\left({a_{ij}-% \mathop{\sum}\limits_{k=1}^{K}w_{i,k}v_{k,j}}\right)v_{k,j}+\beta w_{i,k}=-2e_% {i,j}v_{k,j}+\beta w_{i,k}$ $\displaystyle\frac{\partial}{\partial v_{i,k}}E_{i,j}^{2}=-2\left({a_{ij}-% \mathop{\sum}\limits_{k=1}^{K}w_{i,k}v_{k,j}}\right)w_{i,k}+\beta w_{k,j}=-2e_% {i,j}w_{i,k}+\beta v_{k,j}$

Update according to the direction of the negative gradient.

$\displaystyle w^{\prime}_{i,k}=w_{i,k}-\alpha\frac{\partial}{\partial w_{i,k}}% (e_{i,j}^{2}+\beta w_{i,k})=w_{i,k}+\alpha({2e_{i,j}v_{k,j}-\beta w_{i,k}})$ (6) $\displaystyle v^{\prime}_{k,j}=v_{k,j}-\alpha\frac{\partial}{\partial v_{k,j}}% (e_{i,j}^{2}+\beta v_{k,j})=v_{k,j}+\alpha(2e_{i,j}w_{i,k}-\beta v_{k,j})$ (7)

Iterate and update until convergence. $\mathop{\sum}\limits_{k=1}^{K}w_{i,k}v_{k,j}$ get the score set, and make score recommendation.

3. Bias weight and fast setting recommendation

Due to the different preferences of individual users and the different characteristics of items, the scoring results are different. Different weight conditions can be set according to the bias of various characteristics, such as the average score, user rating tendency, item rating tendency and other characteristic factors.

$\widehat{a_{ij}}=\mu+b_{i}+b_{j}+w_{i}v_{j}^{T}$ , $\mu$ is the existing average scores, $b_{i}$ is the users’ rating tendency $b_{j}$ is Item rating tendency.

$\displaystyle E_{i,j}^{2}=\left({a_{ij}-b_{i}-b_{j}-\mathop{\sum}\limits_{k=1}% ^{K}w_{i,k}v_{k,j}}\right)^{2}+\frac{\beta}{2}\left({\mathop{\sum}\limits_{k=1% }^{K}(w_{i,k}^{2}+v_{k,j}^{2}})+b_{i}^{2}+b_{j}^{2}\right)$ (8)

Some bias conditions have time characteristics, so the time factor should be considered, then:

$\displaystyle\widehat{a_{ij}}(t)=\mu+b_{i}(t)+b_{j}(t)+w_{i}v_{j}^{T}(t)$ (9)

In order to improve the convergence direction, the initial value of the score is quickly set to the average position by using the statistical characteristics of the score. Then randomly initialize the average value to accelerate convergence.

Therefore, combining the characteristics of all parties and improving the fast convergence of the model, a comprehensive recommendation of offset weight fast setting hybrid recommendation is formed, which has better recommendation performance and effect.

4. Fusion CNN hybrid recommendation

According to the bias weight fast position estimation, complete the score matrix vacancy and input into the convolution neural network (CNN). After classification training, personalized recommendation can be made.

(1)
Set the number of hidden features KH, the number of iterations D, the threshold E, initialize the user matrix W and the feature matrix V.
(2)
Within the number of iterations D, unrated, repeat step (3)–(5);
(3)
Compute $\frac{\partial}{\partial w_{i,k}}E_{i,j}^{2}$ and $\frac{\partial}{\partial v_{k,{j}}}E_{i,j}^{2}$ ;
(4)
Update ( $w_{i,k}+\alpha({2e_{i,j}v_{k,j}-\beta w_{i,k}}))\to w_{i,k}$ , $(v_{k,j}+\alpha(2e_{i,j}w_{i,k}-\beta v_{k,j}))\to v_{k,j}$ ;
(5)
For the scored, calculate $E_{i,j}^{2}$ . If it is less than the threshold E, turn to step (6), otherwise turn to step (3).
(6)
Calculate the product A of W and V, get the score matrix A.
(7)
Set the A attribute column and the classification label column, and set the training set XA and the test set XB according to the ratio.
(8)
Input XA and XB into the designed CNN network, and output classification set AA, then make the classified recommendation.

The recommendation algorithm has the following characteristics.

1)
The generalization ability is strong, which can solve the problem of sparse matrix to a certain extent.
2)
The space complexity is low. The algorithm complexity is ${O}({({{m}+{n}})\ast k})$ , which is much lower than collaborative filtering ${O}({{m\ast m}})$ or ${O}({{n\ast n}})$ , or singular value matrix factorization ${O}({m\ast n^{2}+n\ast m^{2}})$ .
3)
Good flexibility and expansibility. To solve matrix vectors $w$ and $v$ , it can be combined with other features and deep learning networks.

5. Experimental verification

5.1 Bias weight recommendation

Six users rated 8 items on a 5-point scale after using them. The value of no rating is 0. Get the data set A $=$ [5, 0, 0, 1, 0, 3, 4, 3; 4, 0, 3, 1, 0, 4, 2, 4; 2, 3, 0, 5, 0, 3, 0, 5; 2, 0, 0, 5, 4, 0, 5, 4; 0, 1, 5, 4, 2, 0, 3, 2; 4, 0, 2, 1, 0, 2, 5, 3].

Parameter setting: iteration number D $=$ 2000, KH $=$ 4, $\alpha=$ 0.0002, $\beta=$ 0.02. The error threshold is 0.001. The recommendation model is established according to Eqs (6) and (7). Improve the model according to the time fluctuation change in the data interval. As shown in Fig. 1, result2 is the convergence of the improved result1 model, the error is reduced by 75%.

Figure 1.

The convergence of the improve model.

Figure 2.

The scoring matrix of result 2.

Figure 3.

Convergence of fast setting model.

Get the scoring matrix in Fig. 2. The unrated items have been completely filled, and personalized recommendation can be realized.

5.2 Fast setting recommendation

In order to speed up the convergence speed, W and V are quickly positioned to a reasonable position at the beginning, which is improved to a fast position recommendation model.

Count score set. According to the average user score and the average item evaluation score, respectively they are:

W0 $=$ [2.00 2.00 2.00 2.00; 2.25 2.25 2.25 2.25; 2.25 2.25 2.25 2.25; 2.50 2.50 2.50 2.50; 2.125 2.125 2.125 2.125; 2.125 2.125 2.125 2.125]; V0 $=$ [2.8333 0.6667 1.6667 2.8333 1.00 2.00 3.1667 3.5; 2.8333 0.6667 1.6667 2.8333 1.00 2.00 3.1667 3.50; 2.8333 0.6667 1.6667 2.8333 1.00 2.00 3.1667 3.50; 2.8333 0.6667 1.6667 2.8333 1.00 2.00 3.1667 3.50];

Table 2
CNN structure diagram

Layer	Name and description	Type	Activation
1	Input, 1211	Input	1211
2	Conv1, 16 2*1, Convolution, stride [1 1], padding [0 0 0 0]	Convolution	11116*1
3	BN1, Batch normalization, 16 channels	Batch normalization	11116*1
4	Relu1	Relu	11116*1
5	Pool_1, 2*1 Maximum pooling, stride [1 1], padding [0 0 0 0]	Max pooling	10116*1
6	Conv2, 32 2116, Convolution, stride [1 1], padding [0 0 0 0]	Convolution	9132*1
7	BN2, Batch normalization, 32 channels	Batch normalization	9132*1
8	Relu2	Relu	9132*1
9	pool_2, 2*1 Maximum pooling, stride [1 1], padding[0 0 0 0]	Max pooling	8132*1
10	Full_1, 4 fully connected layers	Full connection	114*1
11	Softmax	Softmax	114*1
12	Classoutput	Classification output	114*1

Figure 4.

CNN training progress.

Figure 5.

Classification accuracy.

Initialize it as a 5-point scoring matrix according to the average value. Fast recommendation can be achieved by iteratively generating W and V according to section 5.1. The convergence results are shown in Fig. 3. Result3 is the convergence of the improved fast setting recommendation, which quickly drops to the convergence direction and reduces the error by 70%.

5.3 CNN category recommendation

The dataset contains 357 rating samples with incomplete ratings (0 means unrated item). Each sample has 12 rating items and 1 recommended classification mark (4 categories). Based on section 5.1 and section 5.2, a complete scoring set is obtained, and a 12-layer CNN deep neural network is designed. As shown in Table 2, use the SGDM gradient descent method, the maximum number of training is 1000, the initial learning rate is 0.001 and learning rate decline factor is 0.1. After 450 training sessions, the learning rate was 0.001 * 0.1. The training is shown in Fig. 4.

The problem of data sparseness is solved to some extent by using offset weight fast position hybrid recommendation, and the recommendation accuracy is improved from 87.5% to 95.83%, as shown in Fig. 5.

6. Conclusion and future perspective

Aiming at the problems of sparse scoring data, lack of personalized factors and random initial data in traditional personalized recommendation of mechanical products, the characteristics of user scoring and resource scoring tendency are analyzed. According to the principle of matrix decomposition, user rating tendency, product rating tendency and other characteristic factors, different weight bias conditions are established. Based on the statistical characteristics of the score, the score value is quickly set to the position near the average value to accelerate the convergence. Based on the nonlinear learning ability of deep neural network, a hybrid recommendation method of bias weight fast setting is proposed. Compared with similar algorithms, the problem of data sparsity has been solved. the recommendation accuracy has improved by about 10% in the hybrid recommendation method. The recommendation accuracy is improved from 87.5% to 95.83%. The algorithm complexity is ${O}({({{m}+{n}})\ast k})$ , which is much lower than traditional collaborative filtering. It has good flexibility and expansibility, combined with deep neural network technology. The practical results show that the convergence speed is fast and the recommendation effect is good.

In the face of massive data in network application systems, properly combining deep learning technology and improving parallel computing will further improve the recommendation speed and accuracy, and will have better application prospects.

With the widespread application of mechanical products, the increase in application characteristics, the increase in data dimensions, and the imbalance of data will all affect the recommendation effect. At the same time, with the continuous increase in massive data, the calculation speed will be affected. Methods such as dimensionality reduction and parallel computing are used to solve these problems, and achieve high-performance computing goals. The next step is to continue to study the following: 1) Comprehensive analysis of multidimensional data. In future research, comprehensive analysis of data of different dimensions will be carried out, for example, relevant features between project data will be used to extract user preferences. The processing of data imbalance will also bring an undeniable impact on the application effect. The solution is that multiple dimensions of state information should be effectively integrated to improve the generalization and stability of the recognition model based on deep learning technology, to train a more effective state recognition model, and to establish a better deep learning state recognition model system. 2) Reasonable consideration of the impact of multiple auxiliary factors. Since the user-item interaction is related to many auxiliary factors, different auxiliary information for an item has different impacts on the user’s preference. In the future the possible methods need to consider the impacts of multiple auxiliary information on the model’s performance to further optimize the existing model.

Footnotes

Conflict of interest

The author declared no potential conflicts of interest with respect to the research, author-ship, and/or publication of this article.

Data availability

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

References

Liu

. Process knowledge recommendation system for mechanical product design. IEEE Access. 2020; (08): 112795-804. doi: 10.1109/ACCESS.2020.3002922.

Zhou

. Comprehensive evaluation method of wear state of mechanical equipment based on ferrography analysis. Ordnance Weapon Material Science and Engineering. 2022; 45(05): 175-82. doi: 10.14024/j.cnki.1004-244x.20220909.004.

Cao

Geng

. Perceptual evaluation method of mechanical products based on online product review and TextCNN. Machine Design and Research. 2022; 38(05): 189-94. doi: 10.13952/j.cnki.jofmdr.2022.0228.

. Personalized collaborative filtering recommendation algorithm based on weighted slope one. Application Research of Computer. 2017; 34(8): 2264-68. doi: 10.3969/j.issn.1001-3695.2017.08.005.

Hamanaka

Taneishi

Iwata

Pei

Hou

, et al. CGBVS-DNN: Prediction of compound-protein interactions based on deep learning. Molecular Informatics. 2017; 36(1-2): 1600045. doi: 10.1002/minf.201600045.

O’Byrne

Pakrashi

Schoefs

Bidisha Ghosh

. Semantic segmentation of underwater imagery using deep networks trained on synthetic imagery, Journal of Marine Science and Engineering. 2018; 6(3): 93-101. doi: 10.3390/jmse6030093.

Wang

Song

Yang

. Performance comparison and analysis of collaborative filtering recommendation algorithms. Computer Simulation. 2022; 39(09): 435-40. doi: 10.3969/j.issn.1006-9348.2022.09.083.

Zhou

Cui

Zhou

. Review of recommendation system. Journal of Computer Applications. 2022; 42(06): 1898-913. doi: 10.11772/j.issn.1001-9081.2021040607.

Greenspan

van Ginneken

Summers

. Guest editorial deep learning in medical imaging: Overview and future promise of an exciting new Technique. IEEE Transactions on Medical Imaging. 2016; 35(5): 1153-9. doi: 10.1109/TMI.2016.2553401.

10.

Tian

Pan

Yin

Wang

. Deep matrix factorization recommendation algorithm. Journal of Software. 2021; 32(12): 3917-28. doi: 10.13328/j.cnki.jos.006141.

11.

Lee

Mulder

EAB

Hopkins

. Mechanical neural networks: Architected materials that learn behaviors. Science Robotics. 2022; 7(71): eabq7278. doi: 10.1126/SCIROBOTICS.ABQ7278.

12.

Wei

. Research on machinery equipment state recognition based on deep neural network enhancement. Master thesis. Chongqing University of Posts and Telecommunications. 2022.

13.

. Adaptive momentum-based optimization to train deep neural network for computer simulation of hygro-thermo-mechanical vibration performance of laminated plates. Mechanics Based Design of Structures and Machines. 2021; 51(4): 1920-43. doi: 10.1080/15397734.2021.1883442.

14.

Chen

Jin

Zhang

. Deep neural network prediction of mechanical drilling speed. Energies. 2022; 15(9): 3037. doi: 10.3390/EN15093037.

15.

Zhang

Hua

Jia

Chen

. A personalized recommendation system based on deep neural network. Journal of Southwest University (Natural Science). 2019; 41(11): 104-9. doi: 10.13718/j.cnki.xdzk.2019.11.013.

16.

Han

Chen

Shi

. Matrix decomposition recommendation model incorporating item attribute preference. Journal of Xidian University (Natural Science). 2022; 49(03): 147-59. doi: 10.19665/j.issn1001-2400.2022.03.017.

17.

Wang

Jiang

Gong

. Developer hybrid recommendation algorithm based on combination of explicit features and implicit features. Journal of Software. 2022; 33(05): 1635-51. doi: 10.13328/j.cnki.jos.006553.

18.

Lee

Seung

. Algorithms for Non-negative Matrix Factorization. International Conference on Neural Information Processing Systems (NIPS); 2000, Denver, CO, USA. MIT Press. 2000. pp. 556-62.

Application of high performance computing and deep neural network learning in intelligent city mechanical product recommendation

Abstract

Keywords

1. Introduction

Table 1 Recommendation models’ characteristics

5.1 Bias weight recommendation

Table 2 CNN structure diagram

6. Conclusion and future perspective

Footnotes

Conflict of interest

Data availability

References

Table 1
Recommendation models’ characteristics

Table 2
CNN structure diagram