An information recommendation method, device, equipment and storage medium
By constructing a target knowledge graph and using a bidirectional LSTM model, the problems of data sparsity and cold start in traditional recommender systems are solved, enhancing the interpretability and diversity of recommendations and improving data coverage.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI WEIMOB ENTERPRISE DEV CO LTD
- Filing Date
- 2022-04-13
- Publication Date
- 2026-06-02
Smart Images
Figure CN114780839B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of personalized recommendation technology, and in particular to an information recommendation method, apparatus, device, and storage medium. Background Technology
[0002] Currently, with the rapid development of internet technology and applications, recommender systems have been widely used in e-commerce, movies and videos, music, and advertising. However, traditional collaborative filtering recommender systems that rely on user-material interactions to recommend materials of interest to users have three drawbacks: First, since most users do not frequently browse or click on related materials, the user-material records for most users are sparse, meaning the user-material relationship matrix is sparse; second, there is the cold start problem, meaning that for some newly registered users, who have never left any behavioral records on the platform, recommendations cannot be made directly based on traditional user similarity methods; and third, recommended items are often uninterpretable. However, knowledge graphs can describe a large number of user-material interaction relationships, such as click relationships, potentially revealing user preferences and helping to find potential connections, thus improving the accuracy of recommended items; they can also help to reasonably expand user interests, increasing the diversity of recommended materials; and they can connect users' historical browsing information with recommendation records, improving the interpretability of the recommender system. This allows knowledge graph-based recommendation systems to address the shortcomings of traditional collaborative filtering recommendation systems. However, traditional knowledge graph-based recommendation systems treat user behavior sequences as unordered sets, neglecting the temporal order issue. Furthermore, knowledge graph-based recommendation systems use the same random factor in each iteration, which affects the diversity of the recommendation system and the coverage of the data.
[0003] Therefore, how to provide a solution to the above-mentioned technical problems is a problem that needs to be solved by those skilled in the art. Summary of the Invention
[0004] In view of this, the purpose of this invention is to provide an information recommendation method, apparatus, device, and storage medium that can solve the problems of data sparsity and cold start in traditional recommendation systems, enhance the interpretability and diversity of recommendations, and consider time-series issues to improve data coverage. The specific solution is as follows:
[0005] Firstly, this application discloses an information recommendation method, including:
[0006] A target knowledge graph is constructed, and user preference materials corresponding to user historical behavior sequences determined based on the target knowledge graph are modeled to obtain corresponding preference material feature vectors; the user historical behavior sequence is the seed material sequence of the target knowledge graph corresponding to the current user;
[0007] The user's historical behavior sequence is modeled using a pre-defined bidirectional LSTM model to obtain the corresponding user behavior feature vector;
[0008] The user feature vector corresponding to the current user is determined based on the preferred material feature vector and the user behavior feature vector;
[0009] Based on the user feature vector and the material feature vector determined according to the preset materials, the click probability of the recommended material corresponding to the current user is predicted, and the corresponding material recommendation result is determined according to the click probability of the recommended material.
[0010] Optionally, constructing the target knowledge graph includes:
[0011] The data to be processed in the recommendation scenario is preprocessed to determine the corresponding graph entities and the entity relationships corresponding to the graph entities, and the corresponding target knowledge graph is constructed based on the graph entities and the entity relationships.
[0012] Optionally, the step of modeling the user preference material corresponding to the user historical behavior sequence determined based on the target knowledge graph to obtain the corresponding preference material feature vector includes:
[0013] Obtain the historical behavior of the current user within a preset time range to obtain the user's historical behavior sequence;
[0014] Based on the target knowledge graph, the preference material feature vector for each hop of the user's historical behavior sequence is determined, and the preference material feature vectors for each hop are summed to obtain the preference material feature vector corresponding to the current user.
[0015] Optionally, the step of modeling the user's historical behavior sequence using a preset bidirectional LSTM model to obtain the corresponding user behavior feature vector includes:
[0016] The user's historical behavior sequence is input into a preset data encoding layer for encoding, and the encoded user historical behavior sequence is input into a preset bidirectional LSTM model for modeling to obtain the corresponding user behavior feature vector.
[0017] Optionally, after modeling the encoded user historical behavior sequence input values using a pre-set bidirectional LSTM model to obtain the corresponding user behavior feature vector, the method further includes:
[0018] The user behavior feature vector is input into a fully connected layer constructed based on a preset convolutional neural network algorithm for dimensionality reduction, resulting in the dimensionality-reduced user behavior feature vector.
[0019] Optionally, determining the user feature vector corresponding to the current user based on the preferred material feature vector and the user behavior feature vector includes:
[0020] The user feature vector corresponding to the current user is obtained by adding the preferred material feature vector to the user behavior feature vector.
[0021] Optionally, predicting the click probability of the recommended material corresponding to the current user based on the user feature vector and the material feature vector determined according to preset materials includes:
[0022] The user feature vector is multiplied by the material feature vector determined according to the preset materials to obtain the click probability of the recommended material for the current user.
[0023] Secondly, this application discloses an information recommendation device, comprising:
[0024] The knowledge graph construction module is used to construct the target knowledge graph;
[0025] The first feature vector determination module is used to model the user preference material corresponding to the user historical behavior sequence determined based on the target knowledge graph to obtain the corresponding preference material feature vector; the user historical behavior sequence is the seed material sequence of the target knowledge graph corresponding to the current user;
[0026] The second feature vector determination module is used to model the user's historical behavior sequence using a preset bidirectional LSTM model to obtain the corresponding user behavior feature vector.
[0027] The third feature vector determination module is used to determine the user feature vector corresponding to the current user based on the preferred material feature vector and the user behavior feature vector.
[0028] The click probability prediction module is used to predict the click probability of the recommended material corresponding to the current user based on the user feature vector and the material feature vector determined according to the preset material, and to determine the corresponding material recommendation result based on the click probability of the recommended material.
[0029] Thirdly, this application discloses an electronic device, including:
[0030] Memory, used to store computer programs;
[0031] A processor is configured to execute the computer program to implement the steps of the aforementioned disclosed information recommendation method.
[0032] Fourthly, this application discloses a computer-readable storage medium for storing a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the aforementioned disclosed information recommendation method.
[0033] As can be seen, this application provides an information recommendation method, including: constructing a target knowledge graph, and modeling user preference materials corresponding to user historical behavior sequences determined based on the target knowledge graph to obtain corresponding preference material feature vectors; the user historical behavior sequence is the seed material sequence of the target knowledge graph corresponding to the current user; modeling the user historical behavior sequence using a preset bidirectional LSTM model to obtain corresponding user behavior feature vectors; determining the user feature vector corresponding to the current user based on the preference material feature vector and the user behavior feature vector; predicting the click probability of the recommended material corresponding to the current user based on the user feature vector and the material feature vector determined according to the preset material, and determining the corresponding material recommendation result based on the click probability of the recommended material. Therefore, this application determines user preference materials based on the constructed target knowledge graph, solving the problems of data sparsity and cold start in traditional recommendation systems, and enhancing the interpretability and diversity of recommendations. By using a preset bidirectional LSTM model combined with the target knowledge graph, it overcomes the problem that the target knowledge graph treats user historical behavior sequences as an unordered set, considers time series issues, and thus discovers changes in user preferences over time, thereby enhancing material coverage. Attached Figure Description
[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0035] Figure 1 This is a flowchart of an information recommendation method disclosed in this application;
[0036] Figure 2 This is a partial knowledge graph diagram disclosed in this application;
[0037] Figure 3 This is a schematic diagram of a modeling structure disclosed in this application;
[0038] Figure 4 This is a flowchart of a specific recommendation method disclosed in this application;
[0039] Figure 5 This is a schematic diagram of a specific information recommendation model disclosed in this application;
[0040] Figure 6 This is a schematic diagram of the structure of an information recommendation device disclosed in this application;
[0041] Figure 7 This is a structural diagram of an electronic device disclosed in this application. Detailed Implementation
[0042] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0043] Currently, traditional recommender systems suffer from three main drawbacks: data sparsity, the cold start problem, and the lack of interpretability. While knowledge graph-based recommender systems can address these shortcomings, they also treat user history sequences as unordered sets, neglecting time-series considerations and using the same random factor in each iteration, which negatively impacts system diversity and data coverage. Therefore, this application provides an information recommendation method that addresses the data sparsity and cold start problems of traditional recommender systems, enhances interpretability and diversity, and improves data coverage by considering time-series considerations.
[0044] This invention discloses an information recommendation method, see [link to relevant documentation]. Figure 1 As shown, the method includes:
[0045] Step S11: Construct a target knowledge graph and model the user preference materials corresponding to the user historical behavior sequence determined based on the target knowledge graph to obtain the corresponding preference material feature vector; the user historical behavior sequence is the seed material sequence of the target knowledge graph corresponding to the current user.
[0046] In this embodiment, the target knowledge graph is constructed by first preprocessing the data and then using the preprocessed data to build the target knowledge graph. Specifically, the data to be processed in the recommendation scenario is preprocessed to determine the corresponding graph entities and the entity relationships corresponding to the graph entities, and the corresponding target knowledge graph is constructed based on the graph entities and the entity relationships. Then, the user preference materials corresponding to the user's historical behavior sequence are determined according to the above target knowledge graph, and then modeling is performed to obtain the corresponding preference material feature vector. It can be understood that after determining the graph entities, entity recognition methods can be used to extract the data corresponding to the graph entities from the data to be processed. For example, if the current recommendation is a company recommendation, the data in the company resource library is preprocessed, and the company name, investor, company category, and company address are taken as graph entities, and "region", "founded", "includes", "investment relationship" are taken as entity relationships. Thus, the preprocessed data is used to construct a knowledge graph triple containing a head entity, relationship, and tail entity, that is, the knowledge graph triple is (head entity, relationship, tail entity), and the corresponding target knowledge graph G is constructed based on the above knowledge graph triple. Therefore, introducing knowledge graphs for recommendation helps to discover relationships between entities and enhances the interpretability of recommendations. By constructing knowledge graphs for relevant industries, which include various attributes such as categories and regions, the diversity of recommendations is enhanced, and the problems of data sparsity and cold start in traditional recommendations can be solved.
[0047] like Figure 2 As shown, User A clicked on Companies 1, 2, and 3, which they had previously clicked. Since Company 1 and Company 6 are both located in Shanghai, it can be predicted that User A might be interested in companies located in Shanghai, meaning User A might be interested in Company 6. Since Company 2 and Company 5 are both in the clothing category, and the legal representative of Company 2 is Xiaoming, who has invested in Company 5, User A is very likely to be interested in Company 5. Since the investor of Company 3 is Xiaoli, who has invested in Company 4, meaning Xiaoli has invested in both Company 3 and Company 4, and since User A clicked on Company 3, User A might also be interested in Company 4, which Xiaoli has invested in.
[0048] It should be noted that the aforementioned user historical behavior sequence is the seed material sequence of the target knowledge graph corresponding to the current user. In other words, the companies that the current user has clicked or browsed in the past are used as the seed material of the target knowledge graph, and these seed materials serve as the starting point of the knowledge graph. That is, each seed material in the above seed material sequence is used as the starting point of the target knowledge graph. For example, a user set U and a company set V are predetermined, and a matrix Y is defined. When user u in user set U interacts with company v in company set V, i.e., when user u clicks on company v, the elements in matrix Y are 1; otherwise, they are 0. That is, all the companies corresponding to elements of 1 in matrix Y corresponding to the current user represent the user's historical behavior.
[0049] In this embodiment, the step of modeling the user preference material corresponding to the user historical behavior sequence determined based on the target knowledge graph to obtain the corresponding preference material feature vector may include: obtaining the historical behavior of the current user within a preset time range to obtain the user historical behavior sequence; determining the preference material feature vector of each hop of the user historical behavior sequence based on the target knowledge graph, and adding the preference material feature vectors of each hop to obtain the preference material feature vector corresponding to the current user. For example, in the current enterprise recommendation scenario, user historical behavior with the timestamp of the most recent 7 days is selected to obtain the user historical behavior sequence of the most recent 7 days. That is, all entities in the above user historical behavior sequence are used as seed materials of the target knowledge graph, and each seed material is used as the starting point of the target knowledge graph. Then, according to the directed edge of the knowledge graph, it continues to expand outward by one hop to obtain the ripple sets corresponding to the propagation hop. Then, it continues to expand into different i-hop ripple sets (i = 1, 2, 3, ..., h). The larger i is, the farther away from the seed material, which means that the user's preference decays like water ripples.
[0050] Step S12: Use a preset bidirectional LSTM model to model the user's historical behavior sequence to obtain the corresponding user behavior feature vector.
[0051] In this embodiment, the user's historical behavior sequence is input into a preset data encoding layer for encoding, and then the encoded user historical behavior sequence is input into a preset bidirectional LSTM (Long Short-Term Memory) model for modeling, which yields the corresponding user behavior feature vector. That is, by using a bidirectional LSTM model to model the aforementioned user historical behavior sequence with temporal characteristics, user behavior feature vectors can be extracted. However, if the dimension of the user behavior feature vector is too large after the preset bidirectional LSTM model completes the modeling of the aforementioned user historical behavior sequence and outputs the corresponding user behavior feature vector, problems such as slow operation will occur, affecting the updating of recommendation information. Therefore, the user behavior feature vector can be input into a fully connected layer constructed based on a preset convolutional neural network algorithm for dimensionality reduction, obtaining the dimensionality-reduced user behavior feature vector, thereby improving operating efficiency. For example, see... Figure 3 As shown, the user's historical behavior sequence is first input into the Embedding Layer for encoding, then the encoded user historical behavior sequence is input into the bidirectional LSTM model for modeling, and then the user behavior feature vector output by the bidirectional LSTM model is input into the fully connected layer for dimensionality reduction, thereby obtaining the final user behavior embedding (i.e., user behavior feature vector).
[0052] It should be noted that when the encoded user history behavior sequence is input into the preset bidirectional LSTM model for modeling, it is input into the preset bidirectional LSTM model in the order of the execution time of the user behavior in the user history behavior sequence. That is, the user history behavior sequence is regarded as a time-series set.
[0053] Step S13: Determine the user feature vector corresponding to the current user based on the preferred material feature vector and the user behavior feature vector.
[0054] In this embodiment, the preferred materials of the current user are determined based on the target knowledge graph, and a corresponding preference material feature vector is obtained through modeling. Then, based on the preference material feature vector and the user behavior feature vector obtained by modeling the user's historical behavior sequence using a preset bidirectional LSTM model, the user feature vector corresponding to the current user is determined. Specifically, the preference material feature vector and the user behavior feature vector are added together to obtain the user feature vector corresponding to the current user. Figure 4As shown, the data to be processed in the current recommendation scenario is first preprocessed. The preprocessed data is used to construct a target knowledge graph to obtain the user's historical behavior sequence. Then, a bidirectional LSTM is used to model the obtained user historical behavior sequence to obtain the user behavior feature vector. The user's preferred materials are obtained through the target knowledge graph and modeled to obtain the user's preferred material feature vector. The sum of the user behavior feature vector and the user preferred material feature vector is used as the user embedding. Then, the product of the input material embedding and the user embedding is calculated to obtain the click probability.
[0055] Step S14: Based on the user feature vector and the material feature vector determined according to the preset materials, predict the click probability of the recommended material corresponding to the current user, and determine the corresponding material recommendation result according to the click probability of the recommended material.
[0056] Understandably, the user feature vector is multiplied by a material feature vector determined based on preset materials to obtain the click probability of the recommended material for the current user. Then, the corresponding material recommendation result is determined based on this click probability. For example, the user embedding is multiplied by the material embedding to obtain the click probability of the recommended material for the current user. Then, the recommended materials are sorted based on their click probabilities to determine the corresponding material recommendation result. For instance, after sorting based on the click probabilities, a top-k recommendation result list is obtained; that is, the top k recommended materials are obtained after sorting, thus yielding the corresponding recommendation result list.
[0057] It should be noted that since user enterprise updates are relatively slow, unlike news recommendations which are updated in real time and within seconds, user embeddings and material embeddings can be stored in advance to simplify the deployment of knowledge graph models. This eliminates the need for online calculations of the target knowledge graph. In other words, by storing the embeddings in advance, the knowledge graph model can be trained offline and used for prediction online, thus saving computation time and resources. This simplifies the project and makes the model easy to deploy.
[0058] like Figure 5As shown, firstly, given a user u and a material v, the user's historical behavior sequence is used as the seed enterprise. Based on the constructed target knowledge graph G, the seed enterprise is used as the starting point of the target knowledge graph. An initial propagation is performed to obtain the first-hop ripple set1, which is then expanded into k different ripple sets k. Each iteration of the ripple set generates a new random seed, thus resampling is performed during the relevant model training process, increasing data coverage. The preferred material embedding, i.e., seed embedding, is obtained based on the aforementioned user historical behavior sequence used as the seed enterprise. The seed enterprise is the head entity set. The first-hop tail entity set corresponding to the current head entity is obtained based on the directed graph of the target knowledge graph, where the directed graph represents the entity relationships. In other words, the similarity of all tail entities in the tail entity set with the current head entity is calculated and then weighted and summed. The weights are normalized using softmax; the more similar the head entity is to the tail entity, the higher its weight, and the more likely it is to be recommended. Then, a weighted average of all tail entity vectors is calculated to obtain the embedding corresponding to the preferred item in the first hop. This process is repeated to obtain the embedding corresponding to the preferred item in the k-th hop. Finally, the embeddings corresponding to the preferred items obtained in each hop are summed to obtain the user embedding. It is important that k is not too large, as a large k will lead to excessive model complexity, and the larger the number of hops, the more likely entities will overlap, introducing a lot of noise. The entity set for each hop is fixed, and a fixed number of samples can be used instead of selecting all vector triples. However, a new random seed must be generated each time to ensure the recommended entity coverage.
[0059] The seed embedding described above is used to obtain the user behavior embedding through a user behavior sequence model. This adds a user time-based behavior sequence model to the knowledge graph-based recommendation method, allowing the model to detect changes in user interests over time and thus better recommend recently interested materials to users. Then, the user behavior embedding and the user embedding are summed to obtain the final user embedding for the current user u. The user behavior sequence model can be a bidirectional LSTM model. Finally, the user embedding for the current user u is multiplied by the material embedding for a given material v to obtain the click-through rate of the recommended material for the current user.
[0060] As can be seen, the embodiments of this application determine user preference materials based on the constructed target knowledge graph, solving the problems of data sparsity and cold start in traditional recommendation systems. In other words, the knowledge graph can describe the interaction relationships of a large number of user materials, such as click relationships, and can potentially discover user preferences; it also helps to find potential connections, improve the accuracy of recommended materials, and help to reasonably expand user interests and increase the diversity of recommended materials; it connects the user's historical behavior information with the recommendation records, increasing the interpretability of the recommendations; by using a preset bidirectional LSTM model combined with the target knowledge graph, it can overcome the problem that the target knowledge graph treats the user's historical behavior sequence as an unordered set, and consider the impact of user preferences over time, thereby discovering changes in user preferences over time, which can enhance material coverage. That is, the embodiments of this application improve the traditional Ripple Net framework, combining the Ripple Net-based knowledge graph recommendation method with a bidirectional LSTM model to extract complete user embeddings, fully explore the user's potential preferences, and stimulate the propagation of user preferences to the entity set of the knowledge graph.
[0061] Accordingly, this application also discloses an information recommendation device, see [link to relevant documentation]. Figure 6 As shown, the device includes:
[0062] Knowledge graph construction module 11 is used to construct the target knowledge graph;
[0063] The first feature vector determination module 12 is used to model the user preference material corresponding to the user historical behavior sequence determined based on the target knowledge graph to obtain the corresponding preference material feature vector; the user historical behavior sequence is the seed material sequence of the target knowledge graph corresponding to the current user;
[0064] The second feature vector determination module 13 is used to model the user's historical behavior sequence using a preset bidirectional LSTM model to obtain the corresponding user behavior feature vector.
[0065] The third feature vector determination module 14 is used to determine the user feature vector corresponding to the current user based on the preferred material feature vector and the user behavior feature vector.
[0066] The click probability prediction module 15 is used to predict the click probability of the recommended material corresponding to the current user based on the user feature vector and the material feature vector determined according to the preset material, and to determine the corresponding material recommendation result based on the click probability of the recommended material.
[0067] As can be seen from the above, this embodiment determines user preference materials based on the constructed target knowledge graph, which solves the problems of data sparsity and cold start in traditional recommendation systems, and can enhance the interpretability and diversity of recommendations. By using a preset bidirectional LSTM model combined with the target knowledge graph, it can overcome the problem that the target knowledge graph treats the user's historical behavior sequence as an unordered set, takes into account the time series problem, and thus discovers the changes in user preferences over time, thereby enhancing the material coverage.
[0068] In some specific embodiments, the knowledge graph construction module 11 may specifically include:
[0069] The knowledge graph construction unit is used to preprocess the data to be processed in the recommendation scenario to determine the corresponding graph entities and the entity relationships corresponding to the graph entities, and to construct the corresponding target knowledge graph based on the graph entities and the entity relationships.
[0070] In some specific embodiments, the first feature vector determination module 12 may specifically include:
[0071] The historical behavior acquisition unit is used to acquire the historical behavior of the current user within a preset time range to obtain the user's historical behavior sequence;
[0072] The first feature vector determination unit is used to determine the preference material feature vector of each hop of the user's historical behavior sequence based on the target knowledge graph, and to add the preference material feature vectors of each hop to obtain the preference material feature vector corresponding to the current user.
[0073] In some specific embodiments, the second feature vector determination module 13 may specifically include:
[0074] A behavior sequence encoding unit is used to input the user's historical behavior sequence into a preset data encoding layer for encoding;
[0075] The modeling unit is used to input the encoded user history behavior sequence into a preset bidirectional LSTM model for modeling, and obtain the corresponding user behavior feature vector.
[0076] In some specific embodiments, after modeling the encoded user historical behavior sequence input value using a pre-set bidirectional LSTM model to obtain the corresponding user behavior feature vector, it may specifically include:
[0077] The dimensionality reduction unit is used to input the user behavior feature vector into a fully connected layer constructed based on a preset convolutional neural network algorithm to reduce the dimensionality and obtain the dimensionality-reduced user behavior feature vector.
[0078] In some specific embodiments, the third feature vector determination module 14 may specifically include:
[0079] The addition unit adds the preferred material feature vector to the user behavior feature vector to obtain the user feature vector corresponding to the current user.
[0080] In some specific embodiments, the click probability prediction module 15 may specifically include:
[0081] The dot product unit performs a dot product between the user feature vector and the material feature vector determined based on preset materials to obtain the click probability of the recommended material corresponding to the current user.
[0082] Furthermore, embodiments of this application also provide an electronic device. Figure 7 This is a structural diagram of an electronic device 20 according to an exemplary embodiment. The content of the diagram should not be construed as limiting the scope of this application.
[0083] Figure 7 This is a schematic diagram of the structure of an electronic device 20 provided in an embodiment of this application. Specifically, the electronic device 20 may include: at least one processor 21, at least one memory 22, a power supply 23, a communication interface 24, an input / output interface 25, and a communication bus 26. The memory 22 stores a computer program, which is loaded and executed by the processor 21 to implement the relevant steps in the information recommendation method disclosed in any of the foregoing embodiments. Furthermore, the electronic device 20 in this embodiment may specifically be an electronic computer.
[0084] In this embodiment, the power supply 23 is used to provide operating voltage for each hardware device on the electronic device 20; the communication interface 24 can create a data transmission channel between the electronic device 20 and external devices, and the communication protocol it follows can be any communication protocol applicable to the technical solution of this application, and is not specifically limited here; the input / output interface 25 is used to acquire external input data or output data to the outside world, and its specific interface type can be selected according to specific application needs, and is not specifically limited here.
[0085] In addition, the memory 22, as a carrier for resource storage, can be a read-only memory, random access memory, disk or optical disk, etc. The resources stored thereon can include operating system 221, computer program 222, etc., and the storage method can be temporary storage or permanent storage.
[0086] The operating system 221 is used to manage and control the various hardware devices on the electronic device 20 and the computer program 222, which may be Windows Server, Netware, Unix, Linux, etc. In addition to including computer programs capable of performing the information recommendation method executed by the electronic device 20 as disclosed in any of the foregoing embodiments, the computer program 222 may further include computer programs capable of performing other specific tasks.
[0087] Furthermore, this application also discloses a storage medium storing a computer program, which, when loaded and executed by a processor, implements the information recommendation method steps disclosed in any of the foregoing embodiments.
[0088] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since it corresponds to the method disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to in the method section.
[0089] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0090] The above provides a detailed description of the information recommendation method, apparatus, device, and storage medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. An information recommendation method, characterized in that, include: A target knowledge graph is constructed, and user preference materials corresponding to user historical behavior sequences determined based on the target knowledge graph are modeled to obtain corresponding preference material feature vectors; the user historical behavior sequence is the seed material sequence of the target knowledge graph corresponding to the current user; The user's historical behavior sequence is input into a preset data encoding layer for encoding, and the encoded user historical behavior sequence is input into a preset bidirectional LSTM model for modeling to obtain the corresponding first user behavior feature vector; wherein, the preset bidirectional LSTM model processes the user's historical behavior sequence based on the time sequence of user behavior occurrence; The first user behavior feature vector is input into a fully connected layer constructed based on a preset convolutional neural network algorithm for dimensionality reduction, resulting in a dimensionality-reduced second user behavior feature vector. The user feature vector corresponding to the current user is determined based on the preferred material feature vector and the second user behavior feature vector; wherein, the user feature vector is a vector obtained by adding the preferred material feature vector and the second user behavior feature vector together. Based on the user feature vector and the material feature vector determined according to preset materials, predict the click probability of the recommended material corresponding to the current user, and determine the corresponding material recommendation result based on the click probability of the recommended material; The step of modeling the user preference material corresponding to the user historical behavior sequence determined based on the target knowledge graph to obtain the corresponding preference material feature vector includes: Obtain the historical behavior of the current user within a preset time range to obtain the user's historical behavior sequence; Based on the target knowledge graph, the preference material feature vector for each hop of the user's historical behavior sequence is determined, and the preference material feature vectors for each hop are summed to obtain the preference material feature vector corresponding to the current user. The process corresponding to any hop includes: selecting the user's historical behavior within a preset time period as the timestamp, obtaining the user's historical behavior sequence within the preset time period, and using all entities in the user's historical behavior sequence as seed materials of the target knowledge graph, with each seed material serving as the starting point of the target knowledge graph, and then continuing to expand outward by one hop according to the directed edges of the target knowledge graph.
2. The information recommendation method according to claim 1, characterized in that, The construction of the target knowledge graph includes: The data to be processed in the recommendation scenario is preprocessed to determine the corresponding graph entities and the entity relationships corresponding to the graph entities, and the corresponding target knowledge graph is constructed based on the graph entities and the entity relationships.
3. The information recommendation method according to claim 1 or 2, characterized in that, The step of predicting the click probability of the recommended material corresponding to the current user based on the user feature vector and the material feature vector determined according to preset materials includes: The user feature vector is multiplied by the material feature vector determined according to the preset materials to obtain the click probability of the recommended material for the current user.
4. An information recommendation device, characterized in that, include: The knowledge graph construction module is used to construct the target knowledge graph; The first feature vector determination module is used to model the user preference material corresponding to the user historical behavior sequence determined based on the target knowledge graph to obtain the corresponding preference material feature vector; the user historical behavior sequence is the seed material sequence of the target knowledge graph corresponding to the current user; The second feature vector determination module is used to input the user's historical behavior sequence into a preset data encoding layer for encoding, and input the encoded user historical behavior sequence into a preset bidirectional LSTM model for modeling to obtain the corresponding first user behavior feature vector; wherein, the preset bidirectional LSTM model processes the user's historical behavior sequence based on the time sequence of user behavior occurrence; The dimension reduction unit is used to input the first user behavior feature vector into a fully connected layer constructed based on a preset convolutional neural network algorithm to reduce the dimension, thereby obtaining the dimension-reduced second user behavior feature vector. The third feature vector determination module is used to determine the user feature vector corresponding to the current user based on the preferred material feature vector and the second user behavior feature vector; wherein, the user feature vector is a vector obtained by adding the preferred material feature vector and the second user behavior feature vector together; The click probability prediction module is used to predict the click probability of the recommended material corresponding to the current user based on the user feature vector and the material feature vector determined according to the preset material, and to determine the corresponding material recommendation result based on the click probability of the recommended material. The first feature vector determination module includes: The historical behavior acquisition unit is used to acquire the historical behavior of the current user within a preset time range to obtain the user's historical behavior sequence; The first feature vector determination unit is used to determine the preference material feature vector of each hop of the user's historical behavior sequence based on the target knowledge graph, and to add the preference material feature vectors of each hop to obtain the preference material feature vector corresponding to the current user; wherein, the process corresponding to any hop includes: selecting the user's historical behavior within a preset time period as the timestamp, obtaining the user's historical behavior sequence within the preset time period, and using all entities in the user's historical behavior sequence as seed materials of the target knowledge graph, with each seed material serving as the starting point of the target knowledge graph, and then continuing to expand outward by one hop according to the directed edges of the target knowledge graph.
5. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the information recommendation method as described in any one of claims 1 to 3.
6. A computer-readable storage medium, characterized in that, Used to store a computer program; wherein, when the computer program is executed by a processor, it implements the steps of the information recommendation method as described in any one of claims 1 to 3.