CN113342994A - Recommendation system based on non-sampling cooperative knowledge graph network - Google Patents

Abstract

The invention provides a recommendation system based on a non-sampling cooperative knowledge graph network, which comprises the following steps: the embedding module is configured to obtain an initial embedding vector of a triplet in the knowledge-graph; the non-sampling knowledge graph convolution module is set as a single-layer convolution network comprising a plurality of linear aggregators, and performs non-sampling pre-calculation on the initial embedded vector to obtain the deep level information of the triple; combining the embedded vector and the deep level information to serve as an updated embedded vector; the cooperation propagation module is set to simultaneously encode cooperation signals in the interaction of the user and the project as initial preference of the user and the project, and the cooperation signals are combined with the updated embedded vector to serve as an input vector of the prediction module; the prediction module is configured to obtain a recommendation based on the input vector. The invention realizes the performance which is not inferior to the depth model and higher speed only by designing a more complex propagation matrix and pre-calculation operation, and the result is predicted more accurately.

Description

Recommendation system based on non-sampling cooperative knowledge graph network

Technology neighborhood

The invention relates to a technical neighborhood of a recommendation method, in particular to a recommendation system based on a non-sampling cooperative knowledge graph network.

Background

The recommendation system provides commodity information and suggestions to customers by using an e-commerce website, helps the users decide what products should be purchased, and simulates salesmen to help the customers to complete the purchasing process. The personalized recommendation is to recommend information and commodities which are interested by the user to the user according to the interest characteristics and purchasing behaviors of the user. With the continuous expansion of the electronic commerce scale, the number and the variety of the commodities are rapidly increased, and customers need to spend a great deal of time to find the commodities which the customers want to buy. This process of browsing through large amounts of unrelated information and products will undoubtedly result in a constant loss of consumers who are overwhelmed by the problem of information overload. To address these issues, personalized recommendation systems have been developed. The personalized recommendation system is a high-level business intelligent platform established on the basis of mass data mining to help an e-commerce website to provide completely personalized decision support and information service for shopping of customers.

Knowledge Graph (KG) is an emerging Knowledge carrier, which integrates document data into an easily understood triple form and compensates data sparsity through deep semantic association between nodes. For example, (king homer, director, affei normal biography) indicates that king homer is the director of affei normal biography. KG is a directed heterogeneous graph, with nodes and edges corresponding to different types of entities and semantic relationships, respectively. Such graph structure means that KG has strong relationship representation capability and modeling flexibility, and has been successfully applied to many neighborhoods in recent years.

Many current representative studies on KG-binding recommendations have followed GNN-based technical routes. However, the introduction of GNNs also faces the following problems: (1) the exponentially increasing number of nodes in the information dissemination process results in significant memory and time costs. To alleviate this situation, existing approaches typically use sampling strategies to preserve a subset of the node neighbors or subgraphs during training to mitigate computational costs. However, the sampling operation may introduce errors in the optimization process. (2) The problems of gradient disappearance, characteristic smoothness and the like inherent in the neural network architecture of the depth map cause higher difficulty in model training. Although recent work has shown that these problems can be ameliorated to some extent, extensive experimentation has demonstrated that depth often does not lead to significant profitability. Balancing the depth and efficiency of the model, enabling the graph neural network to handle large-scale networks is a current stage challenge.

Disclosure of Invention

The invention aims to provide a knowledge graph recommendation system with simplicity and expressiveness, and the technical scheme is as follows:

a recommendation system based on a non-sampling cooperative knowledge graph network comprises an embedding module, a non-sampling knowledge graph convolution module, a cooperative propagation module and a prediction module which are sequentially connected;

the embedding module is configured to obtain initial embedding vectors of triples in the knowledge-graph;

the non-sampling knowledge graph convolution module is set as a single-layer convolution network comprising a plurality of linear aggregators, and performs non-sampling pre-calculation on the initial embedded vector to obtain the deep-level information of the triple; combining the embedded vector and the deep level information to serve as an updated embedded vector;

the collaboration propagation module is configured to encode collaboration signals in user and project interaction simultaneously as initial preferences of the user and project, and to combine with the updated embedded vector as an input vector to the prediction module;

the prediction module is configured to obtain a recommendation based on the input vector.

In some preferred embodiments, after obtaining the initial embedding vector, the embedding module further includes:

modeling the triples in the knowledge graph into an entity space and a relation space respectively, and evaluating the credibility of the triples according to the following formula:

wherein h and t are entities, r is the relationship existing between the entities h and t, e_h，e_r，e_tAre respectively the embedded representation of h, r, t, W_rA transformation matrix being a relation r;

a lower value of g (h, r, t) means a higher confidence in the triplet (h, r, t); conversely, the lower the confidence of the triplet (h, r, t).

In some preferred embodiments, the no-sample knowledge graph convolution module further comprises: an attention component, an information dissemination component, and a neighborhood aggregation component;

the attention component is configured to determine an attention parameter pi (h, r, t) of the single layer convolutional network through a relational attention mechanism;

the information dissemination component is configured to calculate an initial dissemination matrix B from the attention parameters pi (h, r, t)_i，j＝π(h_i，r，t_j) For the initial propagation matrix B_i，j＝π(h_i，r，t_j) Performing an exponentiation BⁿAcquiring neighborhood information within n hops of the entity; wherein h is_iIs the ith neighbor of the head entity h; t is t_jIs the jth neighbor of the tail entity t;

the neighborhood aggregation component is configured to obtain deep level information for the triplet by using different sized linear aggregators in a single convolutional layer to achieve a non-sampling pre-computation.

In some preferred embodiments, the method for the collaboration propagation module to encode the collaboration signal in the user and project interaction as the initial preference of the user comprises:

aligning a related item set in the user historical interaction with an entity in a knowledge graph, and converting the related item set into a feature set E calculated in the knowledge graph_u：E_u＝{E_e|(v，e)∈A，v∈{v|y_uv1} }; wherein a { (V, E) | V ∈ V, E ∈ E } represents that there is a set of mappings, (V, E) indicates that the item V can align with an entity E in the knowledge graph; y is_uvFor user feedback of parameters, y_uv1 indicates that there is feedback behavior between the user and the item, otherwise y_uv＝0；

Set E of user characteristics_uCarrying out normalization processing to obtain:

in some preferred embodiments, the method for the collaboration propagation module to encode the collaboration signal in the user and project interaction as an initial preference for the project comprises:

acquiring other interactive projects of the user set interactive with the target project V as a collaborative project set V of the target project V_v，

Wherein the content of the first and second substances,

in order to feed back the parameters to the user,

representing user u and item v_uThere is an interaction between them;

set V of collaborative projects_vAligning with the entity in the knowledge graph to obtain the feature set E of the target item v_v，E_v＝{E_e|(v_u，e)∈A，v_u∈V_v}；

Normalizing the initial set of items, and adding the characteristics of the alignment entities of the items to obtain:

advantageous effects

1. In the non-sampling knowledge transmission module, a single graph convolution layer is used for aggregating neighborhood information from multiple layers, and because input parameters are fixed, the step can be directly pre-calculated, neighborhood sampling is not carried out, but all neighborhood information is considered, and errors generated by sampling are avoided;

2. the method for stacking a plurality of graph convolutional layers is different, the model only uses one convolutional layer, but the performance which is not inferior to that of a depth model is realized by designing a more complex propagation matrix and pre-calculation operation, and the speed is better;

3. in the cooperation propagation module, key cooperation signals in user interaction are coded into preferences of users and projects and combined with KG embedding, and finally obtained vectors fully utilize the two key information, so that potential semantics of the users and the projects in a vector space are more effectively represented.

Drawings

FIG. 1 is a simplified schematic diagram of a system architecture in accordance with an embodiment of the present invention;

FIG. 2 is a detailed schematic diagram of a system architecture in another embodiment of the present invention;

FIG. 3 is a comparison graph of predicted results based on last. FM data set in a preferred embodiment of the present invention;

FIG. 4 is a comparison graph of predicted results based on a Book-Cross dataset in a preferred embodiment of the present invention;

FIG. 5 is a comparison graph of predicted results based on MoiveLens dataset in a preferred embodiment of the present invention;

Detailed Description

In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is further described with reference to the accompanying drawings. In the description of the present invention, it is to be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", and the like, indicate orientations or positional relationships based on the orientations or positional relationships shown in the drawings, are merely for convenience in describing the present invention and simplifying the description, and do not indicate or imply that the device or element being referred to must have a particular orientation, be constructed and operated in a particular orientation, and thus, should not be construed as limiting the present invention.

The invention provides a recommendation system based on a non-sampling cooperative knowledge graph network, which comprises an embedding module, a non-sampling knowledge graph convolution module, a cooperative propagation module and a prediction module which are sequentially connected as shown in figure 1; it should be understood that the connection described in the present invention is not specifically illustrated, and the specific connection manner of the connection may be at least one of a wired connection or a wireless connection.

The embedding module is configured to obtain initial embedding vectors for triples in the knowledge-graph. The definition of the triplet of the invention is as follows: given a project knowledge graph G { (h, R, t) | h, t ∈ E, R ∈ R }, where each triplet represents the existence of a relationship R between entity h and entity t, E and R corresponding to the entity and the set of relationships, respectively. For example, the triplet (liu phenanthrene, actor, magnolia) states the fact that liu phenanthrene is the actor of the movie "magnolia flowers". In an actual recommendation scenario, item V may have a mapping relationship with one or more entities in G. For example, the book "Aoxi & prejudice" is synonymous with one entity in KG, while news titled "Lioflavay attends the first ceremony of Magnolia" is related to multiple entities of "Lioflavay" and "Magnolia". A { (V, E) | V ∈ V, E ∈ E } is used to represent a set where there is a mapping, where (V, E) indicates that the item V can be aligned with entity E in the knowledge graph.

Those skilled in the art will appreciate that there are many ways to obtain the embedded vector, which aims to convert the entities and relationships in the knowledge graph into a low-dimensional vector representation. In some preferred embodiments of the present invention, a Translate model may be used to obtain the embedded vector, including methods of TransE, TransH, TransR, TransD, etc. The following application provides an embedded vector acquisition method based on TransR:

considering that entities corresponding to the same relation often have information of different levels, the entities and the relations in the knowledge graph are modeled into two different spaces of the relations and the entities respectively, and the credibility of the two spaces is evaluated by using a credibility scoring function:

wherein e_h,e_r,e_tAre respectively the embedded representation of h, r, t, W_rA transformation matrix of relation r. A lower value of g (h, r, t) means a higher confidence in the triplet; conversely, the less trustworthy the triplet.

The non-sampling knowledge graph convolution module is set as a single-layer convolution network comprising a plurality of linear aggregators, and performs non-sampling pre-calculation on the initial embedded vector to obtain the deep-level information of the triple; and combining the embedded vector and the deep level information to serve as an updated embedded vector.

In which the present invention performs pre-computation only by a single convolutional layer, and achieves performance comparable to deep networks by designing a more efficient propagation matrix and considering all neighbor information, although only shallow networks are used. The pre-calculation without sampling specifically includes: vector values after convolution are calculated in advance, and because the input of the convolution layer is a fixed value and only one convolution layer is provided, the final value can be obtained by directly calculating without training and updating parameters. The deep information of the triples refers to other information related to the triples besides the information of the triples, and the partial information is not easy to directly extract in a conventional mode. In some preferred embodiments, the deep level information includes neighbor information of triples.

In other preferred embodiments, in consideration of the difference in importance of connectivity between entities, a composition structure of a non-sampling knowledge graph convolution module is provided, which includes: an attention component, an information dissemination component, and a neighborhood aggregation component;

the attention component is configured to determine an attention parameter pi (h, r, t) of the single layer convolutional network via a relational attention mechanism. There are many methods for calculating the attention parameter, and the invention provides an example of calculation:

the attention parameter pi (h, r, t) is calculated using a nonlinear activation function tanh: pi (h, r, t) ═ W_re_t)^Ttanh(W_re_h+e_t)；

From this, it can be seen that the attention score is given by e in the relationship space_hAnd e_tIs determined.

Then, the coefficient normalization processing of the whole triple is carried out through a softmax activation function:

wherein N is_hA triple set taking an entity h as a head entity; r 'and t' are other relationships and tail entities in the triple set with the entity h as a head entity.

In this application, the scheme of the present invention only requires that the attention parameters be pre-determined by a small subset of training images and then proceed to the next step, in order not to disrupt the efficient pre-computation of the graph aggregation operation.

The information dissemination component is configured to calculate an initial dissemination matrix B from the attention parameters pi (h, r, t)_i，j＝π(h_i，r，t_j) For the initial propagation matrix B_i，j＝π(h_i，r，t_j) Performing an exponentiation BⁿAcquiring neighborhood information within n hops of the entity; wherein h is_iIs the ith neighbor of the head entity h; t is t_jIs the jth neighbor of the tail entity t.

Because the entity and the neighbor in the knowledge graph have different degrees of relevance, in order to effectively expand the potential preference of users and projects, the invention considers the high-order neighbor information of the node in the propagation process. In some preferred embodiments, the maximum value of the number of hops n is set to 3 in order to balance the integrity of the physical adjacency information with the time required to compute the adjacency matrix.

The neighborhood aggregation component is configured to obtain deep level information for the triplet by using different sized linear aggregators in a single convolutional layer to achieve a non-sampling pre-computation.

The linear aggregator is a continuous function that can compute multiple sets of neighbor node information. In the common scheme of the present neighborhood, an aggregation method, such as mean, sum, max, etc., is mostly used, and such a method cannot distinguish the neighborhood messages. Or using a combination of aggregation methods, such as using mean, maximum, minimum, and standard deviation, etc., but again this approach is not sufficient to accurately describe the context of neighborhood information. The invention proposes to use linear aggregators of different sizes to implement non-sampling pre-computation, and gives a specific example:

in the linear aggregator AX, propagation matrices (A) of different power series are set₀＝B⁰,A₁＝B,A₂＝B²,...,A_n＝Bⁿ) And connecting them. This idea is similar to the initial module in convolutional neural networks (combining convolutional kernels of different sizes on the same convolutional layer). Since AX can be pre-computed, the method considers all neighbor information in the propagation process without selective sampling. The concrete formula is as follows:

wherein E is a knowledge map feature set; x is an initial node characteristic matrix; i is a predetermined power series.

The collaboration propagation module is configured to encode collaboration signals in user and project interaction simultaneously as initial preferences of the user and project, and to combine with the updated embedded vector as an input vector to the prediction module;

unlike the traditional recommendation algorithm which uses independent potential vectors, the invention simultaneously obtains the initial preferences of the users and the projects in the collaborative propagation module so as to obtain the extended preferences of the users and the projects by combining with knowledge embedding. Intuitively, the user's historical interaction items can indicate the user's preferences to some extent. The set of related items in the user's historical interactions is converted into a set of features computed in the knowledge graph by aligning the set of related items with the entities in the knowledge graph.

In some preferred embodiments, a preferred method is provided in which the collaboration propagation module encodes collaboration signals in user and project interactions as initial preferences of a user:

aligning a related item set in the user historical interaction with an entity in a knowledge graph, and converting the related item set into a feature set E calculated in the knowledge graph_u：E_u＝{E_e|(v，e)∈A，v∈{v|y_uv1} }; wherein a { (V, E) | V ∈ V, E ∈ E } represents that there is a set of mappings, (V, E) indicates that the item V can align with an entity E in the knowledge graph; y is_uvFor user feedback of parameters, y_uv1 indicates that there is feedback behavior between the user and the item, otherwise y_uv＝0；

Set E of user characteristics_uCarrying out normalization processing to obtain:

in other preferred embodiments, a similar preferred method of encoding collaboration signals in user and project interactions as initial preferences for a project is presented:

acquiring other interactive projects of the user set interactive with the target project V as a collaborative project set V of the target project V_v，

Wherein the content of the first and second substances,

in order to feed back the parameters to the user,

representing user u and item v_uThere is an interaction between them; where a collection of users that can be considered to have interacted with the same item, their similar behavioral preferences can also describe the underlying representation of the item.

Set V of collaborative projects_vAligning with the entity in the knowledge graph to obtain the feature set E of the target item v_v，E_v＝{E_e|(v_u，e)∈A，v_u∈V_v}；

Normalizing the initial set of items, and adding the characteristics of the alignment entities of the items to obtain:

the prediction module is configured to obtain a recommendation based on the input vector. Those skilled in the art should understand that there are many recommendation algorithms for predicting the input and obtaining the recommendation result in the recommendation system neighborhood, and the recommendation algorithms are not described herein since they are not the key points for the protection of the present invention. For the completeness of the whole invention scheme, the invention provides an example of obtaining a recommendation result by using a prediction function and evaluating the prediction precision by using a loss function, which comprises the following specific steps:

defining the interaction probability of the output user u and the item v as

The prediction function is then:

wherein, theta is a model parameter which can be learned; further, the output prediction score is recorded as

The true prediction score is recorded as y_uv. Will y_uvAnd

the cross entropy loss function of (A) is recorded as

The formula is as follows:

the cross entropy loss function J can reflect the distance between the prediction score and the real score, and can evaluate the performance of the model more accurately.

Examples

The invention provides a recommendation system based on a non-sampling collaborative knowledge graph network, and as shown in fig. 2, the embodiment also provides a method for evaluating model performance aiming at three real data sets of music, books and movies by comparing the technical scheme of the invention with the prior art. For convenience of brief description, a Non-Sampling Collaborative Knowledge Graph Network (Non-Sampling Collaborative Knowledge Graph Network) proposed by the present invention is abbreviated as NCKN.

In this example, the model performance was evaluated using the following three real data sets: fm (Music), Book-cross (Book), MovieLens-20m (movie), as described in table 1, relevant statistical information is given. All three data sets allow public access and vary in size and sparsity.

(1) Fm.fm user song listening behavior and project knowledge provided by the last.fm online music system.

(2) Book-Cross-statistical reader scoring data from the Book community (0 to 10 unequal).

(3) MovieLens-20M is a test data set widely used in the movie recommendation neighborhood, and the file contains feedback information on movie websites, i.e. the user's explicit scores (varying from 1 to 5) for each movie.

Table 1 experimental data set statistics

In view of the fact that implicit feedback can provide richer interactive content, which is beneficial for alleviating the cold start problem, we first convert explicit feedback into implicit feedback in the data preprocessing part. Where 1 represents a sample of the user's positive rating and 0 is a negative sample randomly sampled from the non-interactive set. FM and boost-Crossing interactive data are sparse, so that a threshold value is not set, and a MovieLens-20M front face scoring threshold value is set to be 4.

In addition to pre-processing the interaction data of the user and the project, the present embodiment generates a project knowledge graph for each dataset in microsoft satori. Specifically, triples with confidence higher than 0.9 are first extracted from the entire KG as children KGs. For a certain child KGs, all valid entity ids are collected by matching the names of the head and tail nodes. Finally, the item id is mapped to an entity in KG and the corresponding triplet set is matched in child KGs. Note that to simplify the overall process, we will exclude items where there is no match or multiple matches.

The comparison algorithm adopted by the embodiment comprises the following steps:

BPRMF: a classical CF method optimized using matrix decomposition.

CKE: and the CF and the multiple knowledge graphs are fused for training, and the feature embedding of the structural information, the text information and the visual information in the project knowledge graph is respectively extracted. This embodiment combines only structural knowledge with CF.

PER: and introducing meta-paths to represent the connectivity of the users and the items in different relation paths by utilizing the relation heterogeneity in the project knowledge graph, and recommending the items based on the path similarity. The present embodiment defines the meta path as item-attribute-item attribute.

RippleNet: recently proposed preference propagation based models. By taking the historical user interaction items as an initial set in KG propagation, diffusion is carried out in KG and multi-layer neighbor information is aggregated, so that deeper potential user preference representation is obtained.

KGCN: the most advanced model fusing KG and graph convolution neural network obtains rich item embedding from neighbors of a knowledge graph by utilizing graph convolution, and therefore the recommendation task is greatly improved.

KGAT: and is also the most advanced model of the fused graph convolution network. It combines the project knowledge graph and the user interaction data into a collaborative knowledge graph and recursively propagates neighbors over the graph structure to update the target node's embedding. In addition, an attention mechanism is used during propagation to distinguish the importance of neighboring nodes.

The experimental setup for this example is as follows:

for each data set, the data set is randomly divided into a training set, a testing set and a verification set according to the proportion of 7:2: 1. The present embodiment evaluates in the following two recommendation scenarios: (1) Click-to-Tthrough Rate (CTR), the probability of interaction between a particular user and an item is predicted in a trained recommendation model. (2) And (4) top-k recommendation, namely selecting k items with highest user prediction probability in the test set by using a recommendation model learned from the training set. To verify the effectiveness of these methods, the present example applies the following evaluation indexes:

precision: the model recommends the accuracy of the project. Where R (u) is a list of items recommended to the user according to the training set and T (u) is a list of items recommended to the user according to the testing set.

Recall: hit rate of the candidate recommendation list.

F1: the combination of Precision and Recall weighting makes the F1 value more reflective of the model performance.

AUC: for evaluating the performance of the recommendation system in distinguishing between favorite and disliked items of a user. a is the commodity liked by the user, b is the commodity disliked by the user, the scoring of a and b by the recommendation system is compared each time, n is the total number of comparison, n 'is the number of times that the score of a is greater than that of b, n' is the number of times that the score of a is equal to that of b, and the calculation mode of AUC is as follows.

This embodiment was programmed in the environment of pytorch1.3.0 and the parameters of all comparison algorithms were adjusted. The learning rate is adjusted in [10-3,5 x 10-3,10-2,5 x 10-2] and the embedded dimension size is adjusted in [8,16,32,64,128,256 ].

And (3) comparing experimental results:

this example gives the performance of all methods in three datasets, with the predicted results in CTR and top-k as shown in table 2 below and in figures 3-5, respectively. The results of the experiment are now analyzed as follows:

TABLE 2 CTR prediction based on AUC and F1 metrics

The KG-based recommendation method is far superior to the CF-based method, which indicates that the introduction of additional information in KG greatly improves the recommendation method. However, the performance of BPRMF exceeds CKE on individual indices, suggesting that modeling only the first order relationship in KG may not be able to fully exploit the role of KG. This also verifies the validity of NCKN aggregation from multi-layer higher-order neighbor information.

KGAT achieves significant results in Recall @ k, performing best especially in music and book datasets, but it is noted that in 20M movie datasets, the model NCKN performance of this embodiment outperforms KGAT. Our inference is that KGAT makes more accurate predictions when the data set is small in size and sparse. However, for large-scale and more information-dense movie data sets, the high-order propagation of KGAT in the user interaction graph introduces excessive noise, and NCKN uses a first-order collaboration signal in combination with KG to achieve better results.

The performance ranking of all the methods on the three data sets was found by observation to be movies, music, books, respectively. This may be due to the difference between the average number of user interactions on the three data sets and the average number of links to the entities in the KG. For example, a movie data set has a greater number of interactive behaviors and relational links than a music data set and a book data set, and its rich information allows the recommendation model to learn potential feature representations more accurately.

Our non-sampling collaborative knowledge graph model achieves a competitive advantage over all three datasets compared to all methods. Specifically, the average increase in CTR prediction was 1.2%, 2.3%, 1.5% over other methods. Note that NCKN does not perform well on music data sets but second only to KGAT because the average number of KG links in a music data set is too low and the no-sampling strategy in NCKN does not work its best. In top-k, NCKN performance is excellent, especially performing best in the movie data set, and demonstrates the positive significance of NCKN non-sampling and cooperative propagation compared to rippenet.

The foregoing shows and describes the general principles, essential features, and advantages of the invention. It will be understood by those skilled in the art that the present invention is not limited to the embodiments described above, which are described in the specification and illustrated only to illustrate the principle of the present invention, but that various changes and modifications may be made therein without departing from the spirit and scope of the present invention, which fall within the scope of the invention as claimed. The scope of the invention is defined by the appended claims and equivalents thereof.

Claims (5)

1. A recommendation system based on a non-sampling collaborative knowledge graph network is characterized in that: the system comprises an embedding module, a non-sampling knowledge graph convolution module, a cooperative propagation module and a prediction module which are connected in sequence;

the embedding module is configured to obtain initial embedding vectors of triples in the knowledge-graph;

the non-sampling knowledge graph convolution module is set as a single-layer convolution network comprising a plurality of linear aggregators, and performs non-sampling pre-calculation on the initial embedded vector to obtain the deep-level information of the triple; combining the embedded vector and the deep level information to serve as an updated embedded vector;

the collaboration propagation module is configured to encode collaboration signals in user and project interaction simultaneously as initial preferences of the user and project, and to combine with the updated embedded vector as an input vector to the prediction module;

the prediction module is configured to obtain a recommendation based on the input vector.

2. The non-sampling collaborative knowledge graph network based recommendation system of claim 1, wherein the embedding module, after obtaining the initial embedding vector, further comprises the steps of:

modeling the triples in the knowledge graph into an entity space and a relation space respectively, and evaluating the credibility of the triples according to the following formula:

wherein h and t are entities, r is the relationship existing between the entities h and t, e_h，e_r，e_tAre respectively the embedded representation of h, r, t, W_rA transformation matrix being a relation r;

a lower value of g (h, r, t) means a higher confidence in the triplet (h, r, t); conversely, the lower the confidence of the triplet (h, r, t).

3. The non-sampling collaborative knowledge graph network based recommendation system of claim 2 wherein the non-sampling knowledge graph convolution module further comprises: an attention component, an information dissemination component, and a neighborhood aggregation component;

the attention component is configured to determine an attention parameter pi (h, r, t) of the single layer convolutional network through a relational attention mechanism;

the information dissemination component is configured to calculate an initial dissemination matrix B from the attention parameters pi (h, r, t)_i，j＝π(h_i，r，t_j) For the initial propagation matrix B_i，j＝π(h_i，r，t_j) Performing an exponentiation BⁿAcquiring neighborhood information within n hops of the entity; wherein h is_iIs the ith neighbor of the head entity h; t is t_jIs the jth neighbor of the tail entity t;

the neighborhood aggregation component is configured to obtain deep level information for the triplet by using different sized linear aggregators in a single convolutional layer to achieve a non-sampling pre-computation.

4. The non-sampling collaborative knowledge graph network based recommendation system of claim 1 wherein the method of the collaborative propagation module encoding collaboration signals in user and project interactions as initial preferences of a user comprises:

aligning a related item set in the user historical interaction with an entity in a knowledge graph, and converting the related item set into a feature set E calculated in the knowledge graph_u：E_u＝{E_e|(v，e)∈A，v∈{v|y_uv1} }; wherein a { (V, E) | V ∈ V, E ∈ E } represents that there is a set of mappings, (V, E) indicates that the item V can align with an entity E in the knowledge graph; y is_uvFor user feedback of parameters, y_uv1 indicates that there is feedback behavior between the user and the item, otherwise y_uu＝0；

Set E of user characteristics_uCarrying out normalization processing to obtain:

5. the non-sampling collaborative knowledge graph network based recommendation system according to claim 1 or 4, wherein the method of the collaborative propagation module encoding collaborative signals in user and project interactions as initial preferences of projects comprises:

acquiring other interactive projects of the user set interactive with the target project u as a collaborative project set V of the target project u_v，

Wherein the content of the first and second substances,

in order to feed back the parameters to the user,

representing user u and item v_uThere is an interaction between them;

set V of collaborative projects_vAligning with the entity in the knowledge graph to obtain the feature set E of the target item v_v，E_u＝{E_e|(v_u，e)∈A，v_u∈V_v}；

Normalizing the initial set of items, and adding the characteristics of the alignment entities of the items to obtain:

Images