CN115618108B - Brand recommendation method based on knowledge graph in new retail model - Google Patents
Brand recommendation method based on knowledge graph in new retail model Download PDFInfo
- Publication number
- CN115618108B CN115618108B CN202211305797.7A CN202211305797A CN115618108B CN 115618108 B CN115618108 B CN 115618108B CN 202211305797 A CN202211305797 A CN 202211305797A CN 115618108 B CN115618108 B CN 115618108B
- Authority
- CN
- China
- Prior art keywords
- user
- cigarette
- cigarettes
- knowledge graph
- vector
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 238000000034 method Methods 0.000 title claims abstract description 79
- 235000019504 cigarettes Nutrition 0.000 claims abstract description 258
- 239000011159 matrix material Substances 0.000 claims abstract description 65
- 230000003993 interaction Effects 0.000 claims abstract description 32
- 238000013519 translation Methods 0.000 claims abstract description 24
- 238000005065 mining Methods 0.000 claims abstract description 12
- 239000013598 vector Substances 0.000 claims description 98
- 230000006870 function Effects 0.000 claims description 28
- 230000014509 gene expression Effects 0.000 claims description 27
- 230000007246 mechanism Effects 0.000 claims description 23
- 238000013507 mapping Methods 0.000 claims description 18
- 230000002452 interceptive effect Effects 0.000 claims description 8
- 230000008569 process Effects 0.000 claims description 7
- 230000004931 aggregating effect Effects 0.000 claims description 6
- 238000007781 pre-processing Methods 0.000 claims description 5
- 230000009466 transformation Effects 0.000 claims description 5
- 230000002776 aggregation Effects 0.000 claims description 4
- 238000004220 aggregation Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000004364 calculation method Methods 0.000 claims description 3
- 238000012216 screening Methods 0.000 claims description 3
- 238000001228 spectrum Methods 0.000 abstract 4
- 238000010586 diagram Methods 0.000 description 4
- 235000019640 taste Nutrition 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 238000013459 approach Methods 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000001914 filtration Methods 0.000 description 2
- 241000208125 Nicotiana Species 0.000 description 1
- 235000002637 Nicotiana tabacum Nutrition 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 230000000694 effects Effects 0.000 description 1
- 239000013604 expression vector Substances 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000004927 fusion Effects 0.000 description 1
- 239000000463 material Substances 0.000 description 1
- 239000000203 mixture Substances 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000005096 rolling process Methods 0.000 description 1
- 238000011524 similarity measure Methods 0.000 description 1
- 238000007619 statistical method Methods 0.000 description 1
- 235000019505 tobacco product Nutrition 0.000 description 1
- 230000007704 transition Effects 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/90—Details of database functions independent of the retrieved data types
- G06F16/95—Retrieval from the web
- G06F16/953—Querying, e.g. by the use of web search engines
- G06F16/9535—Search customisation based on user profiles and personalisation
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
- G06F16/30—Information retrieval; Database structures therefor; File system structures therefor of unstructured textual data
- G06F16/36—Creation of semantic tools, e.g. ontology or thesauri
- G06F16/367—Ontology
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02P—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN THE PRODUCTION OR PROCESSING OF GOODS
- Y02P90/00—Enabling technologies with a potential contribution to greenhouse gas [GHG] emissions mitigation
- Y02P90/30—Computing systems specially adapted for manufacturing
Landscapes
- Engineering & Computer Science (AREA)
- Databases & Information Systems (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Animal Behavior & Ethology (AREA)
- Computational Linguistics (AREA)
- Information Retrieval, Db Structures And Fs Structures Therefor (AREA)
Abstract
The invention discloses a cigarette brand recommendation method based on a knowledge graph in a new retail model, which comprises the following steps: constructing a user-project knowledge graph by taking the user, the cigarette and the related attributes among the user and the cigarette as nodes; predicting missing entities in the knowledge graph after proposal set enhancement pretreatment based on a TransR translation model; obtaining the preference information of the current user on cigarettes according to the user-cigarette interaction information and the preference matrix; and generating a cigarette recommendation list according to the implicit preference information of the current user, the inter-node relation obtained by mining based on the user-project graph knowledge graph and the missing entity predicted based on the user-project knowledge graph. According to the cigarette brand recommendation method based on the knowledge spectrum in the new retail mode, the proposal set is used for complementing the knowledge spectrum, predicting tail entities, mining implicit preference, and finally carrying out joint learning with the knowledge spectrum, so that the integrity of the knowledge spectrum and the accuracy of preference between a user and cigarettes are ensured.
Description
Technical Field
The invention relates to the technical field of cigarette recommendation, in particular to a method for recommending cigarette brands based on a knowledge graph in a new retail model.
Background
Because of the particularities of tobacco products, traditional retail brand recommendation channels are relatively limited, and methods of on-site recommendation by brand managers via retailer storefronts have not been satisfactory. The demands of customers in stores are different, so that the customers are required to learn to look like a look, enthusiasm is actively met, the intention of purchasing cigarette brands is judged through the eyes and behaviors of the customers, the actual consumption demands are known, and the customers become forever return customers of the stores, so that the method is not easy to achieve.
Under a new retail mode, a tobacco in-line retail terminal store can intelligently recommend personalized cigarette brands with similar cigarettes and styles and tastes to users by utilizing big data analysis and recommendation technology according to the characteristics of new retail, so that the client needs are guided, the user experience is improved, and therefore targeted marketing is achieved, marketing cost is reduced, and income is improved. The existing cigarette brand recommending party introduces a knowledge graph to improve the interpretability, but does not consider the integrity of the graph and the mining of preference, so that personalized recommendation is difficult to achieve.
Therefore, a method for recommending cigarette brands based on a knowledge graph in a new retail model is needed.
Disclosure of Invention
The invention aims to provide a cigarette brand recommendation method based on a knowledge graph in a new retail model, so as to solve the problems in the prior art, the knowledge graph is complemented and the tail entity is predicted by using a proposal set, the knowledge graph can be complemented, the vector representation of a user and a cigarette is obtained through the scoring matrix learning of the user-cigarette interaction, more accurate implicit preference is mined, and finally the combination learning is carried out with the knowledge graph, so that the completeness of the knowledge graph constructed by the user and the cigarette and the accuracy of the preference between the user and the cigarette are ensured.
The invention provides a cigarette brand recommendation method based on a knowledge graph in a new retail model, which comprises the following steps:
constructing a user-project knowledge graph by taking a user, a cigarette and related attributes between the user and the cigarette as nodes, and carrying out enhanced pretreatment on the user-project knowledge graph through a proposal set;
predicting the missing entity in the user-project knowledge graph after proposal set enhancement pretreatment based on a TransR translation model;
obtaining the preference information of the current user on cigarettes according to the user-cigarette interaction information and a preference matrix, wherein the user-cigarette interaction information represents the cigarette information with the interaction history of the user, and the preference matrix represents the preference information of the user on cigarettes;
And generating a cigarette recommendation list according to the implicit preference information of the current user, the relation among nodes obtained by mining the user-project graph knowledge graph and the missing entity predicted by the user-project knowledge graph.
In the cigarette brand recommendation method based on the knowledge graph in the new retail model, preferably, in the user-project knowledge graph, the relationship between each node comprises an attribute level relationship and a user related relationship, and is used as an edge of the knowledge graph, and the edges of all the knowledge graphs form the proposal set, wherein the attribute level relationship comprises a brand or a manufacturer, and the user related relationship comprises co-purchase or co-view.
The method for recommending cigarette brands based on the knowledge graph in the new retail model, as described above, preferably, the method for recommending cigarette brands based on the knowledge graph in the new retail model predicts missing entities in the user-project knowledge graph after the proposal set enhancement pretreatment, and specifically includes:
learning the expression of the entity and the relation in the user-project knowledge graph after proposal set enhancement pretreatment based on a TransR translation model;
and predicting the missing entity in the user-project knowledge graph after the proposal set is subjected to the enhancement pretreatment according to the learning result of the entity and the relation.
The method for recommending cigarette brands based on the knowledge graph in the new retail model, as described above, preferably, the method for recommending cigarette brands based on the knowledge graph in the new retail model learns the expression of entities and relations in the user-project knowledge graph after the enhancement pretreatment of the proposal set, and specifically includes:
mapping the transformation matrix of the head entity and the tail entity according to the relation in the user-project knowledge graph after the enhancement pretreatment of the proposal set into the same vector space through a TransR translation model so as to set each relation as an independent space, and mapping the head entity and the tail entity into the vector space of the corresponding relation r through the following formula,
e hr ＝e h M r (1)
e tr ＝e t M r (2)
wherein M is r Matrix representing spatial projection, which is a representation of the relationship with the embedding of r itself, e hr Represented by projection matrix M r The original head entity vector e h Mapping toIn the relation space of the corresponding relation r, e tr Representing the tail entity e t Mapping into a relationship space of the corresponding relationship r.
The method for recommending cigarette brands based on the knowledge graph in the new retail model, as described above, preferably predicts the missing entity in the user-project knowledge graph after the enhancement pretreatment of the proposal set according to the learning result of the entity and the relationship, and specifically includes:
The scores for a given entity, each predicted entity and corresponding relationship triplet are calculated by the following formula,
and taking the predicted entity corresponding to the minimum score as a tail entity in the user-project knowledge graph after the enhancement pretreatment of the proposal set.
The method for recommending cigarette brands based on the knowledge graph in the new retail model, as described above, preferably, obtains the preference information of the current user for cigarettes according to the user-cigarette interaction information and the preference matrix, and specifically includes:
aggregating the user and the cigarette information with the interaction history with the user through an attention mechanism to obtain a new user representation;
determining a preference matrix according to the reason that the user has interaction with the cigarettes;
and determining the maximum preference information of the current user according to the similarity between the sum vector of the new user representation and the cigarette vector representation and the preference vector in the preference matrix.
The method for recommending cigarette brands based on the knowledge graph in the new retail mode preferably aggregates the user and the cigarette information with the interaction history with the user through the attention mechanism to obtain a new user representation, and specifically comprises the following steps:
aggregating user expression information and project expression information obtained based on a scoring matrix obtained by an interactive table of users and projects through an attention mechanism to obtain new user expression, wherein the interactive table is used for representing record information of cigarettes with interactive histories of the users, calculating a user cigarette aggregation vector as the new user expression through the following formula,
W ij ＝Softmax(LeakyReLU(W att )(u i ||v ij )) (4)
Wherein W is ij Representing a weight matrix calculated by an attention mechanism for representing the preference degree of the user for different cigarettes, u i Vector representation representing user, v ij Vector representation, W, of cigarettes representing history of interactions with a user att A weight matrix representing the attention mechanism, ||represents a splice vector of the user and the cigarette, leakyReLU represents a nonlinear activation function, the weight of attention is calculated by a Softmax function, u r The user cigarette aggregate vector is represented as a new user vector.
The method for recommending cigarette brands based on the knowledge graph in the new retail mode, as described above, preferably, determines implicit preference information of a current user according to similarity between a summation vector represented by a new user representation and a cigarette vector and a preference vector in the preference matrix, and specifically includes:
according to the similarity of different cigarettes, the scoring score of the current user on each cigarette is calculated through the following formula,
wherein Score ui Representing the score of user u for cigarette i, i and j representing cigarettes, n representing the total number of cigarettes, sim (i, j) representing the similarity of cigarettes i and j calculated by cosine similarity, R uj Representing the score of user u for cigarette j;
The implicit preference information p of the current user is determined through the following formula calculation,
S p∈P ＝(u+v)⊙p (7)
p＝argmax i (log(S i )+b) (8)
where u represents a new user representation of the current user, v represents a cigarette vector representation having an interaction history with the current user, P in equation (7) represents a vector corresponding to the preference matrix P, and the number of preferences is determined from the dataset, and b represents the bias.
The cigarette brand recommendation method based on the knowledge graph in the new retail model, as described above, preferably, generates a cigarette recommendation list according to the implicit preference information of the current user, the relationship between nodes obtained based on the mining of the user-project graph knowledge graph and the missing entity predicted based on the user-project knowledge graph, and specifically includes:
obtaining final preference expression of the current user according to the implicit preference information of the current user and the relation learned based on the TransR translation model through the following formula
And the final preference matrix of the current user->
Wherein r represents the relation in the knowledge graph, and w p Representing an overall weight matrix combining various preferences p, M r A matrix representing a projection of a space is presented,
predicted from the expression i of cigarettes by the scoring matrix and based on the user-project knowledge-graph after the enhanced pretreatment by the proposed set The entity vector e of (2) to obtain the final feature vector of the cigarette
Mapping the feature vectors of the user and the feature vectors of the cigarettes to a vector space under corresponding preferences by the following formula
In (I)>
Wherein u is r ' represents the user cigarette aggregate vector calculated by the attention mechanism, u r Representation mapping to vector space
User representation of->
Representation mapping to vector space +.>
The cigarette of (3) is represented by a cigarette,
for a pair of
A triplet, scored by a scoring function in the following formula,
and selecting partial cold cigarettes and hot cigarettes for scoring, and obtaining a final cigarette recommendation list by using a TopN recommendation method according to the scoring result.
The method for recommending cigarette brands based on the knowledge graph in the new retail mode, as described above, preferably, the selecting part of the cold cigarettes and the hot cigarettes for scoring, and obtaining a final cigarette recommendation list by using a TopN recommendation method according to the scoring result, specifically includes:
for hot cigarettes, a scoring score is obtained by equation (14), all hot cigarettes enter the cigarette recommendation list,
for a cold-door cigarette, the similarity between each cigarette is determined according to the embedded representation of the user and the cigarette by the following formula,
Wherein S (x, y) represents reconstructing a similarity function from the generated embedded representations of the user and the cigarette to obtain a similarity value;
screening out the characteristic representation of the cigarettes with small difference from the vector inner product value through the similarity function;
obtaining the grading score of the user on the cold cigarettes screened out by the similarity function according to the formula (6);
comparing the score of the cold cigarette with the score of each hot cigarette, and adding the cold cigarette into a cigarette recommendation list if the score of the cold cigarette exceeds the score of at least one hot cigarette;
weighting and average dividing according to the preference of the user on similar cigarettes according to the similarity to obtain the number of cigarettes with larger scores, wherein the sum of the number of the lists of the cigarettes of both hot cigarettes and cold cigarettes is N;
obtaining the scores of the cigarettes in the cigarette list with the total number N screened by the current user according to the formula (6), and arranging the cigarettes in a descending order to obtain a final recommendation list;
and a final overall objective loss function through the following formula:
where J represents the cross entropy function, y 'represents the calculated similarity value of equation S (u', i '), y is the label of the current sample, 1 represents the positive sample, 0 represents the negative sample, (u, i') represents the randomly constructed negative sample, y represents the correct set of samples that are present themselves, (u, i) e y represents that the current (u, i) is the positive sample, and 
Representing that the current (u, i') does not belong to the correct sample, the positive sample without the superscript, the negative sample with the superscript, the set behind e represents the positive sample set, +_>
The latter set represents the negative sample set.
The invention provides a cigarette brand recommendation method based on a knowledge graph in a new retail model, which is characterized in that semantic information can be learned by using a translation model TransR after the knowledge graph subjected to proposal set enhancement pretreatment is complemented, so that richer semantic information can be learned, tail entities with higher accuracy can be predicted, and the embedding of one-to-many and many-to-many relations in knowledge graph representation learning is solved; obtaining the preference information of the current user on the cigarettes according to the user-cigarette interaction information and the preference matrix, further mining implicit preference which is more in line with the real situation, and finally recommending by combining with a tail entity, thereby realizing personalized recommendation aiming at high accuracy of the user; and (3) after the representation of the user and the cigarette is obtained in a scoring matrix of the user and the cigarette, introducing an attention mechanism, obtaining an aggregate vector between the user and the cigarette to mine implicit preference, expressing different preferences of the user on the cigarette, and obtaining an end-to-end recommendation model through combination with knowledge graph representation learning.
Drawings
For the purpose of making the objects, technical solutions and advantages of the present invention more apparent, the present invention will be further described with reference to the accompanying drawings, in which:
FIG. 1 is a flowchart of an embodiment of a method for recommending cigarette brands based on a knowledge graph in a new retail model provided by the invention;
FIG. 2 is a schematic diagram of a user-project knowledge graph after enhancement preprocessing of a proposed set;
FIG. 3 is a schematic diagram of a TransR translation model;
fig. 4 is a schematic diagram of preference mining of an embodiment of a method for recommending cigarette brands based on a knowledge graph in a new retail model provided by the invention;
fig. 5 is a schematic diagram of cigarette recommendation according to preference according to an embodiment of a method for recommending cigarette brands based on a knowledge graph in a new retail model provided by the invention.
Detailed Description
Various exemplary embodiments of the present disclosure will now be described in detail with reference to the accompanying drawings. The description of the exemplary embodiments is merely illustrative, and is in no way intended to limit the disclosure, its application, or uses. The present disclosure may be embodied in many different forms and is not limited to the embodiments described herein. These embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art. It should be noted that: the relative arrangement of parts and steps, the composition of materials, numerical expressions and numerical values set forth in these embodiments should be construed as exemplary only and not limiting unless otherwise specifically stated.
"first", "second", as used in this disclosure: and similar words are not to be interpreted in any order, quantity, or importance, but rather are used to distinguish between different sections. The word "comprising" or "comprises" and the like means that elements preceding the word encompass the elements recited after the word, and not exclude the possibility of also encompassing other elements. "upper", "lower", etc. are used merely to denote relative positional relationships, which may also change accordingly when the absolute position of the object to be described changes.
In this disclosure, when a particular element is described as being located between a first element and a second element, there may or may not be intervening elements between the particular element and the first element or the second element. When it is described that a specific component is connected to other components, the specific component may be directly connected to the other components without intervening components, or may be directly connected to the other components without intervening components.
All terms (including technical or scientific terms) used in this disclosure have the same meaning as commonly understood by one of ordinary skill in the art to which this disclosure belongs, unless specifically defined otherwise. It will be further understood that terms, such as those defined in commonly used dictionaries, should be interpreted as having a meaning that is consistent with their meaning in the context of the relevant art and will not be interpreted in an idealized or overly formal sense unless expressly so defined herein.
Techniques, methods, and apparatus known to one of ordinary skill in the relevant art may not be discussed in detail, but where appropriate, the techniques, methods, and apparatus should be considered part of the specification.
Currently, the most common strategy in recommendation methods is collaborative filtering, which uses historical interaction records of users and analyzes similar behavior of the users to make recommendations. However, collaborative filtering-based recommendation methods often suffer from sparsity of interaction data and cold start problems. Researchers often introduce external rich information to address these issues, such as: heterogeneous networks such as social networks, article attributes, knowledge graphs and the like are introduced. Among them, knowledge maps are more widely used. In the recommendation task based on the knowledge graph, the users and the articles can correspond to the nodes in the graph, and the relationship between the articles corresponds to the edges in the graph, so that the knowledge graph can not only provide deep semantic relationship between the articles and improve the accuracy of a recommendation system, but also expand the diversified interests of the users and provide diversity and interpretability for the recommendation system. Up to now, methods of applying a knowledge graph to a recommendation system can be classified into three types of feature-based methods, path-based methods, and graph-based methods. The feature-based method mainly extracts attributes of some users and objects from the knowledge graph as features, puts the attributes and the objects into a traditional model, introduces physical features and does not introduce relational features. DKN for the example of news recommendations, there are typically a large number of entities in the news headlines and text, and semantic relationships between entities can effectively expand user interests. However, such semantic relationships are difficult to explore by traditional methods (topic models, word vectors). The method comprises the steps of constructing a knowledge graph subgraph by entities in news, embedding the subgraph through a translation model, and fusing the knowledge graph and news recommendation by fusing entity vectors, heading word vectors and entity context vectors through a KCNN framework. The path-based approach treats the knowledge graph as a heterogeneous information network (heterogeneous information network) and then constructs meta-path or meta-graph based features between items. Briefly, meta-path is a specific path connecting two entities, such as "actor- > movie- > director- > movie- > actor" which can connect two actors, and thus can be considered as a way to mine potential relationships between actors. The advantage of this type of method is that the network structure of the knowledge graph is fully and intuitively utilized, and the disadvantage is that a meta-path or meta-graph needs to be designed manually, which is difficult to reach the optimum in practice; meanwhile, this type of method cannot be applied in a scenario where entities do not belong to the same domain (e.g., news recommendation), because we cannot predefine meta-path or meta-graph for such scenario. The graph-based method focuses on the association between nodes and neighbors in a knowledge graph, and common practice is to aggregate a certain entity node as a center so as to capture the characteristics of the nodes and the neighbors. The KGCN uses the graph convolution network to aggregate neighbor nodes to acquire k-hop structure information of the center node, and simultaneously weights the neighbors according to the connection relation and the specific user score to represent semantic information of the knowledge graph. NACF uses the attention mechanism in the GCN to compute the user's weight on the historically interacted items, respectively, and then aggregates the items using the graph rolling network to capture the user's preferences. Although the knowledge graph is widely applied to the recommendation system, due to the characteristics of high dimension and isomerism, how to effectively apply the external information in the knowledge graph to the recommendation task is still quite challenging. Most researchers at present divide the recommendation system and the knowledge graph embedding into two independent tasks and train in two different vector spaces respectively, and although the method is convenient for training, errors caused by different vector spaces can reduce the overall performance of the model.
For knowledge graph completion, although there are many methods to find entities and their relationships from text, the existing KG is far from complete. Recent studies on KG completions have shown that learning a low-dimensional representation of entities and relationships while maintaining structural knowledge of the graph is greatly facilitated. Such representation learning methods can be broadly divided into two categories: a translation distance model and a semantic matching model. The transition first proposes the core idea of translating the distance model, i.e. the relationship between two entities corresponds to one translation in their vector space. Although it is simple and effective, it is sometimes confusing because some relationships can convert one entity to a variety of entities, i.e., a 1-to-N problem. Likewise, there are other N-to-1 and N-to-N problems. To address these issues, many approaches extend the TransE by introducing additional hyperplane, vector space, text information, and relationship paths. The second group measures the rationality of facts by matching semantic representations of entities and relationships with a similarity-based scoring function. The RESCAL represents each relationship as a matrix to capture the combined semantics between the entities and uses bilinear functions as similarity measures. To simplify the learning of the relationship matrix, distMult limits them to diagonal lines, holE defines a circular correlation to compress the relationship matrix into vectors, while compiex introduces ComplEx values for asymmetric relationships. Another class of methods does not model the combinatorial relationships, but rather directly introduces NNs for matching. The SME learns the relationship-specific layers of the head and tail entities, respectively, and then feeds them to the final matching layer (e.g., point generation), while the NAM uses a deep architecture for semantic matching.
The existing recommending method based on the knowledge graph has the defects that on one hand, the recommending method based on the knowledge graph assumes that the current knowledge graph is complete, but in practice, the knowledge graph is impossible to complete, the fact that the knowledge graph lacks certain entities and relations is common, the incomplete knowledge graph lacks semantic information, and the recommending result is seriously influenced; on the other hand, the existing recommendation method based on the knowledge graph is characterized in that a recommendation system and the knowledge graph are embedded and divided into two independent tasks and are trained in two different vector spaces respectively, the method is convenient to train, errors brought by the different vector spaces can reduce the overall performance of a model, implicit preference between a user and cigarettes is not focused in the recommendation process, and accurate personalized recommendation cannot be achieved.
In view of this, the invention complements the knowledge graph and digs into the implicit preference between the user and the cigarette. As shown in fig. 1, in the actual implementation process, the method for recommending cigarette brands based on a knowledge graph in the new retail model provided in this embodiment specifically includes:
and S1, constructing a User-Item knowledge Graph (User-Item Graph) by taking the User, the cigarette and the related attributes between the User and the cigarette as nodes, and carrying out enhanced preprocessing on the User-Item knowledge Graph through a proposal set.
Wherein, as shown in fig. 2, in the user-project knowledge graph, the relationship between the nodes comprises an attribute-level relationship and a user-related relationship, and the attribute-level relationship comprises a brand or manufacturer, and the user-related relationship comprises a co-purchase or a co-view, and is taken as an edge of the knowledge graph, and the edges of all the knowledge graphs are formed into the proposal set.
And S2, predicting the missing entity in the user-project knowledge graph after the proposal set enhancement pretreatment based on a TransR translation model.
In one embodiment of the method for recommending cigarette brands based on a knowledge graph in the new retail model of the present invention, the step S2 may specifically include:
and S21, learning the expression of the entity and the relation in the user-project knowledge graph after the proposal set enhancement pretreatment based on a TransR translation model.
Specifically, the transformation matrix of the head entity and the tail entity in the user-project knowledge graph after the proposal set enhancement pretreatment is mapped into the same vector space according to the relation through a TransR translation model so as to set each relation as an independent space, the transformation matrix is expressed through the following formula so as to map the head entity and the tail entity into the vector space of the corresponding relation r,
e hr ＝e h M r (1)
e tr ＝e t M r (2)
Wherein M is r Matrix representing spatial projection, which is a representation of the relationship with the embedding of r itself, e hr Represented by projection matrix M r The original head entity vector e h E in the relation space mapped to the corresponding relation r tr Representing the tail entity e t Mapping into a relationship space of the corresponding relationship r.
As shown in fig. 3, changing the original TransH in the knowledge-enhanced translation-based user preference model (KTUP) to a TransR translation model, the TransR model models entities and relationships in two different spaces, namely, an entity space and a plurality of relationship spaces (relationship-specific entity spaces), and performs conversion in the corresponding relationship spaces has a better effect. While the TransH model allows each entity to have different representations under different relationships, it still holds that the entities and relationships are in the same semantic space, which limits to some extent the representation capabilities of the TransH, but one entity is a complex of multiple attributes, different relationships focus on different attributes of the entity, and different relationships have different semantic spaces. The main technology of TransH is to solve the problem that TransE cannot well process complex relations such as 1-n, n-1 and n-n, so that the Link Prediction on the whole data set is not greatly improved, and the Prediction accuracy on the associated class of the entities with similar attributes is improved with the clear color. However, for the fact that only the head and tail entities are projected into the hyperplane where the relationship exists, the many-to-many relationship still cannot be fully satisfied, the vectors which can be determined in the same hyperplane are far less than the space, and for the fusion of more relationships in a recommendation system, many similar head entities or tail entities can appear so as to influence the accuracy of recommendation, and more vector representations can be obtained by mapping the head and tail entities into the whole space. Specifically, the invention selects mapping the transformation matrix of the head entity and the tail entity according to the relation into the same vector space, and sets each relation as an independent space, so that similar preference or feature vectors obtained by embedding the relation are prevented from being too similar, as shown in a formula (1) and a formula (2).
And S22, predicting the missing entity in the user-project knowledge graph after the enhancement pretreatment of the proposal set according to the learning result of the entity and the relation.
In one embodiment of the method for recommending cigarette brands based on a knowledge graph in the new retail model of the present invention, the step S22 may specifically include:
step S221 of calculating the scores of the given entity, each predicted entity and the corresponding relation triplet by the following formula,
and step S222, taking the predicted entity corresponding to the minimum score as a tail entity in the user-project knowledge graph after the enhancement pretreatment of the proposal set.
For the scoring of triples, the smaller the score of the true triplet, and conversely, the larger the training, the more regular the random gradient drop (Stochastic Gradient Descent, SGD). Most important after this process is the prediction of the missing head or tail entities in the knowledge-graph, given (e h R) find the corresponding N entities in the remaining set of related entities such that f (e h ,e t R) is smallest, thus, the N entities can be treated asTail entity e for a given missing triplet t . According to the invention, the proposal set and the TransR model are introduced, the head and tail entities under a specific relationship are mapped into the relationship space, and under the specific relationship, the obtained tail entities are rich in semantic information of the current relationship, so that the subsequent preference mining is facilitated.
And S3, obtaining the preference information of the current user on the cigarettes according to the user-cigarette interaction information and a preference matrix, wherein the user-cigarette interaction information represents the cigarette information with the interaction history of the user, and the preference matrix represents the preference information of the user on the cigarettes.
In one embodiment of the method for recommending cigarette brands based on a knowledge graph in the new retail model of the present invention, the step S3 may specifically include:
and S31, aggregating the user and the cigarette information with the interaction history of the user through an attention mechanism to obtain a new user representation.
Specifically, as shown in FIG. 4, user expression information and item expression information obtained based on a scoring matrix obtained by an interactive table of users and items for representing recorded information of cigarettes having an interactive history with the users are aggregated by an attention mechanism to obtain a new user representation, and a user cigarette aggregation vector is calculated as the new user representation by the following formula,
W ij ＝Softmax(LeakyReLU(W att )(u i ||v ij )) (4)
wherein W is ij Representing a weight matrix calculated by an attention mechanism for representing the preference degree of the user for different cigarettes, u i Vector representation representing user, v ij Vector representation, W, of cigarettes representing history of interactions with a user att Weight matrix representing the attention mechanism, ||represents the splice vector of user and cigarette, and LeakyReLU represents the nonlinear activation function byThe Softmax function calculates the weight of attention, u r The user cigarette aggregate vector is represented as a new user vector.
When the hidden preference is mined, the information of the user expression u and the item expression i in the scoring matrix (such as the score of the purchased cigarette brand of the user) obtained based on the item interaction table corresponding to the user is aggregated through the attention mechanism, more information of the user for the preference combination of different cigarettes is obtained according to the mechanism, and better preference can be obtained through the feature vector. The specific process is shown in fig. 4.
And step S32, determining a preference matrix according to the reason that the user interacts with the cigarettes.
User preferences are for cigarettes, and in particular implementations may be broadly categorized based on statistical analysis of the data, such as preferences set by the place of origin, manufacturer, taste, smell, etc. The present user preference information for cigarettes is obtained according to the relation between the user and the cigarettes, and is also a vector representation, for example, the user purchases Yuxi, and needs to mine the preference of purchasing cigarettes, for example, like the smell, taste and the like of the cigarettes. In short, the reason for purchasing such cigarettes is an implicit relationship, and the preference matrix P is trained as a parameter in the model, and in some embodiments, the number of preferences for purchasing cigarettes by the user may be set to 5 through data research.
And step S33, determining the maximum preference information of the current user according to the similarity between the sum vector of the new user representation and the cigarette vector representation and the preference vector in the preference matrix.
As can be seen from fig. 5, the cigarette recommendation is performed according to the tail entity prediction and preference, in fig. 5, the cigarettes already purchased by the user are connected by solid lines, the tail entity prediction is performed by training the constructed knowledge graph after the proposed set enhancement pretreatment in the TransR translation model, the cigarettes (many cigarettes to be recommended can be provided in the figure, which is just an example) are connected by the tail entity prediction, and the cigarettes with close relation to the preference can be recommended by the mined preference because the solid lines or the dashed lines represent relations, wherein semantic information of the preference is included. When knowing why the user is buying this type of cigarette we can make accurate personalized recommendations according to the user's preferences. Given a pair of user cigarette interaction pairs (u, v), wherein the pair is a vector expression of the user u and the item v learned by a scoring matrix, wherein the preference of the user is included, and the similarity is obtained through an inner product among vectors so as to find the maximum preference p.
In one embodiment of the method for recommending cigarette brands based on a knowledge graph in the new retail model of the present invention, the step S33 may specifically include:
step S331, calculating the score of each cigarette of the current user according to the similarity of different cigarettes through the following formula,
wherein Score ui Representing the score of user u for cigarette i, i and j representing cigarettes, n representing the total number of cigarettes, sim (i, j) representing the similarity of cigarettes i and j calculated by cosine similarity, R uj Representing the score of user u for cigarette j.
And calculating the average score of the preference of the user on the similar cigarettes according to the scoring matrix, and selecting the similar cigarettes with large scores as recommendation candidates through the descending score.
Step S332 of determining implicit preference information p of the current user by calculation of the following formula,
S p∈P ＝(u+v)⊙p(7)
p＝argmax i (log(S i )+b)(8)
where u represents a new user representation of the current user, v represents a cigarette vector representation having an interaction history with the current user, P in equation (7) represents a vector corresponding to the preference matrix P, and the number of preferences is determined from the dataset, and b represents the bias.
When potential preferences are mined, firstly calculating the similarity between (u, v) and p, and finding the preference p with the maximum similarity result as the preference expression of the current user on cigarettes; by biasing b, the function flexibility can be increased, and the fitting capacity can be improved.
And S4, generating a cigarette recommendation list according to the implicit preference information of the current user, the relation among nodes obtained by mining based on the user-project graph knowledge graph and the missing entity predicted based on the user-project knowledge graph.
In the previous step S31, the user and item expression vectors learned in the scoring matrix are used to obtain a user cigarette aggregation vector (i.e. a mixed feature vector of the user) through an attention mechanism, wherein the preference of the user for cigarettes and the purchased cigarette information u are included r ' based on this, in step S41, the relationship r learned by the TransR and the implicit preference p among the mined cigarettes of the user are integrated to obtain the final preference expression of the user. In one embodiment of the method for recommending cigarette brands based on a knowledge graph in the new retail model of the present invention, the step S4 may specifically include:
step S41, obtaining the final preference expression of the current user according to the implicit preference information of the current user and the relation learned based on the TransR translation model through the following formula
And the final preference matrix of the current user->
Wherein r represents the relation in the knowledge graph, and w p Representing an overall weight matrix combining various preferences p, M r Matrix representing spatial projection。
Step S42, obtaining the final feature vector of the cigarette according to the expression i of the cigarette obtained by the scoring matrix and the entity vector e predicted by the user-project knowledge graph after the enhancement pretreatment based on the proposal set
Step S43, mapping the feature vector of the user and the feature vector of the cigarette to the vector space under the corresponding preference by the following formula
In the process,
wherein u is r ' represents the user cigarette aggregate vector calculated by the attention mechanism, u r Representation mapping to vector space
User representation of->
Representation mapping to vector space +.>
The cigarette of (3) is represented by a cigarette,
step S44, pair
A triplet, scored by a scoring function in the following formula,
and S45, selecting partial cold cigarettes and hot cigarettes for scoring, and obtaining a final cigarette recommendation list by using a TopN recommendation method according to the scoring result.
In one embodiment of the method for recommending cigarette brands based on a knowledge graph in the new retail model of the present invention, the step S45 may specifically include:
step S451, for the hot cigarettes, obtaining a scoring score according to a formula (14), and entering all the hot cigarettes into a cigarette recommendation list.
Step S452, for the cold door cigarettes, according to the embedded representation of the user and the cigarettes, determining the similarity between the cigarettes through the following formula,
where S (x, y) represents a similarity function reconstructed from the generated embedded representations of the user and cigarette to obtain a similarity value.
And step S453, screening out the characteristic representation of the cigarettes with small difference from the vector inner product value through the similarity function.
The invention realizes the recommendation of the cold cigarettes by setting the link prediction definition similarity function.
And step 454, obtaining the grading score of the user on the cold-door cigarettes screened out by the similarity function according to the formula (6).
Step S455, comparing the score of the cold cigarettes with the score of each hot cigarette, and adding the cold cigarettes into the cigarette recommendation list if the score of the cold cigarettes exceeds the score of at least one hot cigarette.
And step 456, weighting average division according to the preference of the user on similar cigarettes according to the similarity to obtain the number of cigarettes with larger scores, wherein the sum of the number of the lists of the cigarettes of both the hot cigarettes and the cold cigarettes is N.
And step 457, obtaining the score of the cigarettes in the cigarette list with the total number N screened by the current user according to the formula (6), and arranging the cigarettes in a descending order to obtain a final recommendation list.
Step S458, the final overall objective loss function is obtained by the following formula:
where J represents the cross entropy function, y 'represents the calculated similarity value of equation S (u', i '), y is the label of the current sample, 1 represents the positive sample, 0 represents the negative sample, (u, i') represents the randomly constructed negative sample, y represents the correct set of samples that are present themselves, (u, i) e y represents that the current (u, i) is the positive sample, and
representing that the current (u, i') does not belong to the correct sample, the positive sample without the superscript, the negative sample with the superscript, the set behind e represents the positive sample set, +_>
The latter set represents the negative sample set.
According to the cigarette brand recommendation method based on the knowledge graph in the new retail mode, provided by the embodiment of the invention, after the knowledge graph subjected to the proposal set enhancement pretreatment is completed, the translation model TransR is used for learning semantic information, so that richer semantic information can be learned, tail entities with higher accuracy can be predicted, and the embedding of one-to-many and many-to-many relations in knowledge graph representation learning is solved; obtaining the preference information of the current user on the cigarettes according to the user-cigarette interaction information and the preference matrix, further mining implicit preference which is more in line with the real situation, and finally recommending by combining with a tail entity, thereby realizing personalized recommendation aiming at high accuracy of the user; and (3) after the representation of the user and the cigarette is obtained in a scoring matrix of the user and the cigarette, introducing an attention mechanism, obtaining an aggregate vector between the user and the cigarette to mine implicit preference, expressing different preferences of the user on the cigarette, and obtaining an end-to-end recommendation model through combination with knowledge graph representation learning.
Thus, various embodiments of the present disclosure have been described in detail. In order to avoid obscuring the concepts of the present disclosure, some details known in the art are not described. How to implement the solutions disclosed herein will be fully apparent to those skilled in the art from the above description.
Although some specific embodiments of the present disclosure have been described in detail by way of example, it should be understood by those skilled in the art that the above examples are for illustration only and are not intended to limit the scope of the present disclosure. It will be understood by those skilled in the art that the foregoing embodiments may be modified and equivalents substituted for elements thereof without departing from the scope and spirit of the disclosure. The scope of the present disclosure is defined by the appended claims.
Claims (8)
1. A knowledge-graph-based brand recommendation method in a new retail model, comprising:
constructing a user-project knowledge graph by taking a user, a cigarette and related attributes between the user and the cigarette as nodes, and carrying out enhanced pretreatment on the user-project knowledge graph through a proposal set;
predicting the missing entity in the user-project knowledge graph after proposal set enhancement pretreatment based on a TransR translation model;
Obtaining the preference information of the current user on cigarettes according to the user-cigarette interaction information and a preference matrix, wherein the user-cigarette interaction information represents the cigarette information with the interaction history of the user, and the preference matrix represents the preference information of the user on cigarettes;
generating a cigarette recommendation list according to the implicit preference information of the current user, the relation among nodes obtained by mining based on the user-project graph knowledge graph and the missing entity predicted based on the user-project knowledge graph,
the method for obtaining the preference information of the current user on the cigarettes according to the user-cigarette interaction information and the preference matrix comprises the following steps:
aggregating the user and the cigarette information with the interaction history with the user through an attention mechanism to obtain a new user representation;
determining a preference matrix according to the reason that the user has interaction with the cigarettes;
determining maximum preference information of the current user according to the similarity between the sum vector of the new user representation and the cigarette vector representation and the preference vector in the preference matrix,
the method for aggregating the cigarette information of the user and the interaction history with the user through the attention mechanism to obtain a new user representation specifically comprises the following steps:
Aggregating user expression information and project expression information obtained based on a scoring matrix obtained by an interactive table of users and projects through an attention mechanism to obtain new user expression, wherein the interactive table is used for representing record information of cigarettes with interactive histories of the users, calculating a user cigarette aggregation vector as the new user expression through the following formula,
W ij ＝Soft max(Leaky ReLU(W att )(u i ||v ij )) (1)
wherein W is ij Representing a weight matrix calculated by an attention mechanism for representing the preference degree of the user for different cigarettes, u i Vector representation representing user, v ij Vector representation, W, of cigarettes representing history of interactions with a user att Weight matrix representing the attention mechanism, ||represents the splice vector of the user and the cigarette, leak ReLU represents the nonlinear activation function, weight of attention is calculated by Softmax function, u r Representing user cigarette aggregationThe vector is used as a new user vector.
2. The brand recommendation method based on a knowledge graph in a new retail model according to claim 1, wherein in the user-item knowledge graph, the relationship between the nodes includes an attribute-level relationship including a brand or vendor and a user-related relationship including a co-purchase or a co-view, and is used as an edge of the knowledge graph, and edges of all knowledge graphs are combined into the proposal set.
3. The brand recommendation method based on knowledge-graph in new retail model of claim 1, wherein said trans-r translation model predicts missing entities in said user-project knowledge graph after proposal set enhancement preprocessing, specifically comprising:
learning the expression of the entity and the relation in the user-project knowledge graph after the proposal set is subjected to the enhancement pretreatment based on a TransR translation model;
and predicting the missing entity in the user-project knowledge graph after the proposal set is subjected to the enhancement pretreatment according to the learning result of the entity and the relation.
4. The brand recommendation method based on knowledge graph in new retail model according to claim 3, wherein said trans-r translation model learns the expression of entities and relations in said user-project knowledge graph after enhancement preprocessing of proposed sets, specifically comprising:
mapping the transformation matrix of the head entity and the tail entity in the user-project knowledge graph subjected to proposal enhancement pretreatment into the same vector space through a TransR translation model so as to set each relation as an independent space, and mapping the head entity and the tail entity into the vector space of the corresponding relation r through the following formula,
e hr ＝e h M r (3)
e tr ＝e t M r (4)
Wherein M is r Matrix representing spatial projection, which is a representation of the relationship with the embedding of r itself, e hr Represented by projection matrix M r The original head entity vector e h E in the relation space mapped to the corresponding relation r tr Representing the tail entity e t Mapping into a relationship space of the corresponding relationship r.
5. The brand recommendation method based on knowledge graph in new retail model of claim 4, wherein predicting missing entities in the user-project knowledge graph after enhancement preprocessing of proposal set according to learning result of entities and relations specifically comprises:
the scores for a given entity, each predicted entity and corresponding relationship triplet are calculated by the following formula,
and taking the predicted entity corresponding to the minimum score as a tail entity in the user-project knowledge graph after the enhancement pretreatment of the proposal set.
6. The brand recommendation method based on a knowledge graph in a new retail model of claim 1, wherein determining implicit preference information of a current user according to similarity of a sum vector of new user representations and cigarette vector representations and preference vectors in the preference matrix specifically comprises:
According to the similarity of different cigarettes, the scoring score of the current user on each cigarette is calculated through the following formula,
wherein Score ui Representing the score of user u for cigarette i, i and j representing cigarettes, n representing the total number of cigarettes, sim (i, j) representing the similarity of cigarettes i and j calculated by cosine similarity, R uj Representing the score of user u for cigarette j;
the implicit preference information p of the current user is determined through the following formula calculation,
S p∈P ＝(u+v)⊙p (7)
p＝argmax i (log(S i )+b) (8)
where u represents a new user representation of the current user, v represents a cigarette vector representation having an interaction history with the current user, P in equation (7) represents a vector corresponding to the preference matrix P, and the number of preferences is determined from the dataset, and b represents the bias.
7. The brand recommendation method based on a knowledge graph in a new retail model of claim 6, wherein the generating a cigarette recommendation list according to the implicit preference information of the current user, the inter-node relationship obtained based on the user-project graph knowledge graph mining, and the missing entity predicted based on the user-project knowledge graph specifically comprises:
obtaining final preference expression of the current user according to the implicit preference information of the current user and the relation learned based on the TransR translation model through the following formula 
And the final preference matrix of the current user->
Wherein r represents the relation in the knowledge graph, and w p Representing an overall weight matrix combining various preferences p, M r A matrix representing a projection of a space is presented,
obtaining the final feature vector of the cigarette according to the expression i of the cigarette obtained by the scoring matrix and the entity vector e predicted by the user-project knowledge graph after the enhancement pretreatment based on the proposal set
Mapping the feature vectors of the user and the feature vectors of the cigarettes to a vector space under corresponding preferences by the following formula
In the process,
wherein u is r ' represents the user cigarette aggregate vector calculated by the attention mechanism, u r Representation mapping to vector space
User representation of->
Representation mapThe space of the shot vector->
The cigarette of (3) is represented by a cigarette,
for a pair of
A triplet, scored by a scoring function in the following formula,
and selecting partial cold cigarettes and hot cigarettes for scoring, and obtaining a final cigarette recommendation list by using a TopN recommendation method according to the scoring result.
8. The brand recommendation method based on a knowledge graph in a new retail model of claim 7, wherein the selecting part of the cold cigarettes and the hot cigarettes for scoring, and obtaining a final cigarette recommendation list by using a TopN recommendation method according to the scoring result, specifically comprises:
For hot cigarettes, a scoring score is obtained by equation (14), all hot cigarettes enter the cigarette recommendation list,
for a cold-door cigarette, the similarity between each cigarette is determined according to the embedded representation of the user and the cigarette by the following formula,
wherein S (x, y) represents reconstructing a similarity function from the generated embedded representations of the user and the cigarette to obtain a similarity value;
screening out the characteristic representation of the cigarettes with small difference from the vector inner product value through the similarity function;
obtaining the grading score of the user on the cold cigarettes screened out by the similarity function according to the formula (6);
comparing the score of the cold cigarette with the score of each hot cigarette, and adding the cold cigarette into a cigarette recommendation list if the score of the cold cigarette exceeds the score of at least one hot cigarette;
weighting and average dividing according to the preference of the user on similar cigarettes according to the similarity to obtain the number of cigarettes with larger scores, wherein the sum of the number of the lists of the cigarettes of both hot cigarettes and cold cigarettes is N;
obtaining the scores of the cigarettes in the cigarette list with the total number N screened by the current user according to the formula (6), and arranging the cigarettes in a descending order to obtain a final recommendation list;
And a final overall objective loss function through the following formula:
where J represents the cross entropy function, y 'represents the calculated similarity value of equation S (u', i '), y is the label of the current sample, 1 represents the positive sample, 0 represents the negative sample, (u, i') represents the randomly constructed negative sample, y represents the correct set of samples that are present themselves, (u, i) e y represents that the current (u, i) is the positive sample, and
representing that the current (u, i') does not belong to the correct sample, the positive sample without the superscript, the negative sample with the superscript, the set following e represents the positive sample set,
the latter set represents the negative sample set. />
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211305797.7A CN115618108B (en) | 2022-10-24 | 2022-10-24 | Brand recommendation method based on knowledge graph in new retail model |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211305797.7A CN115618108B (en) | 2022-10-24 | 2022-10-24 | Brand recommendation method based on knowledge graph in new retail model |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN115618108A CN115618108A (en) | 2023-01-17 |
| CN115618108B true CN115618108B (en) | 2023-05-26 |
Family
ID=84864892
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211305797.7A Active CN115618108B (en) | 2022-10-24 | 2022-10-24 | Brand recommendation method based on knowledge graph in new retail model |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115618108B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116308683B (en) * | 2023-05-17 | 2023-08-04 | 武汉纺织大学 | Knowledge-graph-based clothing brand positioning recommendation method, equipment and storage medium |
| CN116756203B (en) * | 2023-06-13 | 2024-06-11 | 淮阴工学院 | Knowledge graph driving cold chain logistics equipment type selection method and device based on carbon emission reduction target |
| CN117034126B (en) * | 2023-10-09 | 2024-09-20 | 广东电力交易中心有限责任公司 | Green power user classification and feature extraction method and system based on knowledge graph |
Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111507796A (en) * | 2020-04-07 | 2020-08-07 | 桂林电子科技大学 | Online shopping mall commodity recommendation method based on knowledge graph |
| CN113158033A (en) * | 2021-03-19 | 2021-07-23 | 浙江工业大学 | Collaborative recommendation model construction method based on knowledge graph preference propagation |
| CN113190593A (en) * | 2021-05-12 | 2021-07-30 | 《中国学术期刊（光盘版）》电子杂志社有限公司 | Search recommendation method based on digital human knowledge graph |
| CN113342994A (en) * | 2021-07-05 | 2021-09-03 | 成都信息工程大学 | Recommendation system based on non-sampling cooperative knowledge graph network |
| CN113934936A (en) * | 2021-10-22 | 2022-01-14 | 南京邮电大学 | Recommendation algorithm fusing knowledge graph convolutional network and user preference |
| CN114186119A (en) * | 2020-09-15 | 2022-03-15 | 阿里巴巴集团控股有限公司 | Information recommendation method and device and method and device for constructing neighborhood knowledge graph |
| CN114399251A (en) * | 2021-12-30 | 2022-04-26 | 淮阴工学院 | Cold-chain logistics recommendation method and device based on semantic network and cluster preference |
| WO2022116537A1 (en) * | 2020-12-04 | 2022-06-09 | 上海帜讯信息技术股份有限公司 | News recommendation method and apparatus, and electronic device and storage medium |
| CN114820139A (en) * | 2022-05-25 | 2022-07-29 | 重庆大学 | Multi-user recommendation system based on knowledge graph path reasoning |
| CN114817712A (en) * | 2022-04-15 | 2022-07-29 | 华南理工大学 | Project recommendation method based on multitask learning and knowledge graph enhancement |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11017038B2 (en) * | 2017-09-29 | 2021-05-25 | International Business Machines Corporation | Identification and evaluation white space target entity for transaction operations |
| CN110955834B (en) * | 2019-11-27 | 2023-07-07 | 西北工业大学 | Knowledge graph driven personalized accurate recommendation method |
-
2022
- 2022-10-24 CN CN202211305797.7A patent/CN115618108B/en active Active
Patent Citations (10)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN111507796A (en) * | 2020-04-07 | 2020-08-07 | 桂林电子科技大学 | Online shopping mall commodity recommendation method based on knowledge graph |
| CN114186119A (en) * | 2020-09-15 | 2022-03-15 | 阿里巴巴集团控股有限公司 | Information recommendation method and device and method and device for constructing neighborhood knowledge graph |
| WO2022116537A1 (en) * | 2020-12-04 | 2022-06-09 | 上海帜讯信息技术股份有限公司 | News recommendation method and apparatus, and electronic device and storage medium |
| CN113158033A (en) * | 2021-03-19 | 2021-07-23 | 浙江工业大学 | Collaborative recommendation model construction method based on knowledge graph preference propagation |
| CN113190593A (en) * | 2021-05-12 | 2021-07-30 | 《中国学术期刊（光盘版）》电子杂志社有限公司 | Search recommendation method based on digital human knowledge graph |
| CN113342994A (en) * | 2021-07-05 | 2021-09-03 | 成都信息工程大学 | Recommendation system based on non-sampling cooperative knowledge graph network |
| CN113934936A (en) * | 2021-10-22 | 2022-01-14 | 南京邮电大学 | Recommendation algorithm fusing knowledge graph convolutional network and user preference |
| CN114399251A (en) * | 2021-12-30 | 2022-04-26 | 淮阴工学院 | Cold-chain logistics recommendation method and device based on semantic network and cluster preference |
| CN114817712A (en) * | 2022-04-15 | 2022-07-29 | 华南理工大学 | Project recommendation method based on multitask learning and knowledge graph enhancement |
| CN114820139A (en) * | 2022-05-25 | 2022-07-29 | 重庆大学 | Multi-user recommendation system based on knowledge graph path reasoning |
Non-Patent Citations (1)
| Title |
|---|
| 融合实体描述及类型的知识图谱表示学习方法;杜文倩;李弼程;王瑞;;中文信息学报(07);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN115618108A (en) | 2023-01-17 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115618108B (en) | Brand recommendation method based on knowledge graph in new retail model | |
| Chen et al. | Air: Attentional intention-aware recommender systems | |
| Li et al. | A survey of graph neural network based recommendation in social networks | |
| Liu et al. | Trust beyond reputation: A computational trust model based on stereotypes | |
| CN112364976A (en) | User preference prediction method based on session recommendation system | |
| CN112541132A (en) | Cross-domain recommendation method based on multi-view knowledge representation | |
| CN113051468B (en) | Movie recommendation method and system based on knowledge graph and reinforcement learning | |
| CN113918834B (en) | Graph convolution collaborative filtering recommendation method fusing social relations | |
| CN112507246A (en) | Social recommendation method fusing global and local social interest influence | |
| Zarzour et al. | RecDNNing: a recommender system using deep neural network with user and item embeddings | |
| CN115329215A (en) | Recommendation method and system based on self-adaptive dynamic knowledge graph in heterogeneous network | |
| CN116842258A (en) | Group recommendation method for feature enhancement based on meta-path context | |
| CN114329222B (en) | Film recommendation method integrating attention mechanism and double-end knowledge graph | |
| Dridi et al. | Exploiting context-awareness and multi-criteria decision making to improve items recommendation using a tripartite graph-based model | |
| Zhang et al. | Graph-aware collaborative reasoning for click-through rate prediction | |
| Li et al. | A knowledge-based recommendation model utilizing formal concept analysis and association | |
| Zhang et al. | Knowledge graph driven recommendation model of graph neural network | |
| Luo et al. | A hybrid user profile model for personalized recommender system with linked open data | |
| CN116401450A (en) | Recommendation method and device based on knowledge graph and computer equipment | |
| CN116383509A (en) | Movie recommendation method and device based on multi-attribute and knowledge attention fusion | |
| Chen et al. | Trust-based collaborative filtering algorithm in social network | |
| CN115391555A (en) | User-perceived knowledge map recommendation system and method | |
| Yang et al. | A human-in-the-loop approach to social behavioral targeting | |
| CN114970854A (en) | Multi-view dynamic network link prediction method and system based on time sequence neighborhood aggregation | |
| Fu et al. | Preference-aware heterogeneous graph neural networks for recommendation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |