A recommendation system and method based on an e-commerce coupon domain knowledge graph
By constructing a knowledge graph and TransE model in the field of e-commerce coupons, and combining new users' friend information and historical interaction data, the low accuracy and cold start problem of new user recommendations were solved, thus improving the effectiveness of coupon recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- HARBIN ENG UNIV
- Filing Date
- 2023-05-17
- Publication Date
- 2026-07-10
AI Technical Summary
Existing recommendation systems are ineffective at recommending coupons to new users in the field of e-commerce coupons, and suffer from low recommendation accuracy and reliability, as well as data sparsity and cold start problems.
We construct a knowledge graph based on the e-commerce coupon domain. Through user query module, knowledge base module, composite information enhancement module, friend information query module, and popular product rating module, we use the TransE model to recommend coupons and combine the friend information and historical interaction data of new users to build a knowledge graph for new users.
It improved the accuracy and reliability of coupon recommendations, increased the proportion of users using coupons and their exposure, alleviated the cold start problem, and provided an effective recommendation solution for new users.
Smart Images
Figure CN116450854B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of knowledge graph and recommendation technology, and in particular to a recommendation system and method based on a knowledge graph in the field of e-commerce coupons. Background Technology
[0002] In recent years, the application areas of internet technology have been continuously expanding, with e-commerce being one of the most important. Recommendation system technology plays a crucial role in e-commerce, and it has been widely applied to various aspects of our lives, such as online shopping, short video viewing, music listening, and news browsing. These behaviors all rely on the support of recommendation system technology. For example, e-commerce platforms utilize and analyze massive amounts of user purchasing behavior data, using recommendation system technology to summarize user behavior patterns. These patterns are then integrated with the platform's marketing strategies to provide users with personalized product recommendations. This approach not only promotes sales and improves user retention rates but also enables the development of more scientific and effective e-commerce marketing strategies.
[0003] Researchers have proposed several relatively mature recommendation methods, namely content-based recommendation algorithms and collaborative filtering-based recommendation methods. The core idea of content-based recommendation algorithms is to use user preference records to recommend other products with similar attributes to those recorded. Collaborative filtering recommendation algorithms utilize users' historical behavior to recommend products. The core idea is that if two users have similar historical behaviors, their future purchasing behavior should also be similar, meaning they are likely to buy the same products, thus completing the product recommendation. However, due to the significant differences between the coupon domain and conventional domains, and the unclear association between coupon entity types, recommendation methods in other domains have not fully considered the business specificities of the coupon domain, thus making it impossible to directly use conventional methods to recommend coupons. Furthermore, current recommendation system technology mainly faces the following two challenges: First, the number of coupons actually used by users only accounts for a small portion of the total number issued by the platform, and the publicly available user interaction data on e-commerce platforms is extremely limited, leading to data sparsity. This results in the system's inability to accurately capture the implicit information contained in the dataset, and also greatly increases the difficulty for the model to learn features from the dataset. These shortcomings lead to high unreliability and low recommendation accuracy. On the other hand, traditional recommendation systems lack historical interaction data for new users, thus they cannot directly make recommendations for new users, which is known as the cold start problem. Summary of the Invention
[0004] The purpose of this invention is to address the problems of existing business coupon recommendation methods, such as the inability to recommend coupons to new users and low reliability and accuracy. Therefore, this invention proposes a recommendation system and method based on a knowledge graph in the field of e-commerce coupons.
[0005] A recommendation system based on a knowledge graph in the field of e-commerce coupons includes: a user query module, a knowledge base module, a composite information enhancement module, a friend information query module, and a popular product scoring module;
[0006] The knowledge base module stores knowledge graphs corresponding to users of the e-commerce platform.
[0007] The knowledge graph of the e-commerce platform is in the form of: entity-click relationship-entity;
[0008] The user query module is used to obtain the current user information and use the current user information in the knowledge base module to determine whether the current user is a new user. If it is a new user, it will enter the friend information query module; if it is an old user, the knowledge graph corresponding to the current old user stored in the knowledge base module will be input into the composite information enhancement module.
[0009] The composite information enhancement module is used to obtain a user-coupon score list using the user's corresponding knowledge graph, thereby recommending coupons.
[0010] The friend information query module is used to obtain the old user information entered by the new user, obtain the knowledge graph corresponding to the old user in the knowledge base module, and send the knowledge graph corresponding to the old user to the popular product scoring module.
[0011] The popular product rating module is used to obtain the ratings of popular products on the e-commerce platform by current new users, and to construct a new user knowledge graph using the current new user ratings of popular products on the e-commerce platform and the knowledge graphs corresponding to old users. The new user knowledge graph is then input into the knowledge base module and the composite information enhancement module.
[0012] Furthermore, the entities in the knowledge graph include: users, products, and coupons; the click relationships include: use, purchase at the original price, and purchase at a discount.
[0013] Furthermore, the user query module is used to obtain current user information and, in the knowledge base module, to determine whether the current user is a new user, specifically:
[0014] Get the current user ID, and use the user ID to perform a matching query in the knowledge base module. If the user ID is found, the current user is an old user; otherwise, the current user is a new user.
[0015] Furthermore, the composite information enhancement module is used to obtain a user-coupon score list using a knowledge graph, thereby recommending coupons, specifically as follows:
[0016] Step 1: Initialize and normalize the knowledge graph corresponding to the user to obtain a vector representation of the user's behavior information;
[0017] Step 2: Use the vector representation of user behavior information as positive examples, and use the positive examples to construct negative examples. Use the positive examples and negative examples to form a training set.
[0018] The negative example is a vector of erroneous user behavior information that is artificially constructed.
[0019] Step 3: Input the training set into the TransE model to train the TransE model, continuously update the entity and relation vector values in the user behavior information vector, until the preset training rounds are reached, the update ends, and the recommendation tuple is obtained.
[0020] Step 4: Use a scoring function to score the recommended tuples, and sort the recommended tuples from highest to lowest score to obtain a user-coupon score list. Recommend coupon information in order according to the list.
[0021] Furthermore, the TransE model is trained using the cross-entropy loss function.
[0022] A recommendation method based on a knowledge graph in the e-commerce coupon domain includes the following steps:
[0023] S1. Construct a knowledge graph corresponding to the current users of the e-commerce platform and obtain a knowledge graph library;
[0024] The knowledge graph is in the form of: entity-click relationship-entity;
[0025] S2. Obtain the current user information and search for the current user in the knowledge graph database constructed in S1. If the current user is found, the current user is an old user, and S3 is executed; if the current user is not found, the current user is a new user, and S4 is executed.
[0026] S3. Obtain the knowledge graph corresponding to the current user, and use the knowledge graph corresponding to the current user to obtain the user-coupon list, so as to make coupon recommendation information;
[0027] S4. Obtain the information of old user A input by the new user, and obtain the knowledge graph corresponding to A from the knowledge graph library constructed in S1;
[0028] S5. Obtain the ratings of current new users for popular products on the e-commerce platform. Use the knowledge graph corresponding to A obtained in S4 and the ratings of current new users for popular products on the e-commerce platform to construct a new user knowledge graph, and then execute S3; at the same time, save the new user knowledge graph to the knowledge graph library.
[0029] Furthermore, the entities in the knowledge graph corresponding to the current e-commerce platform user in S1 include: user, product, and coupon; the click relationships include: use, purchase at original price, and purchase at a discount.
[0030] Furthermore, in step S3, obtaining the knowledge graph corresponding to the current user and using it to obtain a user-coupon list for coupon recommendation information specifically involves:
[0031] S301. Obtain the knowledge graph corresponding to the current user, initialize and normalize the knowledge graph corresponding to the current user to obtain a vector representation of the user's behavior information;
[0032] S302. Use the vector representation of user behavior information as positive examples, and use the positive examples to construct negative examples. Use the positive examples and negative examples to form a training set.
[0033] The negative example is a vector of erroneous user behavior information that is artificially constructed.
[0034] S303. Input the training set into the TransE model, train the TransE model, continuously update the entity and relation vector values in the user behavior information vector, reach the preset training rounds, the update ends, and the recommendation tuple is obtained.
[0035] S304. Use a scoring function to score the recommended tuples, and sort the recommended tuples from high to low according to their scores to obtain a user-coupon score list. Recommend coupon information in order according to the list.
[0036] Furthermore, the TransE model is trained using the cross-entropy loss function.
[0037] The beneficial effects of this invention are as follows:
[0038] This invention constructs a knowledge graph based on users' past interaction behavior, forming a knowledge base. Utilizing this knowledge base for coupon recommendations effectively increases the proportion of coupons actually used by users compared to those issued by the platform, improving the accuracy and reliability of coupon recommendations and significantly increasing coupon exposure and redemption rates. Furthermore, this invention uses the friend information provided by new users as the foundation for the new user's knowledge graph, constructing historical user interaction data that can be directly used for recommendations. This effectively alleviates the cold start problem of conventional recommendation systems and provides a solution to the cold start issue for new users. Attached Figure Description
[0039] Figure 1 This is a system framework diagram of the present invention;
[0040] Figure 2 This is a schematic diagram illustrating the working principle of the composite information enhancement module;
[0041] Figure 3 This is a structural diagram of the friend information query and popular product rating module. Detailed Implementation
[0042] Specific implementation method one: as follows Figure 1 As shown, this embodiment of a recommendation system based on a knowledge graph in the field of e-commerce coupons includes: a user query module, a knowledge base module, a composite information enhancement module, a friend information query module, and a popular product scoring module;
[0043] The user query module is used to obtain current user information and, in the knowledge base module, to determine whether the current user is a new user. If the user is a new user, the module proceeds to the friend information query module; if the user is an existing user, the knowledge graph corresponding to the current existing user stored in the knowledge base module is input into the composite information enhancement module.
[0044] The system retrieves user information provided by the user and performs a matching query in the knowledge base module (using a query statement to query the user ID). If the user information is matched in the knowledge base module, the current user is an old user, and the system proceeds to the composite information enhancement module. If the user information is not matched in the knowledge base module, the current user is a new user, and the system proceeds to the friend information query module.
[0045] The knowledge base module is used to store a knowledge graph (entity-click relationship) constructed from user-product / coupon interaction data of the e-commerce platform; it is responsible for providing data to the composite information enhancement module, wherein the data needs to be converted into low-dimensional vector form to participate in the training of the composite information enhancement module;
[0046] The knowledge base module stores entities from the knowledge graph, including users, products, coupons, etc.; click relationships include: use, purchase at full price, purchase at a discount; these entities and relationships together form five types of triples, representing user interactions such as using coupons, user discount purchases of products, and user clicks on products. By representing users and items such as coupons as entities in the knowledge graph, and representing the interactions between users and items as relationships in the graph, the rich semantic information of entities and relationships in the graph can effectively alleviate the data sparsity problem.
[0047] The composite information enhancement module is used to obtain a user-coupon score list using the user's corresponding knowledge graph, thereby obtaining coupon recommendation information:
[0048] Step 1: Initialize and normalize the knowledge graph, embed the knowledge graph into a low-dimensional space, and obtain a vector representation (h, l, t) of user behavior information;
[0049] Step 2: Construct positive and negative examples (h, l, t') using the vector representation of user behavior information to obtain the training set;
[0050] Positive examples are triplet information that actually exists in the knowledge base, while negative examples are incorrect triplets that are artificially constructed (i.e. triplets after changing the item embedding vector).
[0051] Step 3: Input the training set into the TransE model to train the TransE model, continuously updating the entity and relation vector values in the user behavior information vector. Once the preset training rounds are reached, the update ends and the recommendation tuple is obtained.
[0052] The TransE model is trained using the cross-entropy loss function;
[0053] The recommendation tuple is in the form of: updated user embedding vector (h*) - updated relation embedding vector (l*) - updated item embedding vector (t*);
[0054] Step 4: Use a scoring function to score the recommended tuples, and sort the recommended tuples from high to low according to their scores to obtain a user-coupon score list. Recommend coupon information in order according to the list.
[0055] The useful information query module is used to obtain information input by new users into old users, retrieve the knowledge graph corresponding to the old users from the knowledge base module, and send the knowledge graph corresponding to the old users to the popular product rating module; it also retrieves personal information of old users input by new users, filters the interaction behavior data of friends (old users) with items from the knowledge base module, and uses the interaction behavior data of friends as the interaction behavior data of the current new user (using coupons, purchasing products at discounted / original prices, etc.);
[0056] The popular product rating module is used to obtain the ratings of popular products on the e-commerce platform by current new users, and to construct a new user knowledge graph using the current new user ratings of popular products on the e-commerce platform and the knowledge graph corresponding to old users. The new user knowledge graph is then input into the knowledge base module and the composite information enhancement module.
[0057] The system obtains ratings from new users for popular products on the platform, constructs a knowledge graph using these ratings and user interaction data, and inputs the constructed knowledge graph into a composite information enhancement module to obtain recommended coupon information.
[0058] like Figure 2 As shown, the model first performs low-dimensional space embedding on entities and relations in the knowledge base, and completes initialization and normalization operations to obtain the mapping vector of text information in low-dimensional space. Then, it constructs negative examples using triples in the knowledge base, obtaining a training dataset consisting of positive and negative triples. This dataset is used as input to the composite information augmentation module for training, continuously updating the entity and relation vector values until the training process ends.
[0059] like Figure 3 As shown, the friend information query allows new users to input personal information of familiar users they know. The recommendation system then uses the historical interaction data of these familiar users to model the new user. The popular product rating module serves the same purpose as the friend information query module: to use historical interaction information within the graph to model the new user. Specifically, new users rate popular products, with higher scores indicating greater interest. The system then uses the interaction data of familiar users who have previously purchased products the new user likes to construct the new user's interaction data.
[0060] Specific Implementation Method Two: A recommendation method based on a knowledge graph in the e-commerce coupon domain, comprising the following steps:
[0061] S1. Construct a knowledge graph corresponding to the current users of the e-commerce platform and obtain a knowledge graph library;
[0062] The knowledge graph is in the form of: entity-click relationship-entity;
[0063] Entities include: users, products, and coupons;
[0064] Click relationships include: Use, Purchase at full price, Purchase at a discount.
[0065] S2. Obtain the current user information and search for the current user in the knowledge graph database constructed in S1. If the current user is found, the current user is an old user, and S3 is executed; if the current user is not found, the current user is a new user, and S4 is executed.
[0066] S3. Obtain the knowledge graph corresponding to the current user, and use the knowledge graph to obtain the user-coupon list, thereby providing coupon recommendation information:
[0067] S301. Obtain the knowledge graph corresponding to the current user, initialize and normalize the knowledge graph corresponding to the current user to obtain a vector representation of the user's behavior information;
[0068] S302. Use the vector representation of user behavior information as positive examples, and use the positive examples to construct negative examples. Use the positive examples and negative examples to form a training set.
[0069] The negative example is a vector of erroneous user behavior information that is artificially constructed.
[0070] S303. Input the training set into the TransE model, train the TransE model, continuously update the entity and relation vector values in the user behavior information vector, reach the preset training rounds, the update ends, and the recommendation tuple is obtained.
[0071] The TransE model is trained using the cross-entropy loss function;
[0072] S304. Use a scoring function to score the recommended tuples, and sort the recommended tuples from high to low according to their scores to obtain a user-coupon score list. Recommend coupon information in order according to the list.
[0073] S4. Obtain the information of old user A input by the new user, and obtain the knowledge graph corresponding to A from the knowledge graph library constructed in S1;
[0074] S5. Obtain the ratings of current new users for popular products on the e-commerce platform. Use the knowledge graph corresponding to A obtained in S4 and the ratings of current new users for popular products on the e-commerce platform to construct a new user knowledge graph, and then execute S3; at the same time, save the new user knowledge graph to the knowledge graph library.
Claims
1. A recommendation system based on a knowledge graph in the field of e-commerce coupons, characterized in that... The system includes: a user query module, a knowledge base module, a composite information enhancement module, a friend information query module, and a popular product rating module; The knowledge base module stores knowledge graphs corresponding to users of the e-commerce platform. The knowledge graph of the e-commerce platform is in the form of: entity-click relationship-entity; The user query module is used to obtain the current user information and use the current user information in the knowledge base module to determine whether the current user is a new user. If it is a new user, it will enter the friend information query module; if it is an old user, the knowledge graph corresponding to the current old user stored in the knowledge base module will be input into the composite information enhancement module. The composite information enhancement module is used to obtain a user-coupon score list using the user's corresponding knowledge graph, thereby recommending coupons. Specifically: Step 1: Initialize and normalize the knowledge graph corresponding to the user to obtain a vector representation of the user's behavior information; Step 2: Use the vector representation of user behavior information as positive examples, and use the positive examples to construct negative examples. Use the positive examples and negative examples to form a training set. The negative example is a vector of erroneous user behavior information that is artificially constructed. Step 3: Input the training set into the TransE model to train the TransE model, continuously update the entity and relation vector values in the user behavior information vector, until the preset training rounds are reached, the update ends, and the recommendation tuple is obtained. Step 4: Use a scoring function to score the recommended tuples, and sort the recommended tuples from high to low according to their scores to obtain a user-coupon score list. Recommend coupon information in order according to the list. The friend information query module is used to obtain the old user information entered by the new user, obtain the knowledge graph corresponding to the old user in the knowledge base module, and send the knowledge graph corresponding to the old user to the popular product scoring module. The popular product rating module is used to obtain the ratings of popular products on the e-commerce platform by current new users, and to construct a new user knowledge graph using the current new user ratings of popular products on the e-commerce platform and the knowledge graphs corresponding to old users. The new user knowledge graph is then input into the knowledge base module and the composite information enhancement module.
2. The recommendation system based on a knowledge graph in the e-commerce coupon domain according to claim 1, characterized in that: The entities in the knowledge graph include: users, products, and coupons; the click relationships include: use, purchase at original price, and purchase at a discount.
3. A recommendation system based on a knowledge graph in the e-commerce coupon domain according to claim 2, characterized in that: The user query module is used to obtain current user information and, in the knowledge base module, to determine whether the current user is a new user. Specifically: Get the current user ID, and use the user ID to perform a matching query in the knowledge base module. If the user ID is found, the current user is an old user; otherwise, the current user is a new user.
4. A recommendation system based on a knowledge graph in the e-commerce coupon domain according to claim 3, characterized in that: The TransE model is trained using the cross-entropy loss function.
5. A recommendation method based on a knowledge graph in the e-commerce coupon domain, characterized in that: The method includes the following steps: S1. Construct a knowledge graph corresponding to the current users of the e-commerce platform and obtain a knowledge graph library; The knowledge graph is in the form of: entity-click relationship-entity; S2. Obtain the current user information and search for the current user in the knowledge graph database constructed in S1. If the current user is found, the current user is an old user, and S3 is executed; if the current user is not found, the current user is a new user, and S4 is executed. S3. Obtain the knowledge graph corresponding to the current user, and use the knowledge graph to obtain the user-coupon list, thereby providing coupon recommendation information. Specifically: S301. Obtain the knowledge graph corresponding to the current user, initialize and normalize the knowledge graph corresponding to the current user to obtain a vector representation of the user's behavior information; S302. Use the vector representation of user behavior information as positive examples, and use the positive examples to construct negative examples. Use the positive examples and negative examples to form a training set. The negative example is a vector of erroneous user behavior information that is artificially constructed. S303. Input the training set into the TransE model, train the TransE model, continuously update the entity and relation vector values in the user behavior information vector, reach the preset training rounds, the update ends, and the recommendation tuple is obtained. S304. Use a scoring function to score the recommended tuples, and sort the recommended tuples from high to low according to their scores to obtain a user-coupon score list. Recommend coupon information in order according to the list. S4. Obtain the information of old user A input by the new user, and obtain the knowledge graph corresponding to A from the knowledge graph library constructed in S1; S5. Obtain the ratings of current new users for popular products on the e-commerce platform. Use the knowledge graph corresponding to A obtained in S4 and the ratings of current new users for popular products on the e-commerce platform to construct a new user knowledge graph, and then execute S3; at the same time, save the new user knowledge graph to the knowledge graph library.
6. The recommendation method based on a knowledge graph in the e-commerce coupon domain according to claim 5, characterized in that: The entities in the knowledge graph corresponding to the current e-commerce platform user in S1 include: user, product, coupon; the click relationships include: use, purchase at original price, purchase at discount.
7. The recommendation method based on a knowledge graph in the e-commerce coupon domain according to claim 6, characterized in that: The TransE model is trained using the cross-entropy loss function.