Method, computer program product and system for interpreting content similarity recommendation results

By generating network diagrams to display content similarity relationships, the problem of lack of transparency of intelligent recommendation algorithms is solved, and intuitive explanation of recommendation results is provided, user trust and experience is improved, and information homogeneity is avoided.

CN118626736BActive Publication Date: 2024-11-26GUANGDONG JINHONG DIGITAL TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202411081799.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-08-08
Publication Date
2024-11-26
Estimated Expiration
2044-08-08

AI Technical Summary

Technical Problem

The existing intelligent recommendation algorithm lacks transparency and interpretation, which leads to a decrease in user trust in recommendation results and makes it difficult to understand recommendation logic, resulting in homogeneity of information push and user disgust.

Method used

The network diagram is generated through the content similarity matrix, and the node position is optimized using force-oriented layout algorithms to display the content similarity relationship. Visual encoding is used to adjust the size and thickness of nodes and edges to provide intuitive explanations of recommendation results.

Benefits of technology

It enhances users' understanding of recommendation logic, improves the trust of recommendation results, avoids the homogeneity of information push, provides clear explanations, and improves user experience.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118626736B_ABST
    Figure CN118626736B_ABST
Patent Text Reader

Abstract

The present invention relates to the technical field of similarity recommendation, and in particular to a method for interpreting content similarity recommendation results, a computer program product, and a recommendation system. The method for interpreting content similarity recommendation results calculates the content similarities of multiple projects related to the recommended project based on text vectors, establishes similarity relationships between the projects based on the content similarities, and visualizes the similarity relationships of the projects to obtain a network diagram, wherein the network diagram uses content entities as nodes, and connects two nodes corresponding to two content entities whose content similarities reach a preset degree to form an edge; a force-directed layout algorithm is used to perform spatial optimization layout on the obtained network diagram, so that nodes with high similarity are close to each other and nodes with low similarity are far away from each other, and complex similarity data is converted into intuitive graphic information, which enables users to intuitively understand the relationship between different projects and enhances users' understanding of the recommendation logic interpretation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of similarity recommendation, and in particular to a method for interpreting content similarity recommendation results, a computer program product, and a recommendation system. Background Art

[0002] With the promotion and application of digital intelligent technologies such as big data, artificial intelligence, cloud computing and blockchain, intelligent recommendation technology has been widely used by enterprises. Enterprises can effectively capture, mine and analyze massive amounts of online user data through intelligent algorithms, accurately grasp information such as users' interests, hobbies, behavioral habits and the environment they live in, thereby helping enterprises provide users with more accurate recommendations and personalized services, and improve user satisfaction and usage experience.

[0003] However, with the continuous improvement of the intelligence level of algorithms, in order to achieve a higher recommendation matching degree, the recommendation results obtained by the algorithm gradually form an information cocoon, causing users to clearly feel the homogeneity of information push while enjoying the intelligent recommendation service. Moreover, the existing recommendation algorithms usually lack transparency and explanation. Users may not be clear about the principles and basis of the recommendation algorithms, or why certain content is recommended to them. Users of intelligent recommendation services lack sufficient understanding of the existence and use of their algorithms, and find it difficult to accurately judge the size of algorithm risks, and be vigilant about possible algorithm risks. This leads to a decrease in user trust in the recommendation results, and may even cause users to dislike intelligent recommendations.

[0004] Therefore, in the process of enterprises providing recommendation system services to users, the transparency of recommendation logic is particularly important. In order to increase users' trust in the recommended content, the system needs to display the logical basis of the recommendation in detail on the recommendation interface. Specifically, when users browse the recommended content, the system should be able to provide a clear explanation of why a certain item is recommended to the user. For example, if the user has previously viewed a specific course A, the system can clearly point out when recommending another course B that this is because course B has a high degree of similarity with course A in content, theme or style. Although this approach can enhance user trust, the explanatory information that needs to be presented to users is too complicated. Therefore, how to provide intuitive and easy-to-understand explanations of recommendation results without affecting the user experience is an urgent problem that needs to be solved. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method for interpreting content similarity recommendation results, a computer program product storing a computer program that implements the steps of the method when executed, and a recommendation system. The method for interpreting content similarity recommendation results can provide users with intuitive and easy-to-understand explanations of recommendation results when making personalized recommendations to users based on content similarity.

[0006] In order to solve the above technical problem, in a first aspect, the present invention provides a method for interpreting content similarity recommendation results, comprising the following steps:

[0007] A similarity relationship acquisition step, calculating the content similarity of multiple items related to the recommended item based on the text vector, and establishing similarity relationships between the items according to the content similarity; the similarity relationships include content entities and content similarity between content entities, and a similarity matrix is ​​generated by the set of similarity relationships of all items;

[0008] In the recommendation result interpretation step, for the generated similarity matrix, a graph generation algorithm is called to construct a network graph, wherein the network graph uses content entities as nodes, and connects two nodes corresponding to two content entities whose content similarity reaches a preset degree to form an edge; the constructed network graph is visually encoded and adjusted: the size of the node represents the type of project content; the thickness of the edge represents the content similarity of the two connected nodes, and the higher the content similarity, the thicker the edge; a force-directed layout algorithm is used to perform spatial optimization layout of the network graph, so that nodes with high similarity are close to each other, and nodes with low similarity are far away from each other.

[0009] Furthermore, the steps for implementing the space optimization layout are as follows:

[0010] For all nodes in the network graph, content similarity is used as the initial distance between two connected nodes;

[0011] For each pair of connected nodes, the attraction between the two nodes is calculated based on the content similarity of the two nodes. The higher the content similarity, the greater the attraction. The repulsion between the two nodes is calculated based on the degree of connection between the types of the two nodes. The smaller the degree of connection between the types, the greater the repulsion.

[0012] Calculate the difference between the attraction and repulsion. If the attraction is greater than the repulsion, subtract the difference from the initial distance to get the new distance between the connected nodes. If the attraction is less than the repulsion, add the difference to the initial distance to get the new distance between the connected nodes.

[0013] Furthermore, the degree of connectivity of the type refers to: searching in the network graph for a node link with the two nodes as the starting point and the end point and the number of connected nodes being within a preset value, and taking the number of nodes of the same type as the two nodes in all the node links searched as the degree of connectivity of the type.

[0014] Furthermore, it also includes interactive supplementation performed after the spatial optimization layout: according to the user's zoom-in / zoom-out operation on any node and edge, a synchronous proportional zoom-in / zoom-out process is performed on all nodes and edges of the network diagram.

[0015] Furthermore, in the similarity relationship acquisition step, content similarity refers to: a value calculated based on one or more of the following similarity calculation methods for the content entity of the project obtained by processing the project data set through a text analysis tool: word vector distance, number of characters or distance, probability statistics, and word embedding model.

[0016] Furthermore, in the similarity relationship acquisition step, keywords and weights of the projects are mined and extracted through a pre-trained content analysis model; the keywords are processed in a high-dimensional manner to obtain feature vectors of the two projects respectively, and the content similarity of the projects is calculated based on the feature vectors.

[0017] Furthermore, the calculating the content similarity of the items based on the feature vectors refers to: calculating the cosine value between the feature vectors of any two items in a high-dimensional space, and using the calculated cosine value as the content similarity of the two items.

[0018] Furthermore, the project is an online course.

[0019] In a second aspect, a computer program product is also provided, which stores a computer program, and when the computer program is executed by a processor, the steps of the above-mentioned method for interpreting content similarity recommendation results can be implemented.

[0020] In a third aspect, a recommendation system is also provided, comprising a processor and a client and a server respectively connected to the processor, and also comprising the above-mentioned computer program product, wherein the computer program on the computer program product can be executed by the processor.

[0021] The recommendation system of the present invention executes the steps of the method for interpreting content similarity recommendation results, confirms recommendation results based on project content similarity, uses a network diagram to display the similarity relationship between contents, and converts complex similarity data into intuitive graphic information, which enables users to intuitively understand the relationship between different projects. A force-directed layout algorithm is used to perform spatial optimization layout of the network diagram, so that nodes with high similarity are close to each other and nodes with low similarity are far away from each other, thereby enhancing the user's understanding of the logical interpretation of the recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the drawings required for describing the embodiments are briefly introduced below.

[0023] Figure 1 It is a flowchart of the steps of the method for explaining the content similarity recommendation results. DETAILED DESCRIPTION

[0024] The present invention is further described in detail below in conjunction with specific implementation methods.

[0025] The recommendation system of this embodiment includes a processor and a client and a server connected to the processor respectively. The recommendation result object of the recommendation system is an online course. The server stores the content of all online courses. The user's client stores a computer program. The computer program is executed by the processor to achieve the following. Figure 1 The invention provides a method for explaining the content similarity recommendation results, thereby providing users with an intuitive and easy-to-understand explanation of the recommendation logic when completing personalized recommendations for users based on content similarity. The specific description is as follows.

[0026] The similarity relationship acquisition step calculates the content similarity of multiple items related to the recommended item based on the text vector, and establishes similarity relationships between each item based on the content similarity. The similarity relationship includes content entities and content similarity between content entities, and a similarity matrix is ​​generated by the collection of similarity relationships of all items.

[0027] Among them, content similarity refers to the value calculated based on one or more of the following similarity calculation methods for the content entity of the project obtained by processing the project data set through a text analysis tool: word vector distance, character number or distance, probability statistics, and word embedding model.

[0028] For example, text analysis technology is used to segment courses and obtain a basic word list. According to the word frequency statistical model, the number of times each word appears in the course is calculated to obtain the frequency of each word. Natural language processing technology is used to analyze the part of speech of each word and filter out content words such as nouns, verbs, and adjectives, which often carry more information. The structural information of the course, such as titles and paragraphs, is obtained to determine the position of the words and give higher weights to the words located in the title, the beginning of the paragraph, and the end of the paragraph. Through dependency syntactic analysis, the dependency relationship between words is identified, the collocation and relationship between words are determined, and the weight of phrases with high correlation is increased. Combined with the domain knowledge base, the professional terms and domain-related vocabulary that appear in the course are annotated and additional weights are given to these words. Through the topic model technology, the subject distribution of the course is analyzed to identify keywords that are closely related to the subject, and these keywords have higher weights. The distribution of words is analyzed. For words that are evenly distributed in the course, they are considered to be keywords of the course and their weights are increased accordingly. By using semantic analysis tools and considering synonyms and hyponyms of words, we can further adjust and optimize the weights of keywords, and finally form a list of representative keywords and their weights.

[0029] Specifically: the keywords and their weights of the projects are mined and extracted through a pre-trained content analysis model, providing basic data for calculating the similarity between the contents; the courses are converted into keyword vectors, in which the weight of each keyword reflects its importance in the course, thereby ensuring the accuracy and efficiency of subsequent similarity calculations; the keywords are processed in a high-dimensional manner to obtain the feature vectors of the two projects respectively, and the content similarity of the projects is calculated based on the feature vectors. For example, the cosine similarity calculation is used to calculate the cosine value between the feature vectors of any two projects in a high-dimensional space, and the calculated cosine value is used as the content similarity of the two projects.

[0030] The implementation process of cosine similarity calculation is as follows: natural language processing technology is used to preprocess the course content, and the basic vocabulary units of the course are obtained by word segmentation and removal of stop words. The purpose of this step is to reduce the noise of subsequent processing and improve the accuracy of analysis. After obtaining the basic vocabulary of each course, the TF-IDF algorithm is used to calculate the weight of each word to obtain the initial feature vector of each course. TF-IDF can effectively reflect the importance of words in the text, thus providing a basis for subsequent similarity calculation. If the dimensions of the obtained vectors are inconsistent, the dimensions of the feature vectors of all courses are ensured to be unified through vector filling or clipping operations. Unified dimensions are the prerequisite for effective vector operations. Vector normalization is applied to standardize the length of the feature vectors of all courses to 1. Through this step, the influence of course length on similarity calculation is eliminated, making the results more fair and comparable. The cosine similarity calculation of the normalized feature vectors in high-dimensional space is performed. This calculation determines the similarity between courses based on the angle between vectors. Set the similarity threshold according to business needs, for example, set the threshold to 8. This threshold is used as a standard to judge whether the courses are similar. Courses above the threshold are considered similar, and courses below the threshold are considered dissimilar. If the cosine similarity between course A and course B exceeds the set threshold, the two courses are judged to have similar content. This judgment helps the system quickly filter out courses with related content. The above similar courses are marked and classified for subsequent use and retrieval by users or systems to ensure the availability and accessibility of data. The TF-IDF model and vocabulary are updated regularly to adapt to new course content and changing language usage habits. During the update process, the feature vectors and similarities are recalculated to ensure the accuracy and timeliness of the system.

[0031] Exemplarily, in the preprocessing stage, the jieba word segmentation tool can be used to segment the courses, and at the same time, the stop word list is used to remove meaningless words such as "de" (的) and "shi" (是). For example, after word segmentation and stop word removal, "I love China" can be transformed into meaningful word units such as "I", "love", and "China". Next, the TF-IDF algorithm is used to calculate the weight of each word, where TF represents the term frequency, that is, the frequency of a certain word appearing in the course; IDF represents the inverse document frequency, which is used to measure the importance of a certain word. The weight of each word can be calculated through the formula tfidf = tf × log(N / df), where N represents the total number of courses in the corpus, and df represents the number of courses containing this word. For the obtained feature vectors, they can be filled with 0 or truncated to unify the dimensions of all vectors to 1000 dimensions. Then, the Euclidean distance formula is used to normalize the vectors so that their length is 1. When calculating the similarity, the cosine similarity formula similarity = A·B / (|A|×|B|) can be used, where A and B represent the feature vectors of two courses. The similarity threshold is set to 8, that is, when the similarity of two courses is greater than or equal to 8, it is considered that their contents are similar, and they are marked and classified. To adapt to the changes in the language, the TF-IDF model and the word library are updated every 1 month using a new corpus, and the feature vectors and similarities are recalculated to ensure the accuracy and timeliness of the system.

[0032] Recommendation result explanation step, construct a network graph based on the generated similarity matrix to achieve visual display. The specific process is described as follows.

[0033] (1) Call the graph generation algorithm to construct a network graph. The network graph uses content entities as nodes, and connects the two nodes corresponding to two content entities with a content similarity reaching a preset degree to form an edge. The network graph converts complex similarity data into intuitive graphical information, enabling users to intuitively understand whether there is a relationship between different courses and enhancing the user experience.

[0034] (2) Adjust the visual encoding of the constructed network graph. The type of project content is characterized by the node size; the content similarity between the two connected nodes is characterized by the thickness of the edge, and the higher the content similarity, the thicker the edge; the force-directed layout algorithm is used to optimize the spatial layout of the network graph, so that nodes with high similarity are close to each other, and nodes with low similarity are far from each other.

[0035] The implementation steps of the spatial optimization layout are as follows:

[0036] For all nodes of the network graph, the content similarity is used as the initial distance between the two connected nodes;

[0037] For each pair of connected nodes, the attraction between the two nodes is calculated based on the content similarity of the two nodes. The higher the content similarity, the greater the attraction. The repulsion between the two nodes is calculated based on the degree of connection between the types of the two nodes. The smaller the degree of connection between the types, the greater the repulsion.

[0038] The type of a node is the type of project content corresponding to the node, including video, voice, text, case, question and answer, etc. The degree of connection of types means: searching for node links in the network diagram that start and end with the two nodes and whose number of connected nodes is within a preset value, and taking the number of nodes of the same type as the two nodes in all the node links found as the degree of connection of the types.

[0039] Calculate the difference between the attraction and repulsion. If the attraction is greater than the repulsion, subtract the difference from the initial distance to get the new distance between the connected nodes. If the attraction is less than the repulsion, add the difference to the initial distance to get the new distance between the connected nodes.

[0040] By analyzing the degree of connection between the types of connected nodes and adjusting the distance between connected nodes based on the similarity of the contents corresponding to each connected node, the similarity relationship between nodes is made more intuitive, further optimizing the effectiveness of the visual information of the network diagram.

[0041] (3) Perform interactive supplementation. According to the user's zoom-in / zoom-out operation on any node or edge, all nodes and edges of the network diagram are synchronously zoomed in / out in proportion. In addition to zooming in / out in proportion, interactive graphic processing such as translating nodes and edges of the network diagram and clicking on nodes to display detailed information can also be realized.

[0042] (4) User data is encrypted through a privacy protection algorithm to ensure that user privacy information is not leaked during the analysis and recommendation process.

[0043] This embodiment implements the above-mentioned method for interpreting content similarity recommendation results through a computer program, and the computer program is stored in a computer program product, and is executed by a computer processor to implement the above-mentioned method for interpreting content similarity recommendation results. The recommendation system embodiment described above is only schematic, wherein the modules described as separate components may or may not be physically separated, and the components displayed as modules may or may not be physical modules, that is, they may be located in one place, or they may be distributed on multiple network modules. Some or all of the modules may be selected according to actual needs to achieve the purpose of the present embodiment. Ordinary technicians in this field can understand and implement it without creative labor.

[0044] Finally, it should be noted that the method for interpreting content similarity recommendation results disclosed in the embodiment of the present invention discloses only a preferred embodiment of the present invention, which is only used to illustrate the technical solution of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments can still be modified, or some of the technical features therein can be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. The explanation method of content similarity recommendation results is characterized by: The following steps are involved: A similarity relationship acquisition step, calculating content similarities between multiple items related to the recommended item based on the text vector, and establishing similarity relationships between the items according to the content similarities; The similarity relationship includes content entities and content similarity between content entities, and a similarity matrix is ​​generated by a collection of similarity relationships of all items; The recommendation result interpretation step is to call a graph generation algorithm to construct a network graph for the generated similarity matrix, wherein the network graph uses content entities as nodes and connects two nodes corresponding to two content entities whose content similarity reaches a preset degree to form an edge; The constructed network graph is visually encoded and adjusted: the size of the node represents the type of project content; the thickness of the edge represents the content similarity of the two connected nodes, and the higher the content similarity, the thicker the edge; the force-directed layout algorithm is used to optimize the spatial layout of the network graph, so that nodes with high similarity are close to each other, and nodes with low similarity are far away from each other; The steps for implementing the space optimization layout are as follows: For all nodes in the network graph, content similarity is used as the initial distance between two connected nodes; For each pair of connected nodes, the attraction between the two nodes is calculated based on the content similarity of the two nodes. The higher the content similarity, the greater the attraction. The repulsion between the two nodes is calculated based on the degree of connection between the types of the two nodes. The smaller the degree of connection between the types, the greater the repulsion. Calculate the difference between the attraction and repulsion. If the attraction is greater than the repulsion, subtract the difference from the initial distance to get the new distance between the connected nodes. If the attraction is less than the repulsion, add the difference to the initial distance to get the new distance between the connected nodes. The degree of connectivity of the type refers to: searching in the network graph for a node link with the two nodes as the starting and ending points and the number of connected nodes being within a preset value, and taking the number of nodes of the same type as the two nodes in all the node links searched as the degree of connectivity of the type.

2. The method for interpreting content similarity recommendation results as claimed in claim 1, characterized in that: It also includes interactive supplements performed after the spatial optimization layout: according to the user's zoom-in / zoom-out operation on any node and edge, all nodes and edges of the network graph are synchronously zoomed in / out in equal proportions.

3. The method for interpreting content similarity recommendation results according to claim 1, characterized in that: In the similarity relationship acquisition step, content similarity refers to: a value calculated based on one or more of the following similarity calculation methods for the content entity of the project obtained by processing the project data set through a text analysis tool: word vector distance, character number or distance, probability statistics, and word embedding model.

4. The method for interpreting content similarity recommendation results according to claim 1, characterized in that: In the similarity relationship acquisition step, keywords and their weights of the items are mined and extracted through a pre-trained content analysis model; the keywords are processed into high dimensions to obtain feature vectors of the two items respectively, and the content similarity of the items is calculated based on the feature vectors.

5. The method for interpreting content similarity recommendation results as claimed in claim 4, characterized in that: The calculating the content similarity of the items based on the feature vectors refers to: calculating the cosine value between the feature vectors of any two items in a high-dimensional space, and taking the calculated cosine value as the content similarity of the two items.

6. The method for interpreting content similarity recommendation results according to claim 1, characterized in that: The program described is an online course.

7. A computer program product storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method for interpreting content similarity recommendation results according to any one of claims 1 to 6 can be implemented.

8. Recommendation system, characterized by, The method comprises a processor and a client and a server respectively connected to the processor, and also comprises a computer program product as claimed in claim 7, wherein the computer program on the computer program product can be executed by the processor.