Progressive learning resource recommendation method and device, storage medium and computer device

Abstract

The application discloses a progressive learning resource recommendation method and device, a storage medium and computer equipment, relates to the field of information technology, and mainly aims at improving the recommendation efficiency and accuracy of learning resources. The method comprises the following steps: obtaining learning resources and user input search information; clustering the learning resources to obtain learning resources under different clustering topics; determining topic characteristic words corresponding to different clustering topics based on the learning resources under different clustering topics; determining a first target clustering topic recommended to the user in different clustering topics based on the search information and the topic characteristic words, and responding to a second target clustering topic selected by the user in the first target clustering topic; determining a topic heat map, a topic distribution map and a feature co-occurrence graph corresponding to the second target clustering topic; receiving target topic characteristic words selected by the user for the three kinds of maps, and recommending corresponding learning resources to the user according to the target topic characteristic words.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of information technology, and in particular to a progressive learning resource recommendation method and device, a storage medium and computer equipment. BACKGROUND

[0002] Learning resource recommendation is based on user needs, and through resource recommendation to achieve the purpose of assisting users in learning and reading.

[0003] At present, related resources are usually pushed by constructing user portraits. However, the construction of user portraits is a process of continuously mining user information, and the user information needs to be continuously updated, resulting in low efficiency of learning resource recommendation. At the same time, if the information for constructing the user portrait is less, or there is no user behavior information, the learning resource recommendation for the user cannot be accurately performed. SUMMARY

[0004] The present application provides a progressive learning resource recommendation method, device, storage medium and computer equipment, which can improve the recommendation efficiency and accuracy of learning resources.

[0005] According to a first aspect of the present application, a progressive learning resource recommendation method is provided, comprising:

[0006] In response to a search signal of a target user, the search information input by the target user is obtained, and a plurality of learning resources are obtained;

[0007] Based on the resource feature words in different learning resources, the different learning resources are clustered to obtain learning resources under different clustering topics;

[0008] Based on the learning resources under the different clustering topics, the topic feature words corresponding to the different clustering topics are determined;

[0009] Based on the search information and the topic feature words, a first target clustering topic recommended to the target user is determined in the different clustering topics, and a second target clustering topic selected by the target user in the first target clustering topic is responded to;

[0010] The topic heat map, topic distribution map and feature co-occurrence graph corresponding to the second target clustering topic are determined, and the topic description information is generated according to the topic feature words corresponding to the second target clustering topic, the topic heat map, the topic distribution map and the feature co-occurrence graph, and the topic description information is sent to the target user end;

[0011] receive a target topic feature word selected by the target user terminal from the topic feature words corresponding to the second target clustering topic, and recommend corresponding learning resources to the target user according to the target topic feature word.

[0012] According to a second aspect of the present application, there is provided a progressive learning resource recommendation device, comprising:

[0013] an acquisition unit configured to acquire search information input by a target user and a plurality of learning resources in response to a search signal of the target user;

[0014] a clustering unit configured to cluster the plurality of learning resources based on resource feature words in the plurality of learning resources to obtain learning resources under different clustering topics;

[0015] a feature word determination unit configured to determine topic feature words corresponding to the different clustering topics based on the learning resources under the different clustering topics;

[0016] a topic determination unit configured to determine a first target clustering topic to be recommended to the target user from the different clustering topics based on the search information and the topic feature words, and to determine a second target clustering topic selected by the target user in the first target clustering topic;

[0017] a sending unit configured to determine a topic heat map, a topic distribution map and a feature co-occurrence graph corresponding to the second target clustering topic, to generate topic description information according to the topic feature words corresponding to the second target clustering topic, the topic heat map, the topic distribution map and the feature co-occurrence graph, and to send the topic description information to a target user terminal;

[0018] a recommendation unit configured to receive a target topic feature word selected by the target user terminal from the topic feature words corresponding to the second target clustering topic, and to recommend corresponding learning resources to the target user according to the target topic feature word.

[0019] According to a third aspect of the present application, there is provided a computer readable storage medium having a computer program stored thereon, the program being executed by a processor to implement the above progressive learning resource recommendation method.

[0020] According to a fourth aspect of the present application, there is provided a computer device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the progressive learning resource recommendation method when executing the program.

[0021] According to the present invention, a progressive learning resource recommendation method, apparatus, storage medium, and computer device, compared with the current method of recommending relevant resources by building user profiles, the present invention obtains the retrieval information input by the target user in response to the retrieval signal of the target user, and obtains multiple learning resources; and clusters the different learning resources based on resource feature words in different learning resources to obtain learning resources under different cluster themes; simultaneously, based on the learning resources under the different cluster themes, it determines the theme feature words corresponding to the different cluster themes; and based on the retrieval information and the theme feature words, it determines the first recommended resource to the target user in the different cluster themes. The system first identifies a target clustering topic and responds to the target user's selection of a second target clustering topic within the first target clustering topic. Then, it determines the topic heatmap, topic distribution map, and feature co-occurrence map corresponding to the second target clustering topic. Based on the topic feature words corresponding to the second target clustering topic, the topic heatmap, the topic distribution map, and the feature co-occurrence map, it generates topic description information and sends the topic description information to the target user's terminal. Finally, it receives the target topic feature words selected by the target user from the topic feature words corresponding to the second target clustering topic based on the topic description information, and recommends corresponding learning resources to the target user based on the target topic feature words. Therefore, by clustering different learning resources, learning resources under different cluster themes are obtained, and the theme keywords corresponding to different cluster themes are determined. Based on the user's input search information and theme keywords, a first target cluster theme is determined among the different cluster themes, and the first target cluster theme is recommended to the user. The user will then select a second target cluster theme from the first target cluster theme. Subsequently, the user is shown a second target cluster theme's theme heatmap, theme distribution map, and feature co-occurrence map. Based on the theme heatmap, theme distribution map, and feature co-occurrence map, the user will select target theme keywords from the theme keywords corresponding to the second target cluster theme, and finally... Based on the target topic's characteristic keywords, this invention presents recommended learning resources to users. During the learning resource recommendation process, it guides users step-by-step to select fine-grained learning resources of interest or need, based on their search information and a comprehensive understanding of the data. This avoids users passively receiving learning materials without understanding the data's knowledge structure, thereby improving the accuracy of learning resource recommendations and enhancing the user experience. Furthermore, this invention does not require prior acquisition of user characteristic information when recommending learning resources, thus improving recommendation efficiency. Moreover, this invention can accurately recommend learning resources to users even when there is no historical user behavior information or significant gaps in user information. Attached Figure Description

[0022] The accompanying drawings, which are included to provide a further understanding of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings:

[0023] Figure 1 A flow chart of a progressive learning resource recommendation method provided by an embodiment of the application is shown in the figure;

[0024] Figure 2 A subject heat map provided by an embodiment of the application is shown in the figure;

[0025] Figure 3 A subject distribution map provided by an embodiment of the application is shown in the figure;

[0026] Figure 4 A feature co-occurrence graph provided by an embodiment of the application is shown in the figure;

[0027] Figure 5 Another flow chart of a progressive learning resource recommendation method provided by an embodiment of the application is shown in the figure;

[0028] Figure 6 A structural schematic diagram of a progressive learning resource recommendation device provided by an embodiment of the application is shown in the figure;

[0029] Figure 7 A structural schematic diagram of another progressive learning resource recommendation device provided by an embodiment of the application is shown in the figure;

[0030] Figure 8 An entity structural schematic diagram of a computer device provided by an embodiment of the application is shown in the figure. DETAILED DESCRIPTION

[0031] The application will be described in detail below with reference to the drawings and in conjunction with embodiments. It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0032] At present, the way of pushing related resources by constructing a user portrait leads to low efficiency of learning resource recommendation. Meanwhile, if there is little information for constructing a user portrait or no user behavior information, the learning resource recommendation for the user cannot be accurately performed.

[0033] To solve the above problems, an embodiment of the application provides a progressive learning resource recommendation method, as shown in the figure, which comprises the following steps. Figure 1 The method comprises the following steps.

[0034] 101、In response to a search signal of a target user, search information input by the target user is acquired, and a plurality of learning resources are acquired.

[0035] The retrieval information is a retrieval requirement of the user, such as artificial intelligence, eyes, and the like, and the learning resource can be a document, an article, a novel, a thesis, and the like.

[0036] Specifically, the target user is a user who needs to obtain a learning resource, and if the target user wants to retrieve a learning resource in a certain aspect in the learning resource retrieval system, the target user needs to first input retrieval information in the learning resource retrieval system, and then the learning resource retrieval system performs progressive retrieval based on the retrieval requirement input by the user. The embodiment of the application is mainly applied to a scenario of performing progressive retrieval on a learning resource, and the execution subject of the embodiment of the application is a device or equipment capable of performing progressive retrieval on a learning resource, which can be specifically arranged on a client side or a server side.

[0037] 102. Based on the resource feature words in different learning resources, the different learning resources are clustered to obtain learning resources under different clustering topics.

[0038] The resource feature words are words related to the learning resource, including an ID, a title, an abstract, a keyword in the resource, a source (a journal name, a patent, a blog, and the like), a type, an author, a publishing institution, a publishing time, a download frequency, a citation frequency, a number of reposts, a number of likes, a topic to which the learning resource belongs, and the like. The different clustering topics are, for example, a biology clustering topic, a history clustering topic, a chemistry clustering topic, and the like.

[0039] For the embodiment of the application, in order to recommend appropriate learning resources to the user, first, different learning resources need to be clustered. Specifically, different learning resources can be clustered according to resource keywords, resource types, title information, and the like of the different learning resources, and various learning resources are divided into different clustering topics. Then, learning resources are recommended to the user according to the learning resources under different clustering topics, so that through clustering of the learning resources, the learning resources can be recommended to the user according to the retrieval information of the user.

[0040] 103. Based on the learning resources under different clustering topics, topic feature words corresponding to the different clustering topics are determined.

[0041] The topic feature words are used to describe the category features corresponding to each learning resource under different clustering categories.

[0042] For the embodiment of the application, after the learning resources under different clustering topics are determined, the topic feature words corresponding to the different clustering topics are determined according to the keywords in each learning resource under the different clustering topics. For example, each learning resource under a certain clustering topic includes "artificial intelligence", and "artificial intelligence" is a keyword in each learning resource, so "artificial intelligence" can be finally determined as the topic feature word under the clustering topic.

[0043] 104. determining a first target clustering theme recommended to the target user in different clustering themes based on the search information and the theme feature words, and responding to a second target clustering theme selected by the target user in the first target clustering theme.

[0044] wherein the number of the first target clustering theme is less than or equal to the total number of different clustering themes, and the number of the second target clustering theme is less than or equal to the number of the second target clustering theme.

[0045] For the embodiment of the present application, after determining different clustering themes and their corresponding theme feature words, the search information is matched with the theme feature words under different clustering themes, and the target theme feature words matching the search information are found in the theme feature words corresponding to different clustering themes, and finally the clustering theme corresponding to the target theme feature words is determined as the first target clustering theme, and then the first target clustering theme and its corresponding theme keywords are displayed to the user, and the user selects the required second target clustering theme in the first target clustering theme according to the displayed information, and then the system determines the theme description information for the second target clustering theme selected by the user, and then the theme description information is displayed to the user again, and the user selects the theme feature words according to the theme description information displayed by the system again, and the system finally recommends corresponding learning resources according to the theme feature words selected by the user. Thus, by guiding the user to select the fine-grained learning resources of interest or required step by step on the premise of understanding the overall data in the process of recommending learning resources according to the preset search information of the user, the defect of passively accepting learning literature without understanding the data knowledge structure is avoided, so that the embodiment of the present application can improve the recommendation efficiency of learning resources.

[0046] 105. determining the theme heat map, the theme distribution map and the feature co-occurrence graph corresponding to the second target clustering theme, and generating the theme description information according to the theme feature words, the theme heat map, the theme distribution map and the feature co-occurrence graph corresponding to the second target clustering theme, and sending the theme description information to the target user end.

[0047] 106. receiving the target theme feature words selected by the target user end in the theme feature words corresponding to the second target clustering theme according to the theme description information, and recommending corresponding learning resources to the target user according to the target theme feature words.

[0048] wherein, as shown in Figure 2 the theme heat map describes the number of learning resources contained in the second target clustering theme, and the larger the number of learning resources, the greater the corresponding theme heat; and as shown in Figure 3As shown, the subject distribution diagram includes an emerging subject diagram and a potential subject diagram. The emerging subject refers to a subject under which resources are distributed in the past 1-2 years. The potential subject refers to a subject under which resources show a continuous annual growth trend, and the journals are in a stable or upward trend. Figure 4 As shown, the feature co-occurrence graph is used to describe which potential subject feature words exist under the subject.

[0049] For example, a student wants to find some learning materials about "artificial intelligence" and "eye" science. The system uses the following process to search for resources:

[0050] 1. The student inputs the search information: "artificial intelligence" and "eye";

[0051] 2. The system recommends the subject according to the search information and the subject feature words corresponding to different cluster subjects:

[0052] The system recommendation result (the first target cluster subject) is T = ["T.490_ artificial intelligence; deep learning; convolutional neural network; glaucoma; big data; endoscopic examination of the digestive system; ethical issues; diabetic retinopathy; computer-aided diagnosis; health management", "T.4706_ diabetic retinopathy; convolutional neural network; deep learning; fundus image; transfer learning; artificial intelligence; computer-aided diagnosis; computer vision; target detection; deep features", "T.4485_ convolutional neural network; dermoscopy image; skin disease diagnosis; deep learning; dense convolutional network; image processing; artificial intelligence; auxiliary diagnosis; pigmented skin disease; neural network (computer)",...];

[0053] 3. Student pre-selection: The student selects T.490 and T.4706 as the initial subject set (the second target cluster subject);

[0054] 4. System analysis: Based on the two subjects selected by the student, the system analyzes the subject popularity, subject literature distribution, and subject feature co-occurrence graph;

[0055] 5. Student data set selection: According to the system analysis results, the student selects the potential sub-subject T_new_ artificial intelligence; diabetic retinopathy in the initial subject set;

[0056] 6. System multi-dimensional analysis and display of subject data set: For the potential sub-subject T_new selected by the student, the system analyzes the literature corresponding to T_new from the dimensions of discipline distribution, author size, institution distribution, journal distribution, and funding distribution, and displays the data information in different sorting ways, including: by literature type, by publication time, by download frequency, by author's G-index, and by journal impact factor.

[0057] For the embodiment of the present application, after the second target clustering theme is determined, the theme heat map, the theme distribution map and the feature co-occurrence graph corresponding to the second target clustering theme are drawn according to the resource information and the theme feature words of each learning resource contained in the second target clustering theme, then the theme description information corresponding to the second target clustering theme is generated according to the theme heat map, the theme distribution map, the feature co-occurrence graph and the theme feature words, and the theme description information is displayed to the user, the user selects the theme feature words needed to be searched according to the theme description information, and finally the system searches the final learning resource according to the theme feature words selected by the user, thereby in the process of learning resource recommendation, the embodiment of the present application guides the user to select the fine-grained learning resource interested or needed step by step on the premise of understanding the overall data, avoids the user to passively accept the learning literature on the premise of not understanding the data knowledge structure, thereby improves the recommendation accuracy of the learning resource, improves the user experience, meanwhile, the present application does not need to acquire the feature information of the user in advance when recommending the learning resource, thereby can improve the recommendation efficiency of the learning resource, and the present application can accurately recommend the learning resource to the user in the case that the user historical behavior information is not available or the user information is seriously missing.

[0058] According to the progressive learning resource recommendation method provided by the present application, compared with the current method of pushing related resources by constructing a user portrait, the present application acquires the search information input by the target user and acquires a plurality of learning resources in response to the search signal of the target user; and based on the resource feature words in different learning resources, the different learning resources are clustered to obtain learning resources under different clustering topics; at the same time, based on the learning resources under the different clustering topics, the topic feature words corresponding to the different clustering topics are determined; and based on the search information and the topic feature words, a first target clustering topic recommended to the target user is determined in the different clustering topics, and a second target clustering topic selected by the target user in the first target clustering topic is responded to; then the topic heat map, the topic distribution map and the feature co-occurrence graph corresponding to the second target clustering topic are determined, and the topic description information is generated according to the topic feature words corresponding to the second target clustering topic, the topic heat map, the topic distribution map and the feature co-occurrence graph, and the topic description information is sent to the target user end; finally, the target user end selects a target topic feature word in the topic feature words corresponding to the second target clustering topic in response to the topic description information, and according to the target topic feature word, the corresponding learning resource is recommended to the target user. Therefore, by clustering different learning resources, learning resources under different clustering topics are obtained, and topic feature words corresponding to different clustering topics are determined, a first target clustering topic is determined in different clustering topics according to the search information input by the user and the topic feature words, and the first target clustering topic is recommended to the user for the first time. The user will select a second target clustering topic in the first target clustering topic, and then the topic heat map, the topic distribution map and the feature co-occurrence graph of the second target clustering topic will be displayed to the user for the second time. The user will select a target topic feature word in the topic feature words corresponding to the second target clustering topic according to the topic heat map, the topic distribution map and the feature co-occurrence graph, and finally the recommended learning resource will be displayed to the user according to the target topic feature word. In the process of recommending learning resources, the present application guides the user to select the fine-grained learning resource of interest or need step by step on the premise of understanding the overall data, thereby avoiding the user from passively accepting learning literature without understanding the data knowledge structure, improving the recommendation accuracy of learning resources, improving the user experience, and at the same time, the present application does not need to acquire the feature information of the user in advance, thereby improving the recommendation efficiency of learning resources, and the present application can accurately recommend learning resources to the user in the case that there is no historical behavior information of the user or the user information is seriously missing.

[0059] Further, in order to better illustrate the above process of progressively recommending learning resources, as a refinement and extension of the above embodiment, the present embodiment provides another progressive learning resource recommendation method, as shown inFigure 5 The method comprises:

[0060] 201. In response to a search signal of a target user, acquiring search information input by the target user and acquiring a plurality of learning resources.

[0061] Specifically, the user inputs search information in a search box of the learning resource search system, and the learning resource search system guides the user to select fine-grained learning resources of interest or required on the premise of understanding the overall data.

[0062] 202. Splicing different resource feature words in different learning resources to obtain resource description information corresponding to different learning resources.

[0063] 203. Inputting different resource description information into a preset semantic information extraction model for semantic extraction to obtain semantic information vectors corresponding to different learning resources.

[0064] 204. Using a preset dimension reduction algorithm to perform dimension reduction processing on the semantic information vectors to obtain dimension-reduced semantic information vectors corresponding to different learning resources.

[0065] The preset semantic information extraction model can be specifically an sBert model. The sBert model comprises an attention layer and a feedforward neural network layer. The preset dimension reduction algorithm can be specifically a uniform manifold approximation and dimension reduction projection algorithm.

[0066] For the embodiment of the application, first, different resource feature words in different learning resources are subjected to data preprocessing, and the specific preprocessing includes eliminating data format errors, missing important fields of data, and low-impact data, etc. Then, the resource IDs, titles, abstracts, keywords, etc. in the preprocessed learning resources are spliced in order to become a sentence as resource description information corresponding to different learning resources, and the resource description information corresponding to different learning resources is input into the sBert model for semantic extraction. The sBert model is a siamese (twin neural network) and triplet (triplet neural network) network architecture used to generate a semantic information sentence embedding representation (semantic information vector) that can be compared using cosine similarity. The semantic sentence embedding representation means that semantically similar sentences are also close in the vector space. The sBert adds a pooling operation on the output of the Bert / RoBerta to generate a fixed-size semantic information vector. There are three pooling operation strategies: using the output of the [CLS] marker, calculating the mean of all output vectors (mean strategy), and calculating the maximum value of the output vector (maximum strategy), wherein the mean strategy is the best.

[0067] Further, the semantic information vector output by the sBert model is a 384-dimensional vector, in order to speed up the calculation speed and reduce the computing power consumption, it is also necessary to use the uniform manifold approximation and dimension reduction projection algorithm (umap algorithm, n_neighbors=x, n_components=y) to reduce the 384-dimensional vector to a 5-dimensional vector, wherein the n_neighbors and n_components parameters can be set according to the user group demand, and finally the dimension-reduced semantic information vector is used to cluster different learning resources.

[0068] 205、based on the dimension-reduced semantic information vector, clustering different learning resources to obtain learning resources under different clustering topics, wherein the different clustering topics constitute a topic set.

[0069] For the embodiment of the application, in order to progressively recommend learning resources to users, it is necessary to first cluster different learning resources, based on which, step 205 specifically includes: based on the dimension-reduced semantic information vector, calculating the distance between different learning resources; based on the distance between different learning resources, constructing a weighted distance graph, wherein the different learning resources are taken as vertices in the weighted distance graph, the connection between the different learning resources is taken as an edge in the weighted distance graph, and the distance between the different learning resources is taken as the weight of the edge in the weighted distance graph; determining the minimum spanning tree corresponding to the weighted distance graph; sorting the edges in the minimum spanning tree from small to large according to the distance, and creating new sub-clusters for each sorted edge to construct a sub-cluster hierarchy; determining the minimum sub-cluster in the new sub-cluster, and using the size of the minimum sub-cluster to compress the sub-cluster hierarchy to generate a compressed spanning tree; based on the compressed spanning tree, determining learning resources under different clustering topics.

[0070] Specifically, the different learning resources can be clustered by using a hierarchical density-based noise application spatial clustering algorithm (hdbscan algorithm, min_cluster_size=x, min_samples=y, cluster_selection_method='eom'), wherein the min_cluster_size and min_samples parameters can be set according to the user group demand, and the specific process of clustering the different learning resources by using the algorithm is as follows: the reachable distance between the different learning resources is calculated two by two, the different learning resources are taken as vertices, the connection lines between the different learning resources are taken as edges, the distance between the different learning resources is taken as the weight of the edge, a weighted distance graph is constructed, the minimum spanning tree of the weighted distance graph is calculated, the edges of the tree are sorted in an increasing order according to the distance, a new sub-cluster is created for each edge in the sorted spanning tree, a sub-cluster hierarchy is constructed, then the sub-cluster hierarchy is compressed according to the minimum sub-cluster size, the hierarchy is traversed, all edges in the minimum spanning tree are deleted in a decreasing order of edge weight, each time the edge is deleted to split the sub-cluster, and it is judged whether the learning resources in the new connected component created by the deleted edge are less than those in the minimum sub-cluster, if yes, the new connected component is declared as a false connected component, the false component is marked as noise, and the sub-cluster is adjusted; if all components generated after the sub-cluster deletes the edge are false, the sub-cluster is deleted; if the components generated after the sub-cluster deletes the edge include both false components and real components, the original sub-cluster label is retained, that is, the sub-cluster before the edge is removed is retained; if the components generated after the sub-cluster deletes the edge are all not false, a new sub-cluster label is assigned to each component, that is, the original sub-cluster is successfully split into new sub-clusters; finally, the tree is traversed in a reverse topological sorting order, it is judged whether the stability of each sub-cluster parent sub-cluster is greater than the sum of its sub-clusters, if yes, the sub-cluster is declared as a selected sub-cluster, after the root node is reached in the traversal, all selected sub-clusters are obtained, and the finally selected sub-cluster and the learning resources contained in the finally selected sub-cluster are determined as the learning resources under different clustering themes.

[0071] In still another embodiment of the present application, the different learning resources can also be clustered by using a K-means algorithm, based on which, step 205 specifically further includes: initializing the different clusters corresponding to the centroid vectors; calculating the distance between the dimensionality-reduced semantic information vectors and the different clusters corresponding to the centroid vectors, and dividing the different learning resources into the different clusters based on the distances corresponding to the different clusters; determining the updated centroid vectors corresponding to the different clusters based on the dimensionality-reduced semantic information vectors of the learning resources in the different clusters; re-dividing the learning resources into the different clusters based on the updated centroid vectors until the updated centroid vectors do not change, and determining the learning resources finally divided into the different clusters as the learning resources under the different clustering themes.

[0072] Specifically, the K clusters respectively correspond to the initial centroid vectors corresponding to the initial centroids, the semantic information vectors corresponding to the different learning resources are calculated, the distance between each semantic information vector and the K centroid vectors is calculated, and each semantic information vector is assigned to the cluster corresponding to the nearest centroid vector. Then, for each cluster, the centroid of each cluster and the corresponding centroid vector are recalculated, and different learning resources are re-divided into different clusters. In this way, the different learning resources are continuously divided until the position of the centroid does not change, i.e. the centroid vector does not change. Finally, the learning resources divided into different clusters are determined as learning resources under different clustering topics.

[0073] 206. Based on the learning resources under different clustering topics, determine the topic feature words corresponding to the different clustering topics.

[0074] For the embodiment of the application, in order to recommend learning resources to users, the topic feature words corresponding to different clustering topics also need to be determined first. Based on this, step 206 specifically includes: performing word segmentation processing on the learning resource keywords under any clustering topic in the different clustering topics to obtain each segmented word contained in the learning resources under the any clustering topic; determining the word frequency of any segmented word in the corresponding learning resource; determining the number of learning resources containing the any segmented word in the learning resources under the any clustering topic; calculating the inverse document frequency corresponding to the any segmented word according to the total number of learning resources under the any clustering topic and the number of resources; multiplying the word frequency and the inverse document frequency to obtain the weight coefficient corresponding to the any segmented word; determining a target weight coefficient greater than a preset weight threshold in the weight coefficients corresponding to each segmented word, and determining the segmented word corresponding to the target weight coefficient as the topic feature word corresponding to the any clustering topic.

[0075] wherein the preset weight coefficient is set according to actual needs. Specifically, taking a clustering topic A in different clustering topics as an example, first, the different learning resource keywords (wherein the learning resource keywords can be the entire learning resource or a key sentence in the learning resource) in A are subjected to word segmentation processing to obtain each segmented word, and then the weight coefficient corresponding to each segmented word is calculated according to the following formula:

[0076]

[0077] Wherein, y represents the weight coefficient corresponding to any word segmentation in the learning resource, TF represents the word frequency corresponding to any word segmentation, IDF represents the inverse document frequency corresponding to any word segmentation, b represents the number of times of any word segmentation appearing in the corresponding learning resource, z represents the total number of word segmentation in the corresponding learning resource, e represents the total number of learning resources under a certain clustering topic, and f represents the number of learning resources containing any word segmentation under a certain clustering topic. Thus, the weight coefficient corresponding to any word segmentation can be calculated according to the above formula, and then the word segmentation corresponding to the weight coefficient greater than the preset weight coefficient is determined in each weight coefficient, and the word segmentation is determined as the topic characteristic word corresponding to the corresponding clustering topic.

[0078] 207. Based on the search information and the topic characteristic word, a first target clustering topic recommended to the target user is determined in different clustering topics, and a second target clustering topic selected by the target user in the first target clustering topic is responded.

[0079] For the embodiment of the application, after determining the topic characteristic word corresponding to the different clustering topics, it is also necessary to determine the first target clustering topic recommended to the user in the first round according to the search information and the topic characteristic word in the different clustering topics. Based on this, step 207 specifically includes: performing full-cut processing on the topic characteristic word corresponding to the different clustering topics to obtain each topic full-cut word corresponding to the different clustering topics, and performing full-cut processing on the search information to obtain each search full-cut word corresponding to the search information, and sorting the each search full-cut word according to the length of the cut word from large to small to obtain the sorted each search full-cut word; determining the target topic full-cut word hit by the each search full-cut word in the each topic full-cut word, and determining the third target clustering topic to which the target topic full-cut word belongs; calculating the cut word weight of each target topic full-cut word under the corresponding third target clustering topic; based on the cut word weight under the third target clustering topic and the order of the sorted each search full-cut word, determining the first target clustering topic recommended to the target user in the different clustering topics. Wherein, the method for calculating the cut word weight of each target topic full-cut word under the corresponding third target clustering topic includes: determining the word segmentation length corresponding to any topic full-cut word in each target topic full-cut word, and the feature word length of the topic characteristic word to which the any topic full-cut word belongs; determining the feature word weight of the topic characteristic word to which the any topic full-cut word belongs under the corresponding third target clustering topic; multiplying the feature word length and the feature word weight to obtain the weight evaluation value corresponding to the any topic full-cut word; dividing the word segmentation length by the weight evaluation value to obtain the cut word weight of the any topic full-cut word under the corresponding third target clustering topic.

[0080] Specifically, the topic feature words corresponding to different clustering topics are processed by full cutting, for example, if two clustering topics and their corresponding topic feature words are: topic 1 artificial intelligence, deep learning; topic 2 convolutional neural network, dermatoscope image, artificial intelligence, each topic feature word in topic topic 1 is processed by full cutting, such as full cutting of "artificial intelligence", to obtain each topic full cut word: "artificial intelligence, artificial, intelligence", full cutting of "deep learning" is performed to obtain each topic full cut word: "deep learning, deep, learning", each topic feature word in topic topic 2 is also processed by full cutting in the above manner, if the search information is "artificial intelligence", the corresponding search full cut word is sorted according to the length of the cut word: "artificial intelligence, artificial, intelligence", the search full cut word "artificial intelligence" hits the target topic full cut word "artificial intelligence" in topic topic 1, and also hits the target topic full cut word "artificial intelligence" in topic topic 2, the search full cut word "artificial" hits the target topic full cut word "artificial" in topic topic 1, and also hits the target topic full cut word "artificial" in topic topic 2, the search full cut word "intelligence" hits the target topic full cut word "intelligence" in topic topic 1, and also hits the target topic full cut word "intelligence" in topic topic 2, then the cut word weight corresponding to each target topic full cut word is calculated, which can be calculated according to the following formula:

[0081]

[0082] wherein h represents the cut word weight, L j represents the length of the cut word corresponding to any topic full cut word in the target topic full cut word, L z represents the feature word length of the topic feature word to which any topic full cut word belongs, and a represents the feature word weight, wherein the feature word weight is assigned according to the position ranking of the topic feature word in the corresponding topic, starting from 1.0 and decreasing by 0.005 intervals, for example, if the topic T.4485_convolutional neural network, dermatoscope image, artificial intelligence, wherein the feature word weight corresponding to the topic feature word "convolutional neural network" is 1, the feature word weight corresponding to the topic feature word "dermatoscope image" is 0.995, and the feature word weight corresponding to the topic feature word "artificial intelligence" is 0.99, the cut word weight corresponding to each target topic full cut word can be calculated by the above formula.

[0083] Further, after determining the weight of the segmented word under the third target clustering theme, the first target clustering theme recommended to the target user is determined in different clustering themes according to the weight of the segmented word and the order of the ranked each search full segmented word. The specific method for determining the first target clustering theme is as follows: for example, if the search full segmented word is “artificial intelligence field”, the corresponding search full segmented word arranged in descending order of word length is “artificial intelligence field”, “artificial intelligence”, “intelligence”, and “field”. If the target theme full segmented word in the third target clustering theme T.50 hit by “artificial intelligence field” corresponds to the segmented word weight 1, which is abbreviated as T.50-1. Taking the above abbreviation as an example, the theme and segmented word weight result hit by “artificial intelligence field” is T50: 1.0, T291: 0.995, the theme and segmented word weight result hit by “artificial intelligence” is T50: 0.667, T241: 1.0, T291: 0.663; the theme and segmented word weight result hit by “artificial” is T50: 0.333, T143: 1.0, T241: 0.5, T291: 0.331; the theme and segmented word weight result hit by “intelligence” is T50: 0.333, T238: 1.0, T241: 0.5, T291: 0.331. Then, the maximum segmented word weight is obtained to obtain T50: 1.0, T291: 0.995, T241: 1.0, T143: 1.0, T238: 1.0. Then, the first target clustering theme is determined by taking the first n themes in the order of the ranked each search full segmented word. If n is 3, the final determined first target clustering theme is T50, T291, and T241. Then, each first target clustering theme is sent to the target user, and the target user selects at least one clustering theme as the second target clustering theme from the first target clustering themes.

[0084] In another embodiment of the present application, the first target clustering theme recommended to the target user can also be determined according to the semantic information vector. Based on this, step 207 can further include: determining the search semantic information vector corresponding to the search information, and determining the theme semantic information vector corresponding to the theme feature word; calculating the cosine similarity between the search information and the theme feature word based on the search semantic information vector and the theme semantic information vector; determining the similar theme feature word in the theme feature word according to the cosine similarity; determining the similar theme feature word weight of the similar theme feature word in the corresponding clustering theme; determining the recommended theme feature word in the similar theme feature word according to the similar theme feature word weight, and determining the clustering theme to which the recommended theme feature word belongs as the first target clustering theme recommended to the target user.

[0085] Specifically, the BERT model can be used to determine the retrieval semantic information vector corresponding to the retrieval information and the topic semantic information vector corresponding to the topic feature word, respectively. Then, the cosine similarity between the retrieval information and the topic feature word is calculated based on the retrieval semantic information vector and the topic semantic information vector. Then, the cosine similarities are sorted in descending order, and the topic feature words corresponding to the top n similarities are determined as similar topic feature words. Then, the similar topic feature word weight of the similar topic feature word in the corresponding topic is determined, wherein the similar topic feature word weight is assigned according to the position ranking of the similar topic feature word weight in the corresponding topic, starting from 1.0 and decreasing by 0.005 intervals, for example, the similar topic feature word is "dermoscopy image", the corresponding topic is T.4485_convolutional neural network; dermoscopy image; skin disease diagnosis; deep learning, and the similar topic feature word weight corresponding to "dermoscopy image" is 0.995. Thus, the similar topic feature word weight corresponding to each similar topic feature word can be calculated. Then, the similar topic feature words are sorted in descending order of weight, and the top x similar topic feature words are determined as recommended topic feature words, and the clustering topic to which the recommended topic feature words belong is determined as the first target clustering topic recommended to the target user.

[0086] 208, determine the topic heat map, topic distribution map and feature co-occurrence graph corresponding to the second target clustering topic, and generate topic description information according to the topic feature words, topic heat map, topic distribution map and feature co-occurrence graph corresponding to the second target clustering topic, and send the topic description information to the target user end.

[0087] For the embodiment of the application, when the system receives the second target clustering topic selected by the target user, the system automatically determines the topic heat map and topic distribution map of the second target topic, and also needs to determine the feature co-occurrence graph of the second target topic. The specific method for determining the feature co-occurrence graph includes: determining each topic feature word contained in the second target clustering topic, and determining the total number of learning resources containing any topic feature word in the second target clustering topic; determining the feature relationship between any two topic feature words in the topic feature words, and determining the co-occurrence frequency of the two topic feature words appearing in the same learning resource; determining the feature co-occurrence graph corresponding to the second target clustering topic based on the topic feature words, the total number of learning resources, the feature relationship and the co-occurrence frequency, wherein the topic feature words constitute the nodes in the feature co-occurrence graph, the total number of learning resources constitutes the node value of each node in the feature co-occurrence graph, the feature relationship constitutes the edge between each node in the feature co-occurrence graph, and the co-occurrence frequency is the edge value of each edge in the feature co-occurrence graph.

[0088] Specifically, each topic feature word included in the second target clustering topic is taken as a node in the feature co-occurrence graph, the total number of learning resources containing any of the topic feature words is determined in the second target clustering topic, and the total number of learning resources is taken as the node value of each node in the feature co-occurrence graph, the feature relationship between each topic feature word is taken as the edge between each node in the feature co-occurrence graph, the co-occurrence frequency of any two topic feature words appearing in the same learning resource is taken as the edge value of each edge in the feature co-occurrence graph, the feature co-occurrence graph is constructed according to the above elements, and the feature co-occurrence graph is displayed to the user. The user can select a target topic feature word for learning resource retrieval based on the potential topic feature word in the feature co-occurrence graph. Thus, by displaying the topic heat map, the topic distribution graph and the feature co-occurrence graph to the user, the user can know the overall data during the retrieval process, thereby improving the accuracy of the user's selection of the retrieval word, and further improving the retrieval accuracy of the learning resource.

[0089] Further, after drawing the feature co-occurrence graph corresponding to the second target clustering topic, in order to enable the user to easily understand the feature co-occurrence graph, the feature co-occurrence graph needs to be trimmed, and the specific trimming method includes: determining a maximum node value and a minimum node value in each node value of the feature co-occurrence graph; subtracting the node value corresponding to any node in the feature co-occurrence graph from the minimum node value to obtain a node difference value, and subtracting the maximum node value from the minimum node value to obtain a node distance value; based on the node difference value, the node distance value, a preset maximum node threshold and a preset minimum node threshold, calculating a normalized node value corresponding to the any node, and replacing the node value corresponding to the any node in the feature co-occurrence graph with the normalized node value; determining a maximum edge value and a minimum edge value in each edge value of the feature co-occurrence graph; subtracting the edge value corresponding to any edge in the feature co-occurrence graph from the minimum edge value to obtain an edge difference value, and subtracting the maximum edge value from the minimum edge value to obtain an edge distance value; based on the edge difference value, the edge distance value, a preset maximum edge threshold and a preset minimum edge threshold, calculating a normalized edge value corresponding to the any edge, and replacing the edge value corresponding to the any edge in the feature co-occurrence graph with the normalized edge value.

[0090] Specifically, the normalized node value corresponding to any node can be calculated according to the following formula:

[0091]

[0092] wherein c represents the normalized node value corresponding to any node, n represents the node value corresponding to any node, n min represents the minimum node value, n max represents the maximum node value, N max represents the preset maximum node threshold, Nmin represents a preset minimum node threshold, n-n min represents a node difference value, n max -n min represents a node distance value, so that the normalized node value corresponding to each node can be calculated according to the above formula, and then the node value corresponding to any node in the feature co-occurrence graph is replaced by the normalized node value.

[0093] Further, the normalized edge value corresponding to any edge can be calculated according to the following formula:

[0094]

[0095] wherein d represents the normalized edge value corresponding to any edge, e represents the edge value corresponding to any edge, e min represents a minimum edge value, e max represents a maximum edge value, E max represents a preset maximum edge threshold, E min represents a preset minimum edge threshold, e-e min represents an edge difference value, e max -e min represents an edge distance value, so that the normalized edge value corresponding to each edge can be calculated according to the above formula, and then the edge value corresponding to any edge in the feature co-occurrence graph is replaced by the normalized edge value.

[0096] 209. receiving the target user's target subject feature word selected from the subject feature words corresponding to the second target clustering subject based on the subject description information, and recommending corresponding learning resources to the target user according to the target subject feature word.

[0097] For the embodiment of the application, after the target user selects the target subject feature word based on the above graphs and feature words after sending the subject heat map, the subject distribution graph, the feature co-occurrence graph and the subject feature word corresponding to the second target clustering subject to the target user, the system performs learning resource retrieval based on the target subject feature word to obtain a plurality of target subject learning resources, and then needs to show the user a plurality of learning resources. Based on this, step 209 specifically includes: determining a plurality of target learning resources based on the target subject feature word; determining the display dimensions corresponding to different target learning resources, wherein the display dimensions include at least one of the time distribution dimension, the subject distribution dimension, the author distribution dimension, the publishing agency distribution dimension, the periodical distribution dimension, and the publishing cost distribution dimension; determining the display order corresponding to the different target learning resources, wherein the display order includes at least one of the publication time order, the download frequency order, the author heat order of the author, and the resource impact factor order; and displaying the different target learning resources and the display dimensions corresponding thereto according to the display order.

[0098] Wherein, the author heat refers to that the average number of references of the article of the author is not less than G, and the resource impact factor is a quantitative index of the size of the resource impact.

[0099] Specifically, at least one display dimension is determined in the time distribution dimension, the discipline distribution dimension, the author distribution dimension, the publishing institution distribution dimension, the periodical distribution dimension, and the publishing cost distribution dimension, while at least one display order is determined in the time sequence of the published article, the download frequency order, the author heat order of the author, and the resource impact factor order, and then the different target learning resources and the corresponding display dimensions are displayed according to the display order. Thus, the learning resources and the corresponding display dimensions are displayed according to different display orders, which can improve the user experience, help the user to select resources according to different dimensions and orders, and save the time of the user to select resources.

[0100] According to the progressive learning resource recommendation method provided by the application, compared with the current method of pushing related resources by constructing a user portrait, the application obtains search information input by a target user and obtains various learning resources in response to a search signal of the target user; different resource feature words in the different learning resources are spliced to obtain resource description information corresponding to the different learning resources; at the same time, different resource description information is input into a preset semantic information extraction model for semantic extraction to obtain a semantic information vector corresponding to the different learning resources; a preset dimension reduction algorithm is used to perform dimension reduction processing on the semantic information vector to obtain a dimension-reduced semantic information vector corresponding to the different learning resources; then, based on the dimension-reduced semantic information vector, the different learning resources are clustered to obtain learning resources under different clustering topics; at the same time, based on the learning resources under the different clustering topics, theme feature words corresponding to the different clustering topics are determined; based on the search information and the theme feature words, a first target clustering topic recommended to the target user is determined in the different clustering topics, and a second target clustering topic selected by the target user in the first target clustering topic is responded to; then, a theme heat map, a theme distribution map and a feature co-occurrence graph corresponding to the second target clustering topic are determined, and theme description information is generated according to the theme feature words corresponding to the second target clustering topic, the theme heat map, the theme distribution map and the feature co-occurrence graph, the theme description information is sent to a target user terminal; finally, target theme feature words selected by the target user terminal in the theme description information are received, and corresponding learning resources are recommended to the target user according to the target theme feature words.According to the semantic information vector, different learning resources are clustered to obtain learning resources under different clustering topics, the clustering accuracy of the learning resources is improved, the topic feature words corresponding to different clustering topics are determined, the first target clustering topic is determined in different clustering topics according to the search information input by the user and the topic feature words, the first target clustering topic is recommended to the user for the first time, the user selects the second target clustering topic in the first target clustering topic, then the topic heat map, the topic distribution map and the feature co-occurrence graph of the second target clustering topic are displayed to the user for the second time, the user selects the target topic feature word in the topic feature words corresponding to the second target clustering topic according to the topic heat map, the topic distribution map and the feature co-occurrence graph, and finally the recommended learning resources are displayed to the user according to the target topic feature word. In the process of recommending the learning resources, the user is guided to select the fine-grained learning resources of interest or required under the premise of understanding the overall data according to the search information of the user, so that the user can passively accept the learning literature without understanding the data knowledge structure, thereby improving the recommendation accuracy of the learning resources, improving the user experience, and meanwhile, the feature information of the user is not required in advance when the learning resources are recommended, so that the recommendation efficiency of the learning resources is improved, and the learning resources can be accurately recommended to the user in the case that the historical behavior information of the user is not available or the user information is seriously missing.

[0101] Further, as a specific implementation of Figure 1 , the embodiment of the present application provides a progressive learning resource recommendation device, as shown in Figure 6 , the device comprises an acquisition unit 31, a clustering unit 32, a feature word determination unit 33, a topic determination unit 34, a sending unit 35 and a recommendation unit 36.

[0102] The acquisition unit 31 can be used to acquire the search information input by the target user and acquire a plurality of learning resources in response to the search signal of the target user.

[0103] The clustering unit 32 can be used to cluster the different learning resources based on the resource feature words in the different learning resources to obtain learning resources under different clustering topics.

[0104] The feature word determination unit 33 can be used to determine the topic feature words corresponding to the different clustering topics based on the learning resources under the different clustering topics.

[0105] The topic determination unit 34 can be used to determine the first target clustering topic recommended to the target user in the different clustering topics based on the search information and the topic feature words, and respond to the second target clustering topic selected by the target user in the first target clustering topic.

[0106] The sending unit 35 can be configured to determine a topic heat map, a topic distribution map and a feature co-occurrence graph corresponding to the second target clustering topic, and generate topic description information according to the topic feature words corresponding to the second target clustering topic, the topic heat map, the topic distribution map and the feature co-occurrence graph, and send the topic description information to the target user terminal.

[0107] The recommendation unit 36 can be configured to receive target topic feature words selected by the target user terminal from the topic feature words corresponding to the second target clustering topic in response to the topic description information, and recommend corresponding learning resources to the target user according to the target topic feature words.

[0108] In a specific application scenario, in order to cluster different learning resources, the clustering unit 32 can be configured to perform the following steps. Figure 7 As shown in the figure, the clustering unit 32 includes a splicing module 321, a semantic extraction module 322, a dimension reduction module 323 and a clustering module 324.

[0109] The splicing module 321 can be configured to splice different resource feature words in the different learning resources to obtain resource description information corresponding to the different learning resources.

[0110] The semantic extraction module 322 can be configured to input different resource description information into a preset semantic information extraction model to perform semantic extraction, and obtain semantic information vectors corresponding to the different learning resources.

[0111] The dimension reduction module 323 can be configured to perform dimension reduction processing on the semantic information vectors by using a preset dimension reduction algorithm to obtain dimension-reduced semantic information vectors corresponding to the different learning resources.

[0112] The clustering module 324 can be configured to cluster the different learning resources based on the dimension-reduced semantic information vectors to obtain learning resources under different clustering topics, wherein the different clustering topics constitute a topic set.

[0113] In a specific application scenario, in order to cluster different learning resources, the clustering module 324 includes a calculation sub-module, a construction sub-module, a first determination sub-module, a sorting sub-module and a compression sub-module.

[0114] The calculation sub-module can be configured to calculate distances between different learning resources based on the dimension-reduced semantic information vectors.

[0115] The construction submodule can be configured to construct a weighted distance graph based on the distances between the different learning resources, wherein the different learning resources are taken as vertices in the weighted distance graph, connections between the different learning resources are taken as edges in the weighted distance graph, and the distances between the different learning resources are taken as weights of the edges in the weighted distance graph.

[0116] The first determination submodule can be configured to determine a minimum spanning tree corresponding to the weighted distance graph.

[0117] The sorting submodule can be configured to sort the edges in the minimum spanning tree from small to large according to the distances, create new sub-clusters for each sorted edge, and construct a sub-cluster hierarchy.

[0118] The compression submodule can be configured to determine a minimum sub-cluster in the new sub-cluster, perform compression processing on the sub-cluster hierarchy by using a size of the minimum sub-cluster, and generate a compressed spanning tree.

[0119] The first determination submodule can also be configured to determine learning resources under different clustering topics based on the compressed spanning tree.

[0120] In a specific application scenario, in order to cluster different learning resources, the clustering module 324 further includes an initialization submodule and a division submodule.

[0121] The initialization submodule can be configured to initialize a centroid vector corresponding to each different cluster.

[0122] The calculation submodule can also be configured to calculate distances between the dimension-reduced semantic information vectors and the centroid vectors corresponding to the different clusters, and divide the different learning resources into the different clusters based on the distances corresponding to the different clusters.

[0123] The first determination submodule can also be configured to determine updated centroid vectors corresponding to the different clusters based on the dimension-reduced semantic information vectors of the learning resources in the different clusters.

[0124] The division submodule can be configured to re-divide the learning resources into the different clusters based on the updated centroid vectors until the updated centroid vectors do not change, and determine learning resources finally divided into the different clusters as the learning resources under the different clustering topics.

[0125] In a specific application scenario, in order to determine topic feature words corresponding to different clustering topics, the feature word determination unit 33 includes a word segmentation module 331, a first determination module 332, a first calculation module 333, and a multiplication module 334.

[0126] The word segmentation module 331 can be configured to perform word segmentation on the keywords of the learning resources under any of the different clustering topics, to obtain each segmented word included in the learning resources under the any of the different clustering topics.

[0127] The first determination module 332 can be configured to determine the word frequency of any of the segmented words in the corresponding learning resources.

[0128] The determination module 333 can be further configured to determine the number of learning resources containing the any of the segmented words in the learning resources under the any of the different clustering topics.

[0129] The first calculation module 333 can be configured to calculate the inverse document frequency corresponding to the any of the segmented words according to the total number of the learning resources under the any of the different clustering topics and the number of the learning resources.

[0130] The multiplication module 334 can be configured to multiply the word frequency and the inverse document frequency to obtain the weight coefficient corresponding to the any of the segmented words.

[0131] The determination module 333 can be specifically configured to determine a target weight coefficient greater than a preset weight threshold value from the weight coefficients corresponding to the segmented words, and determine the segmented word corresponding to the target weight coefficient as the topic feature word corresponding to the any of the different clustering topics.

[0132] In a specific application scenario, in order to determine a first target clustering topic recommended to a target user from the different clustering topics, the topic determination unit 34 includes a full segmentation module 341, a second determination module 342, and a second calculation module 343.

[0133] The full segmentation module 341 can be configured to perform full segmentation on the topic feature words corresponding to the different clustering topics to obtain each topic full segmentation word corresponding to the different clustering topics, and perform full segmentation on the search information to obtain each search full segmentation word corresponding to the search information, and sort the search full segmentation words according to the length of the segmentation words from large to small to obtain the sorted search full segmentation words.

[0134] The second determination module 342 can be configured to determine a target topic full segmentation word hit by the search full segmentation words from the topic full segmentation words, and determine a third target clustering topic to which the target topic full segmentation word belongs.

[0135] The second calculation module 343 can be configured to calculate the segmentation word weight of each target topic full segmentation word under the corresponding third target clustering topic.

[0136] The second determining module 342 can be specifically configured to determine a first target clustering theme recommended to the target user in the different clustering themes based on the word segmentation weight of the segmented word under the third target clustering theme and the sequence of the sorted each search full segmented word.

[0137] In a specific application scenario, in order to calculate the segmented word weight of each target theme full segmented word under the corresponding third target clustering theme, the second calculating module 343 comprises a second determining submodule, a multiplication submodule and a division submodule.

[0138] The second determining submodule can be configured to determine the word segmentation length of an arbitrary theme full segmented word in the target theme full segmented words and the feature word length of the theme feature word to which the arbitrary theme full segmented word belongs.

[0139] The second determining submodule can be further configured to determine the feature word weight of the theme feature word to which the arbitrary theme full segmented word belongs under the corresponding third target clustering theme.

[0140] The multiplication submodule can be configured to multiply the feature word length and the feature word weight to obtain the weight evaluation value corresponding to the arbitrary theme full segmented word.

[0141] The division submodule can be configured to divide the word segmentation length by the weight evaluation value to obtain the segmented word weight of the arbitrary theme full segmented word under the corresponding third target clustering theme.

[0142] In a specific application scenario, in order to determine the first target clustering theme recommended to the target user in the different clustering themes, the second determining module 342 can be further configured to determine the search semantic information vector corresponding to the search information and determine the theme semantic information vector corresponding to the theme feature word.

[0143] The second calculating module 343 can be further configured to calculate the cosine similarity between the search information and the theme feature word based on the search semantic information vector and the theme semantic information vector.

[0144] The second determining module 342 can be specifically configured to determine a similar theme feature word in the theme feature words according to the cosine similarity.

[0145] The second determining module 342 can be specifically configured to determine the similar theme feature word weight of the similar theme feature word in the corresponding clustering theme.

[0146] The second determining module 342 can be specifically configured to determine a recommended topic feature word from the similar topic feature words according to the weight of the similar topic feature word, and determine a cluster topic to which the recommended topic feature word belongs as a first target cluster topic recommended to the target user.

[0147] In a specific application scenario, in order to determine the feature co-occurrence graph corresponding to the second target cluster topic, the sending unit 35 can be specifically configured to determine each topic feature word included in the second target cluster topic, and determine the total number of learning resources containing any topic feature word from the each topic feature word in the second target cluster topic; determine the feature relationship between any two topic feature words from the each topic feature word, and determine the co-occurrence frequency of the any two topic feature words appearing in the same learning resource; and determine the feature co-occurrence graph corresponding to the second target cluster topic based on the each topic feature word, the total number of learning resources, the feature relationship, and the co-occurrence frequency, wherein the each topic feature word constitutes each node in the feature co-occurrence graph, the total number of learning resources constitutes the node value of each node in the feature co-occurrence graph, the feature relationship constitutes the edge between each node in the feature co-occurrence graph, and the co-occurrence frequency is the edge value of each edge in the feature co-occurrence graph.

[0148] In a specific application scenario, in order to optimize the feature co-occurrence graph, the device further includes a node value determining unit 37, a subtraction unit 38, and a calculation unit 39.

[0149] The node value determining unit 37 can be configured to determine the maximum node value and the minimum node value from the each node value of the feature co-occurrence graph.

[0150] The subtraction unit 38 can be configured to subtract the node value corresponding to any node from the each node from the minimum node value to obtain a node difference value, and subtract the maximum node value from the minimum node value to obtain a node distance value.

[0151] The calculation unit 39 can be configured to calculate a normalized node value corresponding to the any node based on the node difference value, the node distance value, a preset maximum node threshold, and a preset minimum node threshold, and replace the node value corresponding to the any node in the feature co-occurrence graph with the normalized node value.

[0152] The node value determining unit 37 can be further configured to determine the maximum edge value and the minimum edge value from the each edge value of the feature co-occurrence graph.

[0153] The subtraction unit 38 can also be configured to subtract the edge value corresponding to any of the edges from the minimum edge value to obtain an edge difference value, and subtract the maximum edge value from the minimum edge value to obtain an edge distance value.

[0154] The calculation unit 39 can also be configured to calculate a normalized edge value corresponding to the any edge based on the edge difference value, the edge distance value, a preset maximum edge threshold, and a preset minimum edge threshold, and replace the edge value corresponding to the any edge in the feature co-occurrence graph with the normalized edge value.

[0155] In a specific application scenario, in order to recommend a corresponding learning resource to a user, the recommendation unit 36 includes a third determination module 361 and a display module 362.

[0156] The third determination module 361 can be configured to determine a plurality of target learning resources based on the target theme feature word.

[0157] The third determination module 361 can be specifically configured to determine a display dimension corresponding to each of the different target learning resources, wherein the display dimension includes at least one of a time distribution dimension, a subject distribution dimension, an author distribution dimension, a publishing organization distribution dimension, a periodical distribution dimension, and a publishing cost distribution dimension.

[0158] The third determination module 361 can also be specifically configured to determine a display order corresponding to each of the different target learning resources, wherein the display order includes at least one of a publication time order, a download frequency order, an author popularity order of an author belonging to the target learning resource, and a resource impact factor order.

[0159] The display module 362 can be configured to display the different target learning resources and the display dimensions corresponding thereto in the display order.

[0160] It should be noted that other corresponding descriptions of the various functional modules involved in the progressive learning resource recommendation device provided by the embodiments of the present application can be referred to the corresponding descriptions of the method shown in the above Figure 1

[0161] Based on the above description of the method shown in the above Figure 1 ​According to the method, the embodiment of the present application also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to realize the following steps: in response to a search signal of a target user, obtaining search information input by the target user and obtaining a plurality of learning resources; clustering the different learning resources based on resource feature words in the different learning resources to obtain learning resources under different clustering themes; determining theme feature words corresponding to the different clustering themes based on the learning resources under the different clustering themes; determining a first target clustering theme recommended to the target user in the different clustering themes based on the search information and the theme feature words, and in response to a second target clustering theme selected by the target user in the first target clustering theme; determining a theme heat map, a theme distribution map and a feature co-occurrence graph corresponding to the second target clustering theme, and generating theme description information according to the theme feature words corresponding to the second target clustering theme, the theme heat map, the theme distribution map and the feature co-occurrence graph, and sending the theme description information to a target user terminal; receiving a target theme feature word selected by the target user terminal in the theme description information in the theme feature words corresponding to the second target clustering theme, and recommending corresponding learning resources to the target user according to the target theme feature word.

[0162] According to the method and the device, the embodiment of the present application also provides a computer device entity structure diagram, as shown in the following table: Figure 1 Figure 6 According to the method and the device, the embodiment of the present application also provides a computer device entity structure diagram, as shown in the following table: Figure 8 ​As shown, the computer device comprises a processor 41, a memory 42, and a computer program stored on the memory 42 and executable on the processor, wherein the memory 42 and the processor 41 are both arranged on a bus 43, and the processor 41 implements the following steps when executing the program: in response to a search signal of a target user, obtaining search information input by the target user, and obtaining a plurality of learning resources; based on resource feature words in different learning resources, clustering the different learning resources to obtain learning resources under different clustering topics; based on the learning resources under the different clustering topics, determining topic feature words corresponding to the different clustering topics; based on the search information and the topic feature words, determining a first target clustering topic recommended to the target user in the different clustering topics, and in response to a second target clustering topic selected by the target user in the first target clustering topic; determining a topic heat map, a topic distribution map, and a feature co-occurrence graph corresponding to the second target clustering topic, and generating topic description information according to the topic feature words corresponding to the second target clustering topic, the topic heat map, the topic distribution map, and the feature co-occurrence graph, and sending the topic description information to a target user end; receiving a target topic feature word selected by the target user end for the topic description information in the topic feature words corresponding to the second target clustering topic, and recommending corresponding learning resources to the target user according to the target topic feature word.

[0163] By the technical scheme of the present application, the target user's search signal is responded to, the search information input by the target user is acquired, and various learning resources are acquired; and based on the resource characteristic words in different learning resources, the different learning resources are clustered to obtain learning resources under different clustering themes; at the same time, based on the learning resources under the different clustering themes, theme characteristic words corresponding to the different clustering themes are determined; and based on the search information and the theme characteristic words, a first target clustering theme recommended to the target user is determined in the different clustering themes, and a second target clustering theme selected by the target user in the first target clustering theme is responded to; then a theme heat map, a theme distribution map and a feature co-occurrence graph corresponding to the second target clustering theme are determined, and theme description information is generated according to the theme characteristic words corresponding to the second target clustering theme, the theme heat map, the theme distribution map and the feature co-occurrence graph, and the theme description information is sent to the target user end; finally, the target user end selects a target theme characteristic word in the theme characteristic words corresponding to the second target clustering theme for the theme description information, and according to the target theme characteristic word, corresponding learning resources are recommended to the target user. Thus, by clustering different learning resources, learning resources under different clustering themes are obtained, theme characteristic words corresponding to different clustering themes are determined, a first target clustering theme is determined in different clustering themes according to the search information input by the user and the theme characteristic words, and the first target clustering theme is recommended to the user for the first time, the user will select a second target clustering theme in the first target clustering theme, then the theme heat map, the theme distribution map and the feature co-occurrence graph of the second target clustering theme are displayed to the user for the second time, the user will select a target theme characteristic word in the theme characteristic words corresponding to the second target clustering theme according to the theme heat map, the theme distribution map and the feature co-occurrence graph, and finally the recommended learning resources are displayed to the user according to the target theme characteristic word. In the process of learning resource recommendation, the present application guides the user to select the fine-grained learning resources of interest or needed step by step on the premise of understanding the overall data, avoiding the user to passively accept learning literature on the premise of not understanding the data knowledge structure, thereby improving the recommendation accuracy of learning resources and enhancing the user experience. At the same time, the present application does not need to acquire the user's feature information in advance when recommending learning resources, thereby improving the recommendation efficiency of learning resources, and the present application can accurately recommend learning resources to the user even in the case of no user historical behavior information or serious user information loss.

[0164] It should be apparent to those skilled in the art that the modules or steps of the application described above can be implemented with a general purpose computing device, which can be centralized on a single computing device or distributed over a network of multiple computing devices, and optionally implemented with program code executable by a computing device, which can be stored in a storage device and executed by a computing device, and in some cases, the steps shown or described can be performed in a different order than shown, or made into individual integrated circuit modules, or multiple modules or steps made into a single integrated circuit module. Thus, the application is not limited to any particular combination of hardware and software.

[0165] The preferred embodiments of the application described above are intended to be merely illustrative, and numerous modifications and adaptations will be apparent to those skilled in the art. Such modifications and adaptations are intended to come within the scope of the present application, which is set forth in the following claims.

Claims

1. A method for recommending a learning resource in a progressive manner, the method comprising: The method comprises the following steps: In response to a search signal of a target user, obtaining search information input by the target user and obtaining a plurality of learning resources; Based on resource feature words in different learning resources, clustering the different learning resources to obtain learning resources under different clustering topics; Based on the learning resources under the different clustering topics, determining topic feature words corresponding to the different clustering topics; Based on the search information and the topic feature words, determining a first target clustering topic recommended to the target user in the different clustering topics, and in response to a second target clustering topic selected by the target user in the first target clustering topic; Determine the topic heat map, topic distribution map and feature co-occurrence graph corresponding to the second target clustering topic, and generate topic description information according to the topic feature words corresponding to the second target clustering topic, the topic heat map, the topic distribution map and the feature co-occurrence graph, and send the topic description information to the target user terminal; Receive the target user terminal for the topic description information, select the target topic feature word in the topic feature word corresponding to the second target clustering topic, and recommend the corresponding learning resource to the target user according to the target topic feature word.

2. The method of claim 1, wherein, The method comprises the following steps: Splice different resource feature words in the different learning resources to obtain resource description information corresponding to the different learning resources; Input different resource description information into a preset semantic information extraction model for semantic extraction to obtain a semantic information vector corresponding to the different learning resources; Use a preset dimension reduction algorithm to perform dimension reduction processing on the semantic information vector to obtain a dimension-reduced semantic information vector corresponding to the different learning resources; Based on the dimension-reduced semantic information vector, the different learning resources are clustered to obtain learning resources under different clustering topics, wherein the different clustering topics constitute a topic set.

3. The method of claim 2, wherein, The method comprises the following steps: Based on the dimension-reduced semantic information vector, the distance between different learning resources is calculated; Based on the distance between the different learning resources, a weighted distance graph is constructed, wherein the different learning resources are taken as vertices in the weighted distance graph, the connection between the different learning resources is taken as an edge in the weighted distance graph, and the distance between the different learning resources is taken as the weight of the edge in the weighted distance graph; Determine the minimum spanning tree corresponding to the weighted distance graph; According to the distance, the edges in the minimum spanning tree are sorted from small to large, and new sub-clusters are created for each sorted edge to construct a sub-cluster hierarchy; Determine the minimum sub-cluster in the new sub-cluster, and use the size of the minimum sub-cluster to compress the sub-cluster hierarchy to generate a compressed spanning tree; Based on the compressed spanning tree, determine the learning resources under different clustering topics.

4. The method of claim 2, wherein, The method comprises the following steps: initializing a center vector corresponding to each cluster; calculating the distance between the dimension-reduced semantic information vector and the center vector corresponding to each cluster, and dividing the different learning resources into the different clusters based on the distance corresponding to each cluster; determining an updated center vector corresponding to each cluster based on the dimension-reduced semantic information vector corresponding to the learning resources in each cluster; re-dividing the learning resources into the different clusters based on the updated center vector until the updated center vector does not change, and determining the learning resources finally divided into the different clusters as the learning resources under the different clustering topics.

5. The method of claim 1, wherein, The method comprises the following steps: performing word segmentation on the learning resource keywords under any clustering topic in the different clustering topics to obtain each segmented word contained in the learning resources under the any clustering topic; determining the word frequency of any segmented word in the corresponding learning resources; determining the number of learning resources containing the any segmented word in the learning resources under the any clustering topic; calculating the inverse document frequency corresponding to the any segmented word according to the total number of learning resources under the any clustering topic and the number of resources; multiplying the word frequency and the inverse document frequency to obtain the weight coefficient corresponding to the any segmented word; determining a target weight coefficient greater than a preset weight threshold in the weight coefficient corresponding to each segmented word, and determining the segmented word corresponding to the target weight coefficient as the topic feature word corresponding to the any clustering topic.

6. The method of claim 1, wherein, The method comprises the following steps: performing full-cut processing on the topic feature words corresponding to the different clustering topics to obtain each topic full-cut segmented word corresponding to the different clustering topics, and performing full-cut processing on the search information to obtain each search full-cut segmented word corresponding to the search information, and sorting the each search full-cut segmented word according to the length of the cut word from large to small to obtain the sorted each search full-cut segmented word; determining a target topic full-cut segmented word hit by the each search full-cut segmented word in the each topic full-cut segmented word, and determining a third target clustering topic to which the target topic full-cut segmented word belongs; calculating the cut word weight of each target topic full-cut segmented word under the corresponding third target clustering topic; determining the first target clustering topic recommended to the target user in the different clustering topics based on the cut word weight under the third target clustering topic and the order of the sorted each search full-cut segmented word.

7. The method of claim 6, wherein, The method comprises the following steps: determine a split word length corresponding to any of the target theme full split words, and a feature word length of a theme feature word to which the any of the target theme full split words belongs; determine a feature word weight of the theme feature word to which the any of the target theme full split words belongs under a third target clustering theme; multiply the feature word length and the feature word weight to obtain a weight evaluation value corresponding to the any of the target theme full split words; divide the split word length and the weight evaluation value to obtain a split word weight of the any of the target theme full split words under the third target clustering theme.

8. The method of claim 1, wherein, determining a first target clustering theme recommended to the target user in the different clustering themes based on the search information and the theme feature word, comprises: determining a search semantic information vector corresponding to the search information, and determining a theme semantic information vector corresponding to the theme feature word; calculating a cosine similarity between the search information and the theme feature word based on the search semantic information vector and the theme semantic information vector; determining a similar theme feature word in the theme feature word according to the cosine similarity; determining a similar theme feature word weight of the similar theme feature word in the corresponding clustering theme; determining a recommended theme feature word in the similar theme feature word according to the similar theme feature word weight, and determining a clustering theme to which the recommended theme feature word belongs as the first target clustering theme recommended to the target user.

9. The method of claim 1, wherein, determining a feature co-occurrence graph corresponding to the second target clustering theme, comprises: determining each theme feature word contained in the second target clustering theme, and determining a total number of learning resources containing any of the theme feature words in the second target clustering theme; determining a feature relationship between any two of the theme feature words, and determining a co-occurrence frequency of the any two of the theme feature words appearing in the same learning resource; determining a feature co-occurrence graph corresponding to the second target clustering theme based on the each theme feature word, the total number of learning resources, the feature relationship and the co-occurrence frequency, wherein the each theme feature word constitutes each node in the feature co-occurrence graph, the total number of learning resources constitutes a node value of each node in the feature co-occurrence graph, the feature relationship constitutes an edge between each node in the feature co-occurrence graph, and the co-occurrence frequency is an edge value of each edge in the feature co-occurrence graph.

10. The method of claim 9, wherein, After determining the feature co-occurrence graph corresponding to the second target clustering theme based on the each theme feature word, the total number of learning resources, the feature relationship and the co-occurrence frequency, the method further comprises: determining a maximum node value and a minimum node value in each node value of the feature co-occurrence graph; subtracting a node value corresponding to any of the nodes from the minimum node value to obtain a node difference value, and subtracting the maximum node value from the minimum node value to obtain a node distance value; Based on the node difference value, the node distance value, a preset maximum node threshold and a preset minimum node threshold, a normalized node value corresponding to the arbitrary node is calculated, and the normalized node value is used to replace the node value corresponding to the arbitrary node in the feature co-occurrence graph; A maximum edge value and a minimum edge value are determined in each edge value of the feature co-occurrence graph; An edge difference value is obtained by subtracting the minimum edge value from an arbitrary edge value in the edges, and an edge distance value is obtained by subtracting the minimum edge value from the maximum edge value; Based on the edge difference value, the edge distance value, a preset maximum edge threshold and a preset minimum edge threshold, a normalized edge value corresponding to the arbitrary edge is calculated, and the normalized edge value is used to replace the edge value corresponding to the arbitrary edge in the feature co-occurrence graph.

11. The method of claim 1, wherein, The method further comprises: Based on the target topic feature word, a plurality of target learning resources are determined; Different target learning resources correspond to different display dimensions, wherein the display dimensions include at least one of a time distribution dimension, a subject distribution dimension, an author distribution dimension, a publishing agency distribution dimension, a periodical distribution dimension, and a publishing cost distribution dimension; Different target learning resources correspond to different display orders, wherein the display orders include at least one of a publication time order, a download frequency order, an author popularity order of an author belonging to the target learning resource, and a resource impact factor order; The different target learning resources and their corresponding display dimensions are displayed in the display order.

12. A progressive learning resource recommendation apparatus, characterized by comprising: The method further comprises: An acquisition unit is configured to acquire search information input by a target user and a plurality of learning resources in response to a search signal of the target user; A clustering unit is configured to cluster the different learning resources based on resource feature words in the different learning resources to obtain learning resources under different clustering topics; A feature word determination unit is configured to determine topic feature words corresponding to the different clustering topics based on the learning resources under the different clustering topics; A topic determination unit is configured to determine a first target clustering topic recommended to the target user among the different clustering topics based on the search information and the topic feature words, and to determine a second target clustering topic selected by the target user in the first target clustering topic; A sending unit is configured to determine a topic heat map, a topic distribution graph and a feature co-occurrence graph corresponding to the second target clustering topic, to generate topic description information based on topic feature words corresponding to the second target clustering topic, the topic heat map, the topic distribution graph and the feature co-occurrence graph, and to send the topic description information to a target user terminal; A recommendation unit is configured to receive a target topic feature word selected by the target user terminal among the topic feature words corresponding to the second target clustering topic based on the topic description information, and to recommend corresponding learning resources to the target user based on the target topic feature word.

13. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by a processor, implements the steps of the method of any one of claims 1 to 11.

14. A computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, characterized in that, The computer program, which is executed by a processor, implements the steps of the method according to any one of claims 1 to 11.

Images