A knowledge graph algorithm enabled deep recommendation system

Collaborative filtering recommendation algorithm

The collaborative filtering recommendation algorithm is a mainstream personalized recommendation algorithm that uses information about the common interests and experiences of groups of people to filter information through cooperative techniques among their agents, viewpoints, and data sources. The system then automatically recommends useful information to users, who filter and judge information according to their preferences.

User-based collaborative filtering algorithm

The user-based collaborative filtering algorithm (UBCF) mainly relies on three links. These include mining the historical interaction behavior data of users and goods, quantitatively evaluating and establishing a quantitative matrix; calculating the similarity of user behavior using Formula (1), and establishing a set:

(1) $$sim(u,v)={\displaystyle \frac{{\sum }_{s-1}^n{r}_{u,s}{r}_{v,s}}{\sqrt{{\sum }_{s-1}^n{r}_{u,s}^2}\sqrt{{\sum }_{s-1}^n{r}_{v,s}^2}}.}$$

Once the data are collected, the algorithm determines the products that exceed the threshold and push the items with ratings above a threshold to the targeted user. This method can perfectly exploit the hidden attributes among users. The recommendation model is shown in Fig. 2.

In Fig. 2, all users have browsed the same product, and according to the data association relationship; product 1 through product 4 have similar attributes, so it is clear that users have similar preferences for each other, then the product can be recommended to a specific user.

(2) $$sim(u,v)={\displaystyle \frac{{\sum }_{s-1}^n{r}_{u,s}{r}_{v,s}}{\sqrt{{\sum }_{s-1}^n{r}_{u,s}^2}\sqrt{{\sum }_{s-1}^n{r}_{v,s}^2}}.}$$

Model-based collaborative filtering algorithm

This algorithm is a learning-based algorithm, similar to a neural network algorithm, with the advantages of good robustness and a fast recommendation process. A parametric model is used to establish the association between users and products, and machine learning techniques are used to obtain a predicted list of recommended resources by establishing and modeling the objective function and successively optimizing the objective function. In general, the algorithm uses a combination of implicit semantics and artificial intelligence for modeling, builds a data matrix of information resources, and trains the data to obtain prediction data. The objective function is:

(3) $$\mathrm{L}\mathrm{o}\mathrm{s}\mathrm{s}=\sum _{i=1}^m{\sum _{j=1}^n(R_{ij}-P_i^T{Q}_j)}^2.$$

In Formula (3), the scoring matrix is R: $R=m\times n$; P is the lower dimensional user representation matrix; Q is the item representation matrix.

Collaborative filtering is an excellent recommendation algorithm. However, there are some drawbacks to its practical application. First, it relies too much on the score value; second, the low scalability of the data set sometimes affects the recommendation effect; finally, the lack of historical information is unfavorable to new users and leads to poor recommendation effect. In view of the existing problems, some scholars began to consider introducing other elements into the algorithm and optimizing it to compensate for its shortcomings.

Content-based recommendation algorithms

Content-based recommendation algorithms (CBR) processes item content to obtain corresponding personalized recommendation resources. The core idea is to perform semantic analysis and modeling based on the attribute features of items, and to recommend items related to the attribute features of entities. Compared with the collaborative filtering-based algorithm, this algorithm has some inherent advantages. Taking the recommended product attributes as input parameters, the algorithm automatically uses the machine learning function to obtain the interconnectedness of products, combines the biased data information of learners individually, and selects similar products with high matching degree to the learners. In this algorithm, the user’s evaluation score of the product is no longer the focus, but the matching degree between the learner and the product to be recommended is dynamically calculated around the product features, learners’ or users’ interests carried out in the product evaluation process. Common algorithms include neural networks, decision trees, and other methods, and their models are shown in Fig. 3.

FEU unit

The conversion of attribute relationships between items into the form of corresponding knowledge graph triads is a key step in the process of knowledge graph processing. The conversion process may result in the phenomenon of data attribute value loss and affect the accuracy of the recommendation. In data feature extraction, we optimize the traditional data processing mode of emphasizing macro data structures and ignoring the basic information of micro data, adopt multimodal thinking, classify the basic information of item attribute data, focus on the accuracy of knowledge graph structure information mining, strengthen the semantics of processing entities themselves, and extract the important data attributes with representative meaning for feature extraction. Sets of item attributes may be divided into three types: text, multi-value conforming attributes and other types of attributes. There are differences in the data processing methods.

Convolution is an important signal processing method widely used in fields such as image processing, speech recognition, and natural language processing, making it important for machine learning and artificial intelligence applications. Its main functions include feature extraction, dimensionality reduction, denoising, image enhancement, and indiscriminate compression of data to improve algorithm performance. Feature extraction is part of data processing and convolution can extract features in the signal, such as edges, textures, etc., through filters. This is very important for image classification and recognition tasks. Convolution is also important in dimension reduction as it can reduce the size of images through pooling operations, thereby reducing the dimensionality of data, which is very useful for processing large-scale image and text data. Denoising and convolution can remove noise from signals through filters, which is very common in the fields of signal processing and image processing and helps to improve the quality of data. Convolutional filtering is helpful for image processing because it can enhance images through filters, which can help improve the visual effect and quality of the image. Convolutional neural networks (CNN) are currently one of the most important models in deep learning to improve algorithm performance, with their basic structure being convolutional layers. Convolution can reduce the size of data through operations such as dimensionality reduction and filtering, thereby achieving data compression. This is very useful for processing large-scale data, achieving data storage and transmission.

Text type data has the characteristic of structural singularity, and in the process of text feature extraction, the convolutional network algorithm is introduced in Markov networks or recurrent neural networks and other methods to effectively reduce the iterative update parameters and achieve the best computation speed with a small amount of computation resources. Assuming that a word is considered as a feature vector of dimension k, a text segment containing m words is transformed into an m * k vector matrix, which is equivalent to the image mapping of vectors.

(11) $$\left[\begin{array}{c}{x}_{11}\cdots x_{1i}\cdots x_{1k}\\ x_{21}\cdots x_{3i}\cdots x_{3k}\\ ⋮\cdots ⋮\cdots ⋮\\ x_{m1}\cdots x_{mi}\cdots x_{mk}\end{array}\right]$$

The matrix x_i denotes a word, and for a text of length n can be expressed as: $T\mathrm{e}\mathrm{x}\mathrm{t}=x_{1:n}=x_1\oplus x_2\oplus x_3\oplus \cdots \oplus x_n$, and the i:i+m words are operated according to the formula to obtain the result of the convolution operation of the feature C_i.

(12) $$c_i=f(w\cdot x_{i:i+q-1}+b)$$

The results of C_i are processed with a convolution kernel to obtain the features $c\in R^{n-q+1}$ of the corresponding layer. The most valuable features $\hat{c}=\{c\}$ are captured using the maximum pooling operation, and regularization is performed using dropout to finally obtain the text feature vector $Vectou{r}_{Title}$ available for learning.

(13) $$Vectou{r}_{Title}=droupt(\hat{c})$$

Multi-valued type attributes are handled by transforming the enumerable attribute values by means of index matrix and one-hot encoding. Suppose the item has a multi-valued attribute Y and has m enumerable attribute values, and the m attribute values are represented as a continuous number, then the attribute Y can be represented as $[y_1,y_1,\cdots ,y_m]$, and the embedding matrix is indexed by a sequence of 1-m. For the i-th item V_i, the attribute values about its attribute Y can be represented as a d-dimensional vector, and the v_i attribute Y of can be represented as

(14) $$Vectou{r}_{mutil}(v_i)=y_{i1}\oplus y_{i2}\oplus \cdots \oplus y_{in}.$$

The vectorization results of text types and multivalued attributes of the same element are merged, fed into the fully connected layer, and processed using the activation function to form an initial vector v_init about the element.

(15) $$v_{init}=f(w1(Vecto{r}_{Title}\oplus y_{i2}Vecto{r}_{mutil})+b_1)$$where $Vecto{r}_{Title}$ represents the vectorized result of the text type attribute of the item, and $Vecto{r}_{mutil}$ represents the vectorized result of the multi-value type attribute of the item.

RS unit

The feature extraction unit performs the multimodal feature representation task and aggregates the vector results generated by multiple modalities as the initial vector of items in the subsequent recommendation unit to ensure that the important attribute value information is not lost. During the recommendation process, a vector matrix of users and projects is formed to provide input parameters for the cross-compression unit.

Assume that the attributes of a user can be represented as $U=[B_1,B_2,\cdots ,B_x]$, where Bi represents an attribute domain of the user. The attribute values of the user’s structured attributes are input to the embedding layer, and the user vector representation is obtained using Eq. (16) and the fully connected layer.

(16) $$u=u_{B1}\oplus u_{B1}\oplus u_{B2}\oplus \cdots \oplus u_{Bx}.$$

CCU unit

In the process of combining the knowledge graph with deep recommendation, CCU realizes the communication between the neighborhood graph embedding and semantic information embedding. The data from the two modules and the information contained in the entities they are associated with learn and improve each other. The recommended item data, and the triad data of the knowledge graph can be associated through this unit, and the triad vectorization and item vectorization results can be gradually improved through learning and training.

CCU is the linking module between V and E. For the potential features $v_L\in R^d$ with $e_l\in R^d$, the cross-feature matrix C_l representing the L-th level is constructed.

(17) $$C_l=v_l{e}_l^T=\left[\begin{array}{c}{v}_l^1{e}_l^d\cdots v_l^1{e}_l^d\\ ⋮\cdots ⋮\\ v_l^d{e}_l^d\cdots v_l^d{e}_l^d\end{array}\right].$$

Eq. (18) describes the crossover operation in the crossover compression cell, form a matrix projection to obtain the final VE feature vector. The cross-compression cell can finally be represented as

(18) $$[v_{l+1,}{e}_{l+1}]=C{v}_l,e_l.$$

KG unit

In the D-KGR model, the triplet of the knowledge map is taken as the input parameter of the model, and valuable head and tail feature representations ( $h_L$ and $t_L$) are extracted, respectively, through the processing links such as cross compression unit and multilayer perceptron, and the estimated $\hat{t}$ of the tail corresponding vector is obtained from Eq. (19) using the multi-layer neural network.

(19) $$\begin{array}{rl}& h_L=E_{v\sim s(h)}\left[C^L(v_{init},h)\left[e\right]\right]\\ & r_L=M^L(r)\\ & \hat{t}=M^k\left[\begin{array}{c}{h}_L\\ r_L\end{array}\right]\end{array}$$where rL: the initial feature of the relation, which is usually in a one-hot coding form. S(h) denotes the set of header relationships. T is the item ID corresponding to the head vector h in the data of the recommender system. C^L represents the corresponding cross-matrix as introduced in the previous section.

Finally, the predicted score of the model can be calculated using Eq. (20).

(20) $$f_{KG}(t,\hat{t})=t^T\times \hat{t}.$$

In practical applications, we can use deep learning methods that integrate knowledge graphs and convolutional neural networks to improve the application effect of recommended learning resources. In this application, the convolutional neural network is divided into two parts: user and book, as shown in Fig. 7.

Generally speaking, data are divided into three types: single valued, multi valued, and text. For different data types, differentiated data preprocessing methods are adopted. The data processing approach for single valued attributes is to encode users or books, use convolutional neural network technology to form an embedding vector matrix, and reconstruct the knowledge graph to obtain the feature vectors of users and books. The user, book ID, gender, age, and occupation of the user are encoded using Onehot and are input into a user book fully connected layer. For the types of books, the combinations can be diverse; for example, the books can be both engineering and philosophy. Assuming there are a total of 18 types of books, a digital index can be used to connect the feature matrix and input it into the fully connected layer. The processing of book names is quite special and requires a text convolutional network. Compose the embedding vector of each word into a matrix, and then use convolutional kernels of different serial port sizes to perform convolution on the matrix, usually using 2 × 2, 3 × 3, 5 × 5. Then, through the pooling layer and using dropout for regularization, the book name embedding vector is obtained.

The specific processing flow is:

1. Users and books are processed using convolutional neural networks to extract user features. Two fully connected layers are used, with the first layer embedding different features into the same dimensionality × 128 vectors.

2. Different features (f) are connected and input into the second fully connected layer, so that the feature vector and book feature network maintain the same dimensionality 1 × 200.

3. For the book feature network, the name features of the book are extracted using a text convolutional neural network, converting words into word vectors, as shown in Fig. 8.

4. Convolutional filtering is applied and used to filter h word windows to obtain feature vectors. Next, the maximum pooling layer is used to extract important feature sentence information and connect the various feature vectors of the book. Then the fully connected layer is used to obtain a dimension of 1 × 200 book feature vectors.

5. The results of the user and book network mentioned above are multiplied and normalized with an actual rating; the loss function is optimized and the feature vectors of the user and book are obtained.

By reconstructing the knowledge graph, effective semantic information can be obtained. By utilizing cross compression units, their functions act like a bridge between the neighborhood graph embedding module and the semantic information embedding module. Information sharing and enhancement can be carried out between the project embedding representation and the corresponding head entity, which can mine higher-order relationships and semantic information in relationships. Moreover, more accurate recommendations can be made through alternating learning patterns of the modules.

The integration of deep learning and knowledge graphs can fully combine and utilize auxiliary information, enrich the description of users and items, enhance the mining ability of recommendation algorithms, and effectively compensate for the sparsity or lack of interactive information. The introduction of convolutional neural network algorithms into the knowledge graph helps resolve the cold start problem. The knowledge graph utilizes its expandable characteristics to carry the user's own attribute information when a new user enters the platform, forming a customized user background. The proposed system is able to determine the corresponding interestes of new users without any preference bias and also recommend similar products to new users.