News recommendation method fusing spatio-temporal features

Abstract

The application provides a news recommendation method fusing space-time features, comprising the following steps: S1, constructing a news knowledge graph; S2, constructing a news entity model; and S3, constructing a recommendation model; on the basis of the constructed news knowledge graph, a news entity model is established through a TransE model, a news click probability of the news entity is calculated by using an attention mechanism to combine the space-time features of the historical time of news release and the news occurrence location based on the historical clicked news of the user, the news associated with the news entity v is sorted from high to low according to the news click probability of the news entity v, a news recommendation list is generated, and the personalized recommendation of the news with the strengthened space-time features is realized, and the news interested by the user is timely pushed.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, in particular to a news recommendation method fusing space-time features. BACKGROUND

[0002] With the development of network technology and news media, various news information media including news portals and news APPs have a large daily circulation and various types. The basic characteristics of news, such as timeliness, universality, openness, variability and pertinence, and the multi-faceted, multi-channel and rapidity of network news, make users receive a large amount of network news every day. On the one hand, a large amount of news that users are not interested in or not interested in causes the user to have news "information overload", on the other hand, the news that the user is interested in or continuously pays attention to cannot be timely pushed to the user. Therefore, how to mine the personalized preferences of users and recommend the news that users are interested in according to the news features is one of the key problems that need to be solved in the current "intelligent media" era.

[0003] Most of the traditional news recommendation methods mine user preferences by analyzing the news type or news theme of the user's historical click news, and some research further analyzes the user's reading time of news, but in fact, users will also pay attention to the news related to the region when reading the news, such as real-time news of the place where they live or news of a specific area of interest; the user's continuous attention to the tracking reports of some current news or the development of the time of major events, such as reading the related news of the project or favorite athlete during the XX Games for several consecutive days. The above phenomena all reflect the user's "space-time" preference for news time and the place where the news event occurs, but the existing news recommendation methods do not consider the "space-time" feature for personalized news recommendation, and cannot timely push the news that the user is interested in.

[0004] In summary, there is an urgent need for a news recommendation method that can timely recommend news that users are interested in to solve the problems in the prior art. SUMMARY

[0005] The present application aims to provide a news recommendation method fusing space-time features, and the specific technical solutions are as follows:

[0006] The news recommendation method fusing space-time features comprises:

[0007] S1: news knowledge graph construction, specifically comprising:

[0008] S1-1: obtaining entities in news, constructing a data model describing news entities, and obtaining a corresponding relationship set according to the relationship between news entities;

[0009] S1-2: Extract news entity v from the unstructured data in the data model of step S1-1 to obtain the corresponding entity set;

[0010] S1-3: Establish a set of triples based on entity sets and relation sets, and construct a news knowledge graph by combining entity sets, relation sets, and triple sets;

[0011] S2: News entity model construction, including: constructing the TransE model based on the news knowledge graph and attention mechanism in step S1, and obtaining a news entity model including news entity v through the TransE model;

[0012] S3: Recommendation model construction, specifically including:

[0013] S3-1: Employ an attention mechanism to calculate the relationship between a given news entity v and entities g from the user's historical clicks on news articles. i The degree of correlation P at historical time T T ;

[0014] S3-2: Calculate the correlation level P using a weighted summation method. T User preferences for a particular news entity

[0015] S3-3: By introducing a time decay term at historical time T and user preferences for a particular news entity. To obtain the degree of user preference for a particular news entity.

[0016] S3-4: Introduce a power-law decay term for the distance between the news location and the user to obtain the probability of a user clicking on a news entity v. Then, sort the news associated with news entity v from high to low according to the probability of clicking on news entity v to generate a news recommendation list.

[0017] Preferably, in step S3-1, the relevance level P T The expression is as follows:

[0018]

[0019] Among them, P T This represents a news entity v and an entity g representing the news items clicked by the user in the past. i The degree of correlation at historical time T, g i This refers to the entities that represent the user's historical clicks on news articles. T represents the historical time, and v represents the news entity, which includes the header entity h and the footer entity t. T G represents a news entity v at a historical time T. i This represents the set of user interactions at historical time T.

[0020] Preferably, the expression of the user's preference for a certain news entity in step S3-2 is as follows:

[0021]

[0022] wherein (g i )∈G i represents the embedding representation of the entity g i clicked by the history, represents the user's preference for a certain news entity, and u represents a certain user.

[0023] Preferably, the expression of the preference degree in step S3-3 is as follows:

[0024]

[0025] wherein, represents the time decay term of the history time T, e represents the natural constant, a2 is the constant parameter of the time decay term, and τ represents the current time, represents the user's preference degree for a certain news entity.

[0026] Preferably, the expression of the news click probability in step S3-4 is as follows:

[0027]

[0028] wherein, represents the news click probability, is the power-law decay term of the distance between the news occurrence location and the user, a1, b1 and c1 are constant parameters of the power-law decay term, and dist represents the user distance.

[0029] Preferably, the news entity model in step S2 includes a correlation relationship tree model of different news events and a space-time evolution tree model of the same news event.

[0030] Preferably, the step S2 specifically includes:

[0031] S2-1: Set different weight values w for different relationships in the news knowledge graph;

[0032] S2-2: Construct a TransE model based on an attention mechanism, and the expression of the TransE model is as follows:

[0033] E(h,r,t)=w·||h+r-t||;

[0034] wherein E(h,r,t) represents the semantic correlation degree between the head entity h and the tail entity t, h represents the head entity, r represents the relationship, t represents the tail entity, and w represents the weight value of the relationship;

[0035] S2-3: The semantic correlation degree between news entities v in different news events is calculated through the TransE model, a threshold value δ1 is set, news entities exceeding the threshold value δ1 are divided into the same type, and an associated relationship tree model of different news events is constructed through the associated relationship r of each entity;

[0036] S2-4: The semantic correlation degree between news entities v in the same news event is calculated through the TransE model, a threshold value δ2 is set, news entities exceeding the threshold value δ2 are divided into the same type, and a space-time evolution tree model of the same news event is constructed through the associated relationship r of each entity.

[0037] The technical scheme of the present application has the following beneficial effects:

[0038] Based on the constructed news knowledge graph, the TransE model is constructed to establish a news entity model, the news click probability of the news entity is calculated by using the attention mechanism to combine the historical time of news release and the space-time features of the news occurrence location based on the historical click news of the user, the news associated with the news entity v is sorted from high to low according to the news click probability of the news entity v, and a news recommendation list is generated, so that the space-time feature enhanced news personalized recommendation is realized, and the news interested by the user is timely pushed.

[0039] In addition to the purposes, features and advantages described above, the present application has other purposes, features and advantages. The present application will be further described below with reference to the drawings. BRIEF DESCRIPTION OF DRAWINGS

[0040] The accompanying drawings, which form a part of this application, are included to provide a further understanding of the application, and are incorporated in and constitute a part of this application. The embodiments of the present application illustrated in the drawings and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application. In the drawings:

[0041] Figure 1 is a structure schematic diagram of the associated relationship tree model obtained by step S2-3 of the preferred embodiment 1 of the present application;

[0042] Figure 2 is a structure schematic diagram of the space-time evolution tree model obtained by step S2-4 of the preferred embodiment 1 of the present application;

[0043] Figure 3 is a recommendation effect comparison diagram based on click rate estimation of the preferred embodiment 1 of the present application;

[0044] Figure 4 is a Precision@K performance index comparison diagram of the preferred embodiment 1 of the present application;

[0045] Figure 5is a Recall@K performance index comparison chart of preferred embodiment 1 of the present application;

[0046] Figure 6 is a F1@K performance index comparison chart of preferred embodiment 1 of the present application. DETAILED DESCRIPTION

[0047] The embodiments of the present application are described in detail below with reference to the accompanying drawings, but the present application can be implemented in various different ways as limited and covered by the claims.

[0048] Embodiment 1:

[0049] Reference Figure 1 , a news recommendation method fusing spatio-temporal features, comprising:

[0050] S1: news knowledge graph construction, specifically comprising:

[0051] S1-1: obtaining entities in news according to five elements of news (including news keywords, news publishing time, news occurrence address, news type and person), constructing a data model describing news entities, and obtaining a corresponding relationship set = ["_contain", "_time", "_address", "_type", "_people"] according to the relationship between news entities;

[0052] S1-2: extracting news entities in unstructured data in the data model of step S1-1, obtaining a corresponding entity set; for example, for each news, an entity set = ["news ID", "keyword", "time", "place", "type", "person"] can be obtained;

[0053] Among them, news ID: in the process of collecting data, a corresponding news ID is set for each news to uniquely identify the news entity;

[0054] Publishing time: the accurate time of publishing each news;

[0055] Occurrence address: the detailed address of each news occurrence, combined with the place contained in the open geographic knowledge graph GeoNames to fuse the geographic information knowledge, and the address area is hierarchically divided;

[0056] Keywords: 2-10 keywords are extracted from the detailed content of news combined with the news title to describe the news event;

[0057] Type: defining news as political news or entertainment news or other types according to news content;

[0058] Person: extracting the related person of each news, which can be null;

[0059] For example, for the news "On XX, XX, XX, at XX o'clock and XX minutes, XX accident occurred in XX district of XX city of XX province. After the accident, a rescue was quickly launched. On XX, XX, at dawn, the reporter followed XX fire rescue team into the accident site and witnessed the on-site rescue.", the corresponding entity set will be extracted = ["D101", "XX accident", "fire rescue team rescue", "XXXX year XX month XX day", "XX province", "XX city", "news of the time and politics"];

[0060] S1-3: Establish a set of triples based on the entity set and the relationship set, and construct a news knowledge graph combining the entity set, the relationship set, and the set of triples;

[0061] For example, for the above news, some triples contained in the news knowledge graph can be obtained = [(D101, _time, XXXX year XX month XX day), (D101, _address, XX province), (D101, _address, XX city)]

[0062] S2: News entity model construction, including: constructing a TransE model based on a news knowledge graph and an attention mechanism, and constructing a news entity model (i.e., a news knowledge graph with spatiotemporal vector features) including a news entity v through the TransE model; the construction process includes:

[0063] S2-1: Set different weight values w for different relationships in the news knowledge graph; for example, the weight value of "_contain" in the relationship set = ["_contain", "_time", "_address", "_type", "_people"] is w1, the weight value of "_time" is w2, the weight value of "_address" is w3, the weight value of "_type" is w4, and the weight value of "_people" is w5;

[0064] S2-2: Construct a TransE model based on an attention mechanism, and the expression of the TransE model is as follows:

[0065] E(h, r, t) = w·||h+r-t||;

[0066] Wherein, E(h, r, t) represents the semantic association degree between the head entity h and the tail entity t, h represents the head entity, r represents the relationship, t represents the tail entity, and w represents the weight value of the relationship;

[0067] S2-3: Calculate the semantic correlation degree between news entities v in different news events through the TransE model, set threshold value δ1, the value of δ1 is greater than or equal to 0.6 and less than or equal to 0.8, divide the news entities with E(h, r, t) value exceeding the threshold value δ1 into the same type, and construct the correlation relationship tree model of different news events through the correlation relationship r of each entity;

[0068] For example, model the news entities related to "sports meeting" in the above manner, see Figure 1 Divide the news entities of the sports meeting in previous years into the same type of event, and describe the correlation relationship between related events according to the news release time.

[0069] S2-4: Calculate the semantic correlation degree between news entities v in the same news event through the TransE model, set threshold value δ2, the value of δ2 is greater than or equal to 0.6 and less than or equal to 0.8, divide the news entities with E(h, r, t) value exceeding the threshold value δ2 into the same type, and construct the spatio-temporal evolution tree model of the same news event through the correlation relationship r of each entity.

[0070] For example, model the related news entities reported for consecutive days of "XX sports meeting" in the above manner, see Figure 2 Take "XX sports meeting" as the root node, calculate other news entities of "XX sports meeting" through the TransE model, and describe the correlation relationship between related news entities according to the news release time;

[0071] S3: Recommendation model construction, specifically including:

[0072] S3-1: Use attention mechanism to calculate the correlation degree P i of a news entity v and the entities g T of the user's historical clicked news at historical time T. T The expression of the correlation degree P

[0073]

[0074] Wherein, P T represents the correlation degree of a news entity v and the entities g i of the user's historical clicked news at historical time T, g i is the entity of the user's historical clicked news, T represents the historical time, v represents the news entity and includes the head entity h and the tail entity t, v T represents the news entity v at historical time T, G i represents the interaction set of the user at historical time T, and the historical clicked interaction set is G = {G1, G2,..., G i}, the user's interaction set at historical time T is G i = {g1, g2,..., g i} ;

[0075] S3-2: The degree of relevance P T of a user to a news entity is calculated by weighted summation The expression of the degree of relevance P of a user to a news entity is as follows:

[0076]

[0077] Wherein, (g i )∈G i represents the embedding representation of the historical clicked news entity g i , represents the degree of preference of a user to a news entity, and u represents a user;

[0078] S3-3: By introducing the time decay term at historical time T and the degree of preference of a user to a news entity The degree of preference of a user to a news entity is obtained The expression of the degree of preference of a user to a news entity is as follows:

[0079]

[0080] Wherein, represents the time decay term at historical time T, e represents a natural constant, a2 is a constant parameter of the time decay term, which can be set according to actual requirements, and τ represents the current time, represents the degree of preference of a user to a news entity;

[0081] S3-4: A power-law decay term of news occurrence location and user distance is introduced, the news click probability of a user to a news entity v is obtained, the news associated with the news entity v is sorted according to the news click probability from high to low, a news recommendation list is generated, and the expression of the news click probability is as follows:

[0082]

[0083] Wherein, represents the news click probability, is a power-law decay term of news occurrence location and user distance, a1, b1 and c1 are constant parameters of the power-law decay term, which can be set according to actual requirements, and dist represents the user distance.

[0084] Experimental results

[0085] Data set fusing spatio-temporal characteristics

[0086] In the experiment, the entity in the Yelp2018 dataset is taken as a seed node, the geographic position longitude and latitude information of the entity matched in the Yelp:business dataset are extracted, and the time information of the entity is combined to establish a knowledge graph with spatio-temporal characteristics, namely YelpST, as shown in Table 1.

[0087] Table 1: Domain knowledge graph with spatio-temporal characteristics

[0088] Dataset #users #projects #interactions #relationships #triples Size (MB) Yelp:business 1637138 192609 6685900 ─ ─ 8845.45 Yelp2018 45919 192609 1177186 42 1853704 143.55 YelpST 45919 192609 1177186 42 1853704 37.38

[0089] In the experiment, 80% of the user interaction data in the constructed YelpST dataset is taken as a training set, 10% is taken as a validation set for training hyperparameters, and the remaining 10% is taken as a test set.

[0090] Experimental environment and evaluation index

[0091] The operating system of the experimental environment is Ubuntu 20.04, the GPU is GeForce RTX3080, and the RAM is 32GB.

[0092] In the training process, the setting of the hyperparameters is shown in Table 2. Among them, the vector dimension of the relationship is d=8, the sliding data window is H=3, the neighbor sampling number is K=4, the L2 regularization loss is λ=10-7, and the learning rate is η=2×10-2.

[0093] Table 2: Hyperparameter setting

[0094] Parameter Description Setting d Vector dimension of relationships 8 H Sliding data window 3 K Number of neighbors to sample 4 L2 regularization loss Learning rate 10-7 Figure 3 Figure 3 2×10-2

[0095] In order to verify the recommendation effect of the news recommendation method with fused spatio-temporal characteristics, AUC (Area Under Curve) and Accuracy (accuracy) are used as the recommendation performance evaluation indexes based on click rate estimation; Precision@K, Recall@K and F1@K three performance indexes are used as the Top-K recommendation performance evaluation indexes.

[0096] Experimental result analysis

[0097] The news recommendation method with fused spatio-temporal characteristics disclosed in the present application can effectively improve the recommendation effect when embedding the knowledge graph by fusing the time feature and the space feature of the event. Figure 4It can be seen from the experimental results that the news recommendation method fusing spatio-temporal features (STAKGE in the figure) has the best recommendation effect, with an AUC of 93.5% and an accuracy of 87.3%; the recommendation effect of the method SAKGE fusing only spatial features is second; and the recommendation effect of the method TAKGE fusing only temporal features is worse than the above two methods.

[0098] In order to verify the effectiveness of the news recommendation method fusing spatio-temporal features, the recommendation effects of the method are compared with those of two types of benchmark recommendation methods. Among them, the non-knowledge graph recommendation methods include SVD, LibFM, and the knowledge graph-based recommendation methods include LibFM+TransE, PER, CKE and RippleNet.

[0099] (1) SVD is a collaborative filtering recommendation model using matrix decomposition, which models the potential features of user-item interaction data.

[0100] (2) LibFM is a feature-based matrix decomposition model, which takes user ID and item ID as input to the model for recommendation.

[0101] (3) The LibFM+TransE recommendation model is based on the LibFM model, which uses the TansE model to perform entity embedding on user-item entity pairs.

[0102] (4) The PER recommendation model uses the knowledge graph as a heterogeneous information network, and uses the extracted meta-path features to represent the connectivity between users and items.

[0103] (5) CKE is a recommendation method based on collaborative filtering and knowledge graph feature learning, which combines structured knowledge, image knowledge and text knowledge for learning and recommendation.

[0104] (6) The RippleNet model uses the knowledge graph to realize an end-to-end framework of the recommendation system, which continuously discovers the user's potential hierarchical interests by propagating user preferences in the knowledge graph for recommendation.

[0105] From Figure 5 It can be seen that in the recommendation performance comparison based on click rate evaluation, the AUC of the RippleNet model is 89.9% and the accuracy is 87.3% among the six methods other than STAKGE, SAKGE and TAKGE, which is better than the other five methods. However, the news recommendation method fusing spatio-temporal features STAKGE improves by 3.6% and 2.4% respectively in the two indicators compared with the RippleNet model, and its performance is better than the above six methods.

[0106] Referring to Figure 6 ,​ and ​ In the Top-K recommendation performance comparison, the Precision@K, Recall@K and F1@K three performance indicators are used to compare STAKGE and RippleNet, and it can be seen from the experimental results that the performance of STAKGE is better than RippleNet in the three indicators. At the same time, through experimental analysis, it is found that when K=4, the recommendation effect of STAKGE is best, which shows that when K=4, STAKGE can better integrate the neighbor information of the nodes in the knowledge graph, and too large K value will introduce more noise and affect the recommendation effect.

[0107] The above merely describes preferred embodiments of the present application but is not intended to limit the present application. Various modifications and changes can be made by those skilled in the art to the present application, and any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A news recommendation method fusing spatio-temporal features, characterized in that, Comprise: S1: news knowledge graph construction, specifically comprising: S1-1: obtaining entities in news, constructing a data model describing news entities, and obtaining a corresponding relationship set according to the relationship between news entities; S1-2: extract news entities in the unstructured data in the data model of step S1-1 , to obtain a corresponding entity set; S1-3: establishing a triple set based on the entity set and the relationship set, and constructing a news knowledge graph combining the entity set, the relationship set and the triple set; S2: news entity model construction, comprising: constructing a TransE model based on the news knowledge graph of step S1 and an attention mechanism, and constructing a news entity model comprising news entities ​ S3: recommendation model construction, specifically comprising: S3-1: Employ an attention mechanism to calculate a specific news entity. Entities related to users' historical clicks on news In historical time The degree of relevance ; S3-2: Calculate the relevance degree of the news entity to the user by weighted summation The user's preference for a news entity The expression of the user's preference for a news entity is as follows: ； wherein, an entity representing a historically clicked news an embedding representation of the entity representing a historically clicked news, representing a certain user; S3-3: By introducing a time decay term at history time and the user's preference for a news entity , the user's preference degree for a news entity is obtained ; the expression of the preference degree is as follows: ； wherein, denotes a history time a time decay term, denotes a natural constant, is a constant parameter of the time decay term, denotes a current time, denotes a user's preference degree for a certain news entity; S3-4: Introducing a power-law decay term of the distance between the news occurrence location and the user, obtaining the news click probability of a user on a certain news entity, and sorting the news associated with the news entity according to the news click probability from high to low to generate a news recommendation list; ;​​ The expression of news click probability is as follows: ； wherein, represents the news click probability, is a power-law decay term of the distance between the news occurrence location and the user, , and are constant parameters of the power-law decay term, represents the user distance.

2. The news recommendation method of claim 1, wherein, The degree of relevance in the step S3-1 The expression is as follows: ； in, Indicates a news entity Entities related to users' historical clicks on news In historical time The degree of relevance, It is the entity of the user's historical clicks on news. Indicates historical time. This refers to a news entity, which includes a head entity. Tail-end entity , Indicates historical time News entities , Indicates the user's historical time The following is an interaction set.

3. The news recommendation method of claim 1, wherein, The news entity model in step S2 comprises a correlation relationship tree model of different news events and a space-time evolution tree model of the same news event.

4. The news recommendation method according to claim 3, characterized in that, The step S2 specifically comprises: S2-1: Set different weight values for different relations in the news knowledge graph ; S2-2: constructing a TransE model based on an attention mechanism, and the expression of the TransE model is as follows: ； wherein, represents the degree of semantic association between a head entity and a tail entity, represents the head entity, represents the relationship, represents the tail entity, represents the weight value of the relationship;​ S2-3: Calculating news entities in different news events using the TransE model. The degree of semantic association between them, setting a threshold It will exceed the threshold. News entities are categorized into the same type, and their relationships are determined by the relationships between these entities. Construct a tree model of the relationships between different news events; S2-4: Calculate the semantic correlation degree between news entities in the same news event through the TransE model, set a threshold , and divide news entities exceeding the threshold into the same type and construct a spatio-temporal evolution tree model of the same news event through the associated relationship of each entity . ​

Images