Research on performance improvement of personalized recommendation algorithm based on deep neural network optimization
Data publikacji: 21 mar 2025
Otrzymano: 29 paź 2024
Przyjęty: 06 lut 2025
DOI: https://doi.org/10.2478/amns-2025-0623
Słowa kluczowe
© 2025 Xianglin Xiao, published by Sciendo
This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.
In personalized recommendation algorithms, whether they are content-based recommendation algorithms, collaborative filtering recommendation algorithms or hybrid recommendation algorithms combining the two, when faced with massive multi-source heterogeneous data, they are all subject to the limitations of shallow models, relying on manually-designed features, and failing to tap into the deeper features of the user and the item [1-4], which makes the phenomena such as cold-start and data sparsity leading to the problem of low accuracy in the recommendation algorithm revealed, and the research on hybrid recommendation algorithms fusing heterogeneous auxiliary information from multiple sources still faces serious challenges [5-7]. Deep learning, which has made good progress in many fields such as NLP and image processing in recent years, is able to mine the feature information implied in images, text and other information [8-10], and its feature characterization ability is very powerful and it can automatically carry out feature learning, which brings new opportunities for the research of personalized recommendation algorithms [11-13].
On the one hand, deep learning can effectively fuse heterogeneous data from multiple sources by mapping different data into a hidden space, and automatically perform feature learning [14-15]. On the other hand, deep learning has a powerful feature extraction capability, which can be used to represent a large amount of heterogeneous data of users and items by learning a deep nonlinear network structure, so as to obtain a deep feature representation of users and items [16-19]. Incorporating deep learning into traditional recommendation algorithms can effectively deal with massive multi-source heterogeneous data, thus alleviating the problem of low recommendation accuracy caused by cold-start and data sparsity of traditional recommendation algorithms [20-22], and the deep learning model has strong noise immunity and good robustness, so that the performance of the recommendation system with the addition of the deep learning model is higher [23-24].
This paper is devoted to the research of personalized recommendation algorithm, which improves the existing matrix decomposition recommendation algorithm and combines the deep neural network model on its basis to improve the quality of personalized recommendation. After combining matrix decomposition, deep neural networks, and other related principles, a matrix decomposition model based on deep neural network improvement is proposed. We use neural network to extract the user information features and item information features, and extract the text feature information in the user and item through Doc2vec model, use text convolution network to transform the text into word vectors, calculate the similarity of the vectors and apply it to the matrix decomposition algorithm. After obtaining the user and item features, the user-item feature matrix is fused through the fully connected layer, and the feature matrix with auxiliary information is taken as the original input and incorporated into the matrix decomposition algorithm. Finally, iterative training is performed using Adam’s algorithm to optimize the model function by minimizing the error between the real data and the generated data to arrive at the best model. The personalized recommendation performance of the matrix decomposition personalized recommendation model proposed in this paper is examined through personalized recommendation experiments to explore the effects of various factors such as the number of MLP layers, pre-training, model parameters, and so on, on its recommendation performance.
With the development of the Internet and electronic devices, recommendation systems have filled people’s lives. However, with the arrival of the big data era, the traditional recommendation learning method is not as effective as before in the process of dealing with a large amount of data. In order to effectively improve the degree of user personalized satisfaction in the process of massive information, this paper proposes a matrix decomposition personalized recommendation model based on deep neural network improvement to improve the accuracy of personalized recommendation for users.
Traditional recommendation algorithms are usually categorized into three types: content-based recommendation algorithms, collaborative filtering recommendation algorithms, and hybrid recommendation algorithms, among which collaborative filtering recommendation algorithms can be divided into two types of recommendations, which are content-based and model-based.
The main idea of this recommendation algorithm is based on the user’s past behavior towards the item. The core of this algorithm is to extract and analyze the features of the item or content and calculate its similarity, so as to recommend the item or content that the user may like. Content-based recommendation algorithms usually use natural language processing methods to deal with the relevant content, through which the user’s favorite content can be obtained, so as to recommend to the user the content that may be of interest.
Collaborative filtering recommendation algorithms are usually divided into three types: user-based collaborative filtering recommendation algorithms, project-based collaborative filtering recommendation algorithms, and model-based collaborative filtering recommendation algorithms.
User-based collaborative filtering recommendation algorithms are mainly used to discover the interests of a user through the user’s historical behavior on the project, and then evaluate the degree of similarity between the user and the user to find users with similar interests, and recommend to the target user the content that similar users have interacted with.
Item-based collaborative filtering recommendation algorithm is based on the user’s historical rating information to discover the user’s preferences, find the similarity between the items and the items, and then recommend similar items to the target user based on the user’s interests and preferences.
Model-based collaborative filtering recommendation algorithm is to analyze the user’s historical operation behavior of the item with the help of machine learning method, and to predict the ratings of the items that the user has not over-rated by learning the items that the user has over-rated, and to build mathematical models in the process.
Hybrid recommendation algorithms combine multiple recommendation algorithms to improve their accuracy and coverage.It can combine the results of different algorithms through a model-based approach or a weighting-based approach to produce better recommendations.Hybrid recommendation algorithms have the potential to address difficult problems in recommender systems, such as the cold-start problem and the variety of recommendation results.
In collaborative filtering recommendation algorithms, matrix decomposition is one of the most frequently used methods, which mainly utilizes the features of machine learning and implicit semantic modeling to mine potential user-item relationships [26]. Matrix decomposition is a method to decompose the user-item scoring matrix into two low-dimensional potential feature matrices, which are the user potential feature matrix and the item potential feature matrix, so that the result of multiplying the two potential feature matrices is approximately equal to the user’s original scoring matrix, e.g., the user-item scoring matrix is
An objective function is constructed based on the loss between the predicted and true values, and the objective function is shown in Equation (2):
Where
Deep neural network is a hierarchical model which is composed of multiple neural network layers, each layer consists of multiple neurons which are used to process the input data and output the result produced [27]. Deep neural network mainly consists of input layer, multiple hidden layers and output layer. Each layer changes the input data into a higher level representation and uses it as input data for the next layer. During model training, the deep neural network learns how to adjust the connection weights between each layer to ensure that the error between the predicted value and the actual value is made as small as possible, and the adjustment of the weights between layers needs to be accomplished by utilizing backpropagation, which is mainly used to update the individual weight values by calculating the gradient of the error. Deep neural networks are mainly categorized into three types: feed-forward deep networks, feedback deep networks and two-way deep networks.
Matrix decomposition is mostly applied in the rating prediction part of recommendation algorithms, which is an important aspect of recommendation algorithms.The matrix decomposition method is generally used to decompose the user-item rating matrix to obtain two implied feature matrices, and to assess the likelihood of a user liking an item based on these features.In specific application scenarios, the algorithm is capable of solving the recommendation problem in several scenarios.The five categories below are the main classification of recommendation algorithms based on matrix decomposition.
1) Collaborative filtering recommendation algorithms based on matrix decomposition
The algorithm decomposes the high-dimensional user-item rating matrix into two low-dimensional matrices, i.e., user latent feature matrix and item latent feature matrix, and multiplies these two latent feature matrices to predict the ratings of the items not rated by the user, and then ranks these ratings in descending order to recommend to the user the top
2) Matrix Decomposition Algorithm Based on Probabilistic Models
This algorithm views the matrix of users’ historical scores on items as a hybrid model consisting of multiple probabilistic models and uses maximum likelihood or Bayesian inference methods to learn the parameters in the model for recommendation.
3) Recommendation algorithm based on tensor decomposition
The algorithm is to view the user-item rating matrix as a three-dimensional tensor, and by decomposing the tensor, the potential feature vectors of multiple dimensions such as user, item and time are obtained, and the tensor multiplication is used to predict the user’s rating of the unrated item, so as to recommend the relevant content to the user.
4) Recommendation algorithm based on non-negative matrix decomposition
The algorithm obtains non-negative potential feature vectors of users and potential feature vectors of items by decomposing the user-item scoring matrix, and then makes predictions and recommendations. Common recommendation algorithms based on non-negative matrix decomposition include non-negative matrix decomposition algorithms, attribute-based non-negative matrix decomposition algorithms, and so on.
5) Deep learning based matrix decomposition algorithm
The algorithm models the user’s scoring matrix through a deep neural network and learns the potential feature vectors of the user and the item to make predictions and recommendations. Common deep learning-based matrix decomposition algorithms include neural network-based matrix decomposition algorithms, self-encoder-based matrix decomposition algorithms, and so on.
This chapter proposes an improved hybrid recommendation model based on deep learning. The whole framework is divided into two parts, the first part contains feature extraction from deep neural networks and text convolutional neural networks; the extracted features are then put into the second part, matrix decomposition based algorithm for rating prediction.
The first layer of this network architecture is the embedding layer, where user features are extracted from the input user and item information data by inputting user ID, user gender, user age, and user occupation, and item features are extracted by inputting item
First of all, the input layer performs the operation of inputting
Where,
Finally, the full join operation is performed and the output is obtained as user attribute feature vector. The same process is performed for the item attribute feature vector.
In this chapter, text convolutional neural network is used to extract text features from users and projects. Doc2vec model has excellent property of extracting text semantics and can find the similarity between text vectors by calculating the distance between vectors. Therefore, in this paper, the Doc2vec model is used to pre-train the word vectors for the embedding of the text to retain the correlation feature between the words. [28]. The embedding vectors are convolved using a convolutional layer, then the main features are selected using maximum pooling, and finally, the feature vectors are outputted into a specific dimensional space by a fully connected layer. Take the text feature extraction method of item name as an example:
The Doc2vec model transforms the word vectors into a two-dimensional matrix, and the pre-trained matrix is input into the embedding layer. Specifically, firstly,
The convolutional layer is used to do convolution on the text sequence. By using convolution kernels with different window sizes to operate on the embedding matrix, the step size of the sliding window is set to
Since the use of ReLU does not create the problem of oversaturation and the convergence obtained by ReLU is better, this activation function uses ReLU as the activation function.
Since the length of the convolution kernel is not a fixed value, the length of the feature vectors after convolution is different and it is difficult to construct the next layer, so the vector length of the text needs to be fixed. The pooling layer operates by performing max-pooling pooling from the convolutional layer to construct a feature vector with a fixed length. In this chapter, max-pooling is used to select the maximum value of each pooled region, i.e., the most salient features of the user or item are retained.
As shown in Eq. (9),
where
The fully connected layer receives the output pass from the pooling layer and integrates the feature vectors. The output of the pooling layer is mapped into the dimension space using the Relu activation function, and finally the text feature vector representation is output, as shown in Equation (10):
After the fully connected layer, the output in the model is shown in equation (11) below:
In this paper, a convolutional neural network is used to extract textual features from item textual information, aiming for better extraction of user attribute features and item attribute features. Therefore, the input user and item features in this chapter are the user and item feature matrices extracted using the neural network, rather than directly obtained from the scoring matrix using the traditional matrix decomposition algorithm, thus allowing the model to be denser in terms of real data.
In model optimization, in order to solve the overfitting problem after training due to sparse data in matrix decomposition, the L2 regular term is used to reduce the size of the weights, and the ability of the model to solve the overfitting problem is improved by using the reduction of model complexity. The loss function is shown in equation (12):
At this time, the user feature vector
The project eigenvector
The Adam algorithm is used to optimize the parameter training and
In order to verify the personalized recommendation performance of the matrix decomposition personalized recommendation model proposed in this paper, two public datasets, MovieLens and Pinterest1, are selected for experimental verification in this chapter in comparison with other existing recommendation algorithms. The statistical properties of the datasets are specifically shown in Table 1. The MoiveLens dataset used in this chapter contains ratings data collected from multiple users for multiple movies.The Pinterest dataset used in this chapter contains interaction data of multiple users with images. As can be seen from the table, both datasets are sparse, especially the Pinterest dataset, which has a sparsity of 99.68%.
Dataset
| Dataset | Number of interactions | Number of users | Number of projects | Sparsity |
|---|---|---|---|---|
| Movie Lens | 1000309 | 3807 | 6050 | 95.42% |
| 1500709 | 9948 | 55175 | 99.68% |
In order to investigate the factors affecting the performance of the proposed deep neural network-based matrix decomposition personalized recommendation model, three experiments are designed to compare the differences in the recommendation performance of the model with the number of MLP layers, pre-trained and non-pre-trained with Adam’s algorithm, and different model parameter settings, respectively.
In the model based on deep neural network, the depth of the network and the effect of the model have a great correlation, generally speaking, the higher the number of layers of the network, the better the effect of the corresponding model, therefore, this experiment mainly explores whether the number of layers of the neural network in the recommendation algorithm will have a significant impact on the recommendation accuracy. In this experiment, the number of MLP hidden layers is set to 0, 1, 2 and 4 layers respectively, and the results of this model on two data sets are shown in Table 2. From the table, it can be found that with the increase of the number of MLP hidden layers, the recommendation accuracy is gradually improved, so it can be assumed that increasing the number of layers of the neural network can improve the effect of the recommendation model. At the same time, when there is no hidden layer, the recommendation accuracy is extremely low, which also shows that the direct connection between user embedding and item embedding can’t tap the user-item interaction information well, and it must be converted by the hidden layer.
HR values of different MLP layers
| Dataset | Factors | MLP-0 | MLP-1 | MLP-2 | MLP-4 |
|---|---|---|---|---|---|
| MovieLens | 8 | 0.46 | 0.629 | 0.668 | 0.665 |
| 16 | 0.45 | 0.678 | 0.685 | 0.69 | |
| 32 | 0.464 | 0.693 | 0.702 | 0.709 | |
| 8 | 0.298 | 0.85 | 0.853 | 0.862 | |
| 16 | 0.301 | 0.867 | 0.856 | 0.862 | |
| 32 | 0.298 | 0.87 | 0.861 | 0.86 |
Pre-training the model helps to find the global optimal solution faster, and the model in this paper is trained using Adam’s algorithm to optimize its parameters. In this experiment, two datasets are used to verify the actual impact of the model pre-trained by Adam’s algorithm on the recommendation effect, which is mainly reflected by the two evaluation indexes of hit rate (HR) and normalized discount cumulative gain (NDCG), and the specific experimental results are shown in Table 3. From the results in the table, it can be seen that there is some overall improvement in the recommendation effect of the pre-trained model compared with that of the non-pre-trained model, and the model improves the HR value by up to 4.12% after pre-training, while the NDCG value improves by up to 5.7%. The experimental results indicate that pre-training Adam’s algorithm before initializing the matrix decomposition recommendation model has a facilitating effect on the recommendation of the model.
Model pretraining
| Training | Factors | MovieLens | |||
|---|---|---|---|---|---|
| HR | NDCG | HR | NDCG | ||
| With Pre_training | 8 | 0.707 | 0.402 | 0.885 | 0.564 |
| 16 | 0.722 | 0.445 | 0.88 | 0.567 | |
| 32 | 0.721 | 0.446 | 0.891 | 0.548 | |
| Without Pre_training | 8 | 0.679 | 0.409 | 0.879 | 0.545 |
| 16 | 0.704 | 0.421 | 0.875 | 0.559 | |
| 32 | 0.71 | 0.442 | 0.868 | 0.545 | |
In this section, the effect of deep neural network hyperparameters on the matrix decomposition-based personalized recommendation model of this paper is investigated. In each comparison, keeping other settings unchanged and adjusting the corresponding parameter values, the new evaluation metrics used are Precision (P@K), Mean Reverse Ranking (MRR), Recall (R@K), Mean Average Precision Mean (MAP).
This time, the performance of the model in this paper will be evaluated using the number of convolution kernels with 8, 16, 24, 32, 40, 48 respectively, and the evaluation results are shown in Fig. 1. Fig. (a) represents the effect of the number of convolutional kernels on ranking quality and Fig. (b) represents the effect of the number of convolutional kernels on recommendation accuracy. It can be seen that the trends of the line graphs of MRR and NDCG are basically the same, both of them are increasing first and then decreasing. The performance of MRR and NDCG is optimized at the same time when the number of convolutional kernels is 32.The trend of change in the foldplot performance of P@5 and R@5 is consistent, but not with that of MAP. However, P@5 and R@5 as well as MAP finally reach the optimal performance when the convolutional kernel is 32. Considering the ranking quality and recommendation accuracy together, the model performance is optimal when the number of convolutional kernels is 32.
Recommendation accuracy
The impact of output dimensions on ranking quality and recommendation accuracy is shown specifically in Fig. 2. Figure (a) represents the impact of output dimension on ranking quality, and Figure (b) represents the impact of output dimension on recommendation accuracy. In terms of output dimensions, the trends of the line graphs of MRR and NDCG are basically the same, and the performance is optimal when the output dimension is 64.The trends of the line graphs of the performance of P@5 and R@5 are the same, but they are not the same as that of the line graphs of MAP. However, the performance of P@5 and R@5 as well as MAP is almost optimal when the output dimension is 64. Together, the best performance in terms of sorting quality and prediction accuracy can be seen at an output dimension of 64.
The influence of output dimension
There are three common activation functions, namely the Sigmoid function, Tanh function, and ReLU function.The activation function chosen for the convolutional layer has a significant impact on the model’s performance. The impact of the activation function on recommendation accuracy and ranking quality is specifically shown in Figure 3. It can be seen that the evaluation indexes P@5, R@5, MAP, MRR, and NDCG perform best when the activation function is ReLU, and the effect of recommendation can reach the best degree, which are 0.425, 0.406, 0.499, 0.626, and 0.664, respectively.
The influence of activation function
In this section, some other mainstream recommendation models based on matrix decomposition, BPR, eALS, and MLP models, are selected for comparison. In the embedded feature representation process, the selection of potential dimensions will have a significant impact on the recommendation effect. While keeping other optimal hyperparameters unchanged at the same time, the influence of each potential dimension of the embedded information in the neural network input process on the recommendation effect is shown in Fig. 4. Figure (a) shows the effect of potential dimensions on MAP on the MovieLens dataset, while Figure (b) shows the effect of potential dimensions on MAP on the Pinterest1 dataset. On MovieLens, which is relatively denser, larger potential dimensions do not always lead to better model performance. The model’s optimal performance is achieved when the latent dimensions are chosen correctly, but it performs worse when the latent dimensions are larger. On the relatively less dense Pinterest1 dataset, each model requires more latent dimensions to achieve the best results. By analyzing the experiments on these two datasets, with relatively fewer potential dimensions, the models in this paper outperform the other comparison models experimentally.
The impact of potential dimensions on the recommendation effect
In this section, the deep neural network-based matrix decomposition personalized recommendation model constructed in this paper is compared with some other mainstream recommendation algorithms based on matrix decomposition, including BPR, eALS, MLP, and NeuMF models. The performance of this paper’s model and other models on the MovieLens dataset with varying number of predictors is specifically shown in Fig. 5. Where figure (a) shows the trend of the hit rate with the number of predictors and figure (b) plots the trend of the normalized discount cumulative gain with the number of predictors. From the figure, it can be found that the model of this paper improves the recommendation accuracy relative to other mainstream algorithm models. When the predictor is 8, the NDCG value of this paper’s model is 0.394, which is slightly lower than that of NeuMF, but in general, its recommendation performance is better than that of other algorithms.
Performance of the model
The recommendation performance performance of this paper’s model with other algorithmic models on the Pinterest dataset varying with the number of predictors is specifically shown in Fig. 6. Figure (a) shows the trend of the hit rate with the number of predictors, and Figure (b) plots the trend of the normalized discount cumulative gain with the number of predictors. The recommendation accuracy of this paper’s model is improved compared with other models, and the NDCG value of this paper’s model is always higher than other models, up to 0.559, regardless of the number of predictors.
The variation trend of HR and NDCG
This paper combines deep neural networks and matrix decomposition, and proposes a matrix decomposition personalized recommendation model based on deep neural networks to improve the accuracy and quality of personalized recommendations. Two public datasets, MovieLens and Pinterest1, are selected to carry out personalized recommendation experiments to test the personalized recommendation performance and performance of this paper’s model, and to explore the factors affecting the recommendation performance.
In the experiment on the effect of the number of MLP layers on recommendation accuracy, with the increase of the number of MLP hidden layer layers, the recommendation accuracy of the model in this paper is gradually improving, indicating that the increase of the number of neural network layers can improve the effect of the recommendation model. For this paper, the Adam algorithm is used for pre-training, and we compare and analyze the performance of the model in both training and untraining.The results prove that the HR and NDCG values of the pre-trained model have been improved by up to 4.12% and 5.7%, which obviously promotes the effect. As for the experiments on the influence of model parameters such as the number of convolution, output dimension, activation function, potential dimension and other model parameters on the recommendation accuracy, firstly, from the comprehensive consideration of various aspects such as recommendation accuracy and ranking quality, the performance of this paper’s model reaches the optimum when the number of convolution kernels is 32 and the output dimension is at 64. Secondly, among the three common activation functions, Sigmoid function, Tanh function and ReLU function, the model in this paper can achieve the best recommendation effect when the activation function is ReLU, and the evaluation indexes of P@5, R@5, MAP, MRR, and NDCG reach the optimum, which are 0.425, 0.406, 0.499, 0.626, and 0.664, respectively. Finally, in terms of potential dimension, when the potential dimension is larger, the performance of this paper’s model is worse, and the experimental results of this paper’s model are better compared with other models in the case of a relatively small number of potential dimensions. Obviously, both the number of MLP layers, pre-training, and model parameters have a significant impact on the recommendation performance of this paper’s model.
Other mainstream recommendation algorithms based on matrix decomposition, such as BPR, eALS, MLP, and NeuMF model, are selected as comparisons to further explore the comprehensive performance of the recommendation performance of this paper’s model. On the MovieLens dataset, the recommendation accuracy of this paper’s model is improved compared with other models, and the NDCG value is slightly lower than NeuMF only when the predictor is 8, and the overall recommendation performance is still better than other models. On the Pinterest dataset, the recommendation accuracy of this paper’s model is still better than that of other models, and the NDCG value is always higher than that of other models, reaching a maximum of 0.559.