A Location-Aware Conversational News Recommendation Method and System Based on Knowledge Graph

Abstract

The present invention proposes a method and system for location-aware conversational news recommendation based on a knowledge graph, which relates to the field of news recommendation. It includes: data collation; obtaining text features; obtaining geographical location features; constructing a preference extraction network based on GRU to obtain the user's interest preferences and geographical location preferences in a set session; constructing a feature matrix, using a multi-head attention mechanism to interactively learn the feature matrix to obtain the feature matrix after feature interaction; constructing a predictor to obtain the probability that the user clicks on a candidate news. The present invention uses a knowledge graph representation learning algorithm to learn the embedded representation of geographical locations, making the geographical location information rich and accurate; uses a gated recurrent neural network to accurately extract interest features and geographical location preference features, constructs a feature interaction unit based on multi-head attention to realize the interaction between features, and finally realizes the improvement of the performance of conversational news recommendation.

Description

Technical Field

[0001] The present invention belongs to the technical field of news recommendation, and in particular, relates to a location-aware conversational news recommendation method and system based on a knowledge graph. Background Art

[0002] The statements in this section merely provide background technical information related to the present invention and do not necessarily constitute prior art.

[0003] Recently, with the development of the Internet, more and more information floods into people's lives. The problem of information overload makes it difficult for people to obtain the information they need from a large amount of information. As an important tool to solve the problem of information overload, personalized recommendation systems are widely used in various aspects of real life, such as Taobao product recommendations, Douyin video recommendations, news recommendations, and so on. News recommendation is a very important application of personalized recommendation systems, and its purpose is to recommend a set of news that users are interested in. The purpose of conversational news recommendation is to recommend the news that the user is interested in at the next moment by means of the short-term preferences shown by the user when reading news in a short time. Existing conversational news recommendation works often use natural language processing models to process news text information, then input the news sequence into a recurrent neural network (RNN) to learn the short-term preferences of users, and finally perform news recommendations by means of the similarity between user features and news features.

[0004] Existing conversational news recommendation methods extract news text features very fully. However, it is far from enough to only use the text features of news to mine the short-term preferences of users in the short conversation of users reading news. News is not only closely related to news text, but also closely related to the location where it occurs. For example, when a major event occurs in a certain place, this piece of news may attract the continuous attention of users. The news that users read at different locations is also different. When users are in different geographical locations, they often have different reading preferences. For example, in the office, people tend to like reading economic and political news. However, at home, people tend to like reading entertainment and sports news. Therefore, geographical location information plays a very important role in mining users' reading preferences. It is very important to fully explore the relationship between news content, news geographical location, and user geographical location to improve the performance of conversational news recommendation. Summary of the Invention

[0005] To overcome the deficiencies of the above-mentioned prior art, the present invention provides a location-aware conversational news recommendation method and system based on a knowledge graph, which fully explores the relationship between news content, geographical location, and user geographical location, and improves the performance of the conversational news recommendation system.

[0006] To achieve the above object, one or more embodiments of the present invention provide the following technical solutions:

[0007] In the first aspect of the present invention, a location-aware conversational news recommendation method based on a knowledge graph is provided;

[0008] A location-aware conversational news recommendation method based on a knowledge graph includes the following steps:

[0009] Process the news dataset to obtain a news text set, a set of geographical locations related to each news, a session set, and a session geographical location set;

[0010] Extract features from each news text in the news text set respectively to obtain a news text feature set;

[0011] Construct a geographical location knowledge graph, extract the features of each geographical location related to the news in the knowledge graph to obtain a news geographical location feature set, and based on the news geographical location feature set and the session geographical location set, obtain a session geographical location feature set;

[0012] Construct a preference extraction network based on GRU, and based on the text features and news geographical location features of each news in the set session read by the user, obtain the interest preference and geographical location preference of the user in the set session;

[0013] Combine the text features and geographical location features of the candidate news, the interest preference and geographical location preference of the user in the set session, and the user geographical location features into a feature matrix, and use the multi-head attention mechanism to perform interactive learning on the feature matrix to obtain a feature matrix after feature interaction;

[0014] Based on the feature matrix after feature interaction, obtain the probability that the user clicks on the candidate news, and recommend news with a high click-through rate to the user.

[0015] Preferably, use a pre-trained Bert model to extract features from each news text in the news text set respectively, and combine the text features of all news to obtain a news text feature set.

[0016] Preferably, obtain all geographical locations related to the news dataset, combine the existing knowledge graph, construct a geographical location knowledge graph by means of entity connection, use the knowledge graph representation learning method to learn the embedding representation of the location entity of each geographical location in the knowledge graph as the feature of each geographical location, and fuse the features of multiple geographical locations corresponding to each news in the dataset into the geographical location feature of this news, and combine the geographical location features of all news in the dataset to obtain a news geographical location feature set.

[0017] Preferably, the GRU-based preference extraction network consists of multiple GRU computing units. The text features and geographical location features of the set session read by the user are respectively input into the interest preference extraction network and the geographical location preference extraction network based on GRU to obtain the interest preference and geographical location preference of the user in the set session.

[0018] Preferably, the input of each GRU computing unit includes the hidden state at the previous moment, the text features or geographical location features of the news read at the current moment, and the hidden state output by the last GRU computing unit is the interest preference or geographical location preference of the user in the set session.

[0019] Preferably, before using the multi-head attention mechanism for interactive learning of the feature matrix, the candidate news text features and geographical location features, the interest preference and geographical location preference of the user in the set session, and the user geographical location features are mapped to the same dimension, and the mapped features are combined into a feature matrix.

[0020] Preferably, the multi-head attention mechanism consists of multiple self-attention units, and each self-attention unit is responsible for feature interaction in one feature space. The process of using a single self-attention unit for feature interaction calculation is as follows:

[0021]

[0022] Q ua =E ua W Q

[0023] K ua =E ua W K

[0024] U ua =E ua W U

[0025] Among them, E ua is the feature matrix, is the feature matrix after feature interaction, are the query, key, and value matrices respectively, is the weight matrix of the self-attention unit, d att =d map is the mapped feature dimension, is the regularization factor.

[0026] The second aspect of the present invention provides a location-aware conversational news recommendation system based on a knowledge graph.

[0027] A location-aware conversational news recommendation system based on a knowledge graph, comprising:

[0028] A data sorting module, configured to: process a news data set to obtain a news text set, a set of geographical locations related to each news, a session set, and a session geographical location set;

[0029] A text feature extraction module, configured to: perform feature extraction on each news text in the news text set respectively to obtain a news text feature set;

[0030] A geographical location feature extraction module, configured to: construct a geographical location knowledge graph, extract the features of each geographical location related to the news in the knowledge graph to obtain a news geographical location feature set, and based on the news geographical location feature set and the session geographical location set, obtain a session geographical location feature set;

[0031] A user preference extraction module, configured to: construct a preference extraction network based on GRU, and based on the text features and news geographical location features of each news in the set session read by the user, obtain the interest preference and geographical location preference of the user in the set session;

[0032] An interaction learning module, configured to: combine the text features and geographical location features of the candidate news, the interest preference and geographical location preference of the user in the set session, and the user geographical location features into a feature matrix, and use a multi-head attention mechanism to perform interaction learning on the feature matrix to obtain a feature matrix after feature interaction;

[0033] A click-through rate prediction module, configured to: based on the feature matrix after feature interaction, obtain the probability that the user clicks on the candidate news, and recommend news with a high click-through rate to the user.

[0034] The third aspect of the present invention provides a computer-readable storage medium, on which a program is stored, and when the program is executed by a processor, the steps in the location-aware conversational news recommendation method based on a knowledge graph as described in the first aspect of the present invention are implemented.

[0035] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored on the memory and executable on the processor. When the processor executes the program, the steps in the location-aware conversational news recommendation method based on a knowledge graph as described in the first aspect of the present invention are implemented.

[0036] The above one or more technical solutions have the following beneficial effects:

[0037] 1. The present invention constructs a geographical location information knowledge graph, realizes the complement of news geographical location information, describes news features more comprehensively, and thus can learn more accurate user geographical location preferences, making the learning of user preferences more comprehensive.

[0038] 2. The present invention uses a gated recurrent neural network to separately learn the short-term interest preferences and geographical location preferences of users, more fully mining the preference features of users and making user features richer.

[0039] 3. The present invention constructs a feature interaction unit based on multi-head attention, realizing the interaction of user interest preferences, geographical location preferences, user location information features, news text features, and news geographical location features, mining the relationships between features, and making the final features more accurate.

[0040] Advantages of additional aspects of the present invention will be partly given in the following description, partly will become obvious from the following description, or will be understood through the practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] The accompanying drawings forming a part of this specification are used to provide a further understanding of the present invention. The schematic embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation to the present invention.

[0042] Figure 1 It is a flowchart of the method for the first embodiment.

[0043] Figure 2 It is a structure diagram of the prediction period for the first embodiment.

[0044] Figure 3 It is a system structure diagram for the second embodiment. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] It should be noted that the following detailed description is exemplary and is intended to provide further illustration of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention belongs.

[0046] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention.

[0047] In the case of no conflict, the embodiments in the present invention and the features in the embodiments can be combined with each other.

[0048] The general idea proposed by the present invention:

[0049] First, in order to make the geographical location information more abundant and accurate, the present invention constructs a geographical location information knowledge graph by combining the geographical location information involved in users and news with Wikidata, and uses a knowledge graph representation learning algorithm to learn the embedded representation of geographical locations. Then, in order to more accurately extract the interest features from news texts and the geographical location preference features from geographical location embeddings, considering that the two parts of features have a certain degree of independence, two gated recurrent neural networks are used to separately process and extract the user's interest preferences and geographical location preferences from the text features of news and the geographical location embedding features respectively.

[0050] Meanwhile, in order to realize the interaction between various features of users and various features of candidate news, the present invention constructs a feature interaction unit based on multi-head attention. For a user, the feature combination after interaction by the user feature interaction unit obtains the user's final short-term preference features; for a candidate news, its features consist of its text features and geographical location features. After interaction by the news feature interaction unit, the feature matrix is obtained by combination.

[0051] Finally, the above feature matrix is input into a click-through rate predictor based on a multi-layer neural network to predict the click-through rate.

[0052] Embodiment 1

[0053] This embodiment discloses a location-aware conversational news recommendation method based on a knowledge graph;

[0054] As Figure 1 shown, the location-aware conversational news recommendation method based on a knowledge graph includes:

[0055] A location-aware conversational news recommendation method based on a knowledge graph includes the following steps:

[0056] Process the news dataset to obtain a news text set, a set of geographical locations related to each news, a session set, and a session geographical location set;

[0057] Extract features from each news text in the news text set respectively to obtain a news text feature set;

[0058] Construct a geographical location knowledge graph, extract the features of each geographical location related to the news in the knowledge graph to obtain a news geographical location feature set, and based on the news geographical location feature set and the session geographical location set, obtain a session geographical location feature set;

[0059] Construct a preference extraction network based on GRU, and based on the text features and news geographical location features of each news in the set session read by the user, obtain the interest preferences and geographical location preferences of the user in the set session;

[0060] Combine the text features and geographical location features of candidate news, the interest preferences and geographical location preferences of the user in the set session, and the user geographical location features into a feature matrix, and use the multi-head attention mechanism to perform interactive learning on the feature matrix to obtain the feature matrix after feature interaction;

[0061] Based on the feature matrix after feature interaction, obtain the probability that the user clicks on the candidate news, and recommend the news with a high click-through rate to the user.

[0062] Specifically, it includes the following steps:

[0063] 1) Data collation

[0064] Taking the Adressa news dataset as an example for data collation, this dataset was collected by the Norwegian University of Science and Technology and 《Adressavisen》 (a local newspaper in Trondheim, Norway). After processing, the news text set A = {a1, a2,..., a |A|}, where |A| is the set length. Taking the news a i as an example, its related geographical location set is Use the dataset containing 1-week versions, and divide the news reading records into sessions at 30-minute intervals for news, and delete sessions with a length exceeding 20 and sessions with a length less than 3. After session segmentation, the session set S = {s1, s2,..., s |S|} and the session geographical location set where |S| is the set length. For any session is a set containing |s i | news, and the news is arranged in the reading order.

[0065] 2) Text feature extraction

[0066] To extract more accurate news text features, use a pre-trained Bert model to extract text features. First, the text composed of the title and content in the news a i is t i , and then, input it into the Bert model to obtain the text feature i of the news a Perform the above processing on the text features of all news to obtain the text feature set V = {v1, v2,..., v A} of all news.

[0067] 3) Geographical location feature extraction

[0068] First, let the geographical location set related to the news a i be where |L i| is the set length. The set of relationship weights between each geographical location and the article in the Adressa dataset is Constructed by entity connection. The geographical location knowledge graph containing all geographical location entities and their related entities is Then, use the knowledge graph embedding representation learning method RotatE based on relational reasoning to learn the embedding representations of entities in the knowledge graph, and finally obtain the feature set E of geographical locations KG .

[0069] For a news article, there may be multiple related geographical locations. To ensure the unity of news geographical location features, it is necessary to fuse the features of multiple geographical locations into a unified feature. With the help of the weights of geographical locations provided in the dataset, a weighted fusion strategy is designed. For news a i The geographical location feature of The calculation formula is as follows:

[0070]

[0071] Among them, is the feature of the j-th geographical location entity related to news a in the knowledge graph. For a user, the geographical location in a session does not change. Therefore, use its related geographical location feature i as the user geographical location feature p of the current session That's it. u

[0072] 4) Construction of preference extraction network based on GRU

[0073] GRU consists of many GRU computational units. Each computational unit is responsible for processing the input data at a certain moment, that is, the news text feature (news geographical location feature) at this moment in the session. The input of the GRU processing unit includes two parts, namely the hidden state at the previous moment and the data at this moment After that, the two are calculated through the following formula to obtain the alternative hidden state at time t

[0074]

[0075] Among them, r t is the reset gate, W r x , W r h , W x , W h are weight matrices, σ is the sigmoid activation function, tanh is the non-linear activation function, and ⊙ is the Hadamard product. After obtaining the alternative hidden state, through the update gate z​t The output hidden state h of this computing unit is obtained by the control t , and the calculation process is as follows:

[0076]

[0077] A sequence length of GRU units constitutes a user interest preference (geographical location preference) feature extraction network, and the hidden state output by the last computing unit is the user's interest preference feature (geographical location preference feature).

[0078] 5) User preference extraction

[0079] The text features and news geographical location feature sequences of the news read by the user in the session are respectively and To ensure that the same preference extraction network can be used for all sessions, zero vectors are used to supplement the above two feature sequences to the maximum session length N. Let the user interest preference extraction network be Interest_Net(·), and the geographical location preference extraction network be Location_Net(·). The following formula is used to calculate the user u's interest preference and geographical location preference

[0080]

[0081] where, [|s i |] is to take the hidden state output by the |s i |th GRU unit.

[0082] 6) Feature interaction unit based on multi-head attention mechanism

[0083] The candidate news is also called the news to be predicted. For a session, the candidate news refers to the other news except the news that appears in the set session read by the user. Its geographical location feature and text feature are included in the news geographical feature set and news text feature set.

[0084] In the session s i , the various features of user u include the geographical location feature interest preference and geographical location preference |s i | + 1 moment, the various features of the candidate news a include the text feature and geographical location feature Since the calculation of the multi-head attention mechanism requires that the various features have the same dimension, before performing feature interaction, it is necessary to map the various features to the same dimension, and the mapping formula is as follows:

[0085]

[0086] Among them and are weight matrices, and are biases, and relu is an activation function.

[0087] Combine the mapped features into a feature matrix Each row in the feature matrix represents a kind of feature. The multi-head attention mechanism is used for the interactive learning of the above features, which consists of multiple self-attention units, and each self-attention unit is responsible for the feature interaction in a feature space. The calculation process of using a single self-attention unit for feature interaction is as follows:

[0088]

[0089] Q ua = E ua W Q

[0090] K ua = E ua W K

[0091] U ua = E ua W U (2)

[0092] Among them, is the feature matrix after feature interaction. are the query, key, and value matrices respectively, is the weight matrix of the self-attention unit, d att = d map is the dimension of the mapped feature, is a regularization factor used to avoid gradient vanishing.

[0093] The representational ability of a single representational space is limited, so multiple self-attention units are selected to perform feature interaction in multiple feature spaces, that is, the multi-head attention mechanism.

[0094] The calculation process of the feature interaction unit based on the multi-head attention mechanism is as follows:

[0095]

[0096] Among them, is the weight matrix, and, in order to ensure that the computational amount of the multi-head attention is the same as that of the single-head attention, this method sets d att to d / n h . n his the number of attention heads. After completing the feature interaction, unfolds it by rows to obtain the feature and uses it for the prediction of the probability that the user clicks on the candidate news at time |s i | + 1.

[0097] 7) Click-through rate prediction

[0098] Since the multi-layer neural network is non-linear and has higher generalization ability, this method chooses to use it to construct a predictor, and the structure of this predictor is as Figure 2 shown. The calculation process of the predictor is as follows:

[0099] g (0) = f ua

[0100] ...

[0102]

[0103] where W p and b p are the weight matrix and bias, and relu, sigmoid are activation functions.

[0104] Embodiment 2

[0105] This embodiment discloses a location-aware conversational news recommendation system based on a knowledge graph;

[0106] As Figure 3 shown, the location-aware conversational news recommendation system based on a knowledge graph includes:

[0107] A location-aware conversational news recommendation system based on a knowledge graph, characterized in that it includes:

[0108] A data sorting module, configured to: process the news data set to obtain a news text set, a set of geographical locations related to each news, a session set, and a session geographical location set;

[0109] A text feature extraction module, configured to: extract features from each news text in the news text set respectively to obtain a news text feature set;

[0110] A geographical location feature extraction module, configured to: construct a geographical location knowledge graph, extract the features of each geographical location related to the news in the knowledge graph to obtain a news geographical location feature set, and based on the news geographical location feature set and the session geographical location set, obtain a session geographical location feature set;

[0111] The user preference extraction module is configured to: construct a GRU-based preference extraction network, and obtain the user's interest preferences and geographical location preferences in the set session based on the text features and news geographical location features of each news in the set session read by the user;

[0112] The interactive learning module is configured to: combine the text features and geographical location features of the candidate news, the user's interest preferences and geographical location preferences in the set session, and the user's geographical location features into a feature matrix, and use the multi-head attention mechanism to perform interactive learning on the feature matrix to obtain the feature matrix after feature interaction;

[0113] The click-through rate prediction module is configured to: based on the feature matrix after feature interaction, obtain the probability that the user clicks on the candidate news, and recommend news with a high click-through rate to the user.

[0114] Embodiment III

[0115] The purpose of this embodiment is to provide a computer-readable storage medium.

[0116] A computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, it implements the steps in the knowledge graph-based location-aware conversational news recommendation method described in Embodiment 1 of the present disclosure.

[0117] Embodiment IV

[0118] The purpose of this embodiment is to provide an electronic device.

[0119] An electronic device, including a memory, a processor, and a program stored on the memory and executable on the processor, and when the processor executes the program, it implements the steps in the knowledge graph-based location-aware conversational news recommendation method described in Embodiment 1 of the present disclosure.

[0120] The steps involved in the devices in the above Embodiments II, III, and IV correspond to those in Method Embodiment 1. For specific implementation manners, reference may be made to the relevant description part of Embodiment 1. The term "computer-readable storage medium" should be understood to include a single medium or multiple media including one or more instruction sets; it should also be understood to include any medium that can store, encode, or carry an instruction set for execution by a processor and enable the processor to execute any method in the present invention.

[0121] Those skilled in the art should understand that the various modules or steps of the present invention described above can be implemented by a general-purpose computer device. Optionally, they can be implemented by program codes executable by a computing device, so that they can be stored in a storage device and executed by the computing device, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. The present invention is not limited to any specific combination of hardware and software.

[0122] Although the specific implementation manners of the present invention have been described above in conjunction with the accompanying drawings, it is not a limitation on the protection scope of the present invention. Those skilled in the art should understand that, based on the technical solution of the present invention, various modifications or deformations that can be made without creative efforts by those skilled in the art are still within the protection scope of the present invention.

Claims

1. A location-aware conversational news recommendation method based on a knowledge graph, characterized in that, Including the following steps: Process the news dataset to obtain a news text set, a set of geographical locations related to each news, a session set, and a session geographical location set; Extract features from each news text in the news text set respectively to obtain a news text feature set; Construct a geographical location knowledge graph, extract the features of each geographical location related to the news in the knowledge graph to obtain a news geographical location feature set, and based on the news geographical location feature set and the session geographical location set, obtain a session geographical location feature set; Construct a preference extraction network based on GRU, and based on the text features and news geographical location features of each news in the set session read by the user, obtain the interest preference and geographical location preference of the user in the set session; Combine the text features and geographical location features of the candidate news, the interest preference and geographical location preference of the user in the set session, and the user geographical location features into a feature matrix, and use the multi-head attention mechanism to perform interactive learning on the feature matrix to obtain the feature matrix after feature interaction; Based on the feature matrix after feature interaction, obtain the probability that the user clicks on the candidate news, and recommend news with a high click-through rate to the user; Among them, obtain all geographical locations related to the news dataset, combine the existing knowledge graph, and construct a geographical location knowledge graph by means of entity connection. Use the knowledge graph representation learning method to learn the embedding representation of the location entity of each geographical location in the knowledge graph as the feature of each geographical location, and fuse the features of multiple geographical locations corresponding to each news in the dataset into the geographical location feature of this news. Aggregate the geographical location features of all news in the dataset to obtain a news geographical location feature set.

2. The method for location-aware conversational news recommendation based on a knowledge graph according to claim 1, wherein, Use the pre-trained Bert model to extract features from each news text in the news text set respectively, and aggregate the text features of all news to obtain a news text feature set.

3. The method for location-aware conversational news recommendation based on a knowledge graph according to claim 1, characterized in that The preference extraction network based on GRU consists of multiple GRU computing units. Input the text features and geographical location features of the set session read by the user into the interest preference extraction network and geographical location preference extraction network based on GRU respectively to obtain the interest preference and geographical location preference of the user in the set session.

4. The method for location-aware conversational news recommendation based on a knowledge graph according to claim 3, wherein The input of each GRU computing unit includes the hidden state of the previous moment and the text feature or geographical location feature of the news read at the current moment. The hidden state output by the last GRU computing unit is the interest preference or geographical location preference of the user in the set session.

5. The method for location-aware conversational news recommendation based on a knowledge graph according to claim 1, wherein Before using the multi-head attention mechanism to perform interactive learning on the feature matrix, map the candidate news text features and geographical location features, the interest preference and geographical location preference of the user in the set session, and the user geographical location features to the same dimension, and combine the mapped features into a feature matrix.

6. The method for location-aware conversational news recommendation based on a knowledge graph according to claim 1, wherein The multi-head attention mechanism consists of multiple self-attention units. Each self-attention unit is responsible for the feature interaction of a feature space. The process of using a single self-attention unit to perform feature interaction calculation is: Q ua = E ua W Q K ua = E ua W K U ua = E ua W U Among them, E ua is the feature matrix, is the feature matrix after feature interaction, are the query, key, and value matrices respectively, is the weight matrix of the self-attention unit, d att = d map is the feature dimension after mapping, Regularization factor.

7. A location-aware conversational news recommendation system based on a knowledge graph, characterized in that: Including: A data sorting module, configured to: process the news dataset to obtain a news text set, a set of geographical locations related to each news, a session set, and a session geographical location set; The text feature extraction module is configured to: perform feature extraction on each news text in the news text collection respectively to obtain a news text feature set; The geographical location feature extraction module is configured to: construct a geographical location knowledge graph, extract the features of each news-related geographical location in the knowledge graph to obtain a news geographical location feature set, and based on the news geographical location feature set and the session geographical location set, obtain a session geographical location feature set; The user preference extraction module is configured to: construct a preference extraction network based on GRU, and based on the text features and news geographical location features of each news in the set session read by the user, obtain the interest preference and geographical location preference of the user in the set session; The interaction learning module is configured to: combine the text features and geographical location features of the candidate news, the interest preference and geographical location preference of the user in the set session, and the user geographical location features into a feature matrix, and use the multi-head attention mechanism to perform interaction learning on the feature matrix to obtain a feature matrix after feature interaction; The click-through rate prediction module is configured to: based on the feature matrix after feature interaction, obtain the probability that the user clicks on the candidate news, and recommend news with a high click-through rate to the user; Among them, all geographical locations related to the news data set are obtained, combined with the existing knowledge graph, and a geographical location knowledge graph is constructed by means of entity connection. The knowledge graph representation learning method is used to learn the embedding representation of the location entity of each geographical location in the knowledge graph as the feature of each geographical location. The features of multiple geographical locations corresponding to each news in the data set are fused into the geographical location feature of this news, and the geographical location features of all news in the data set are aggregated to obtain a news geographical location feature set.

8. A computer-readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the knowledge graph-based location-aware conversational news recommendation method according to any one of claims 1-6.

9. An electronic device, comprising a memory, a processor, and a program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the knowledge graph-based location-aware conversational news recommendation method according to any one of claims 1-6.

Images