Method for recommending control operation by using classification information and electronic device for performing same

Abstract

The present invention relates to a method for recommending a control operation, and an electronic device. The method may extract a first hashtag corresponding to first content. The method may generate a first embedding vector corresponding to the first hashtag on the basis of the first hashtag and the first content. The method may determine a first cluster corresponding to the first embedding vector among a plurality of clusters. The method may generate personalized classification information corresponding to the first hashtag on the basis of the first embedding vector and a first center point vector corresponding to a center point of the first cluster. The method may provide a recommendation application and a recommendation control operation from personalized classification information by using a neural network model. The first embedding vector may include a first hashtag vector corresponding to the first hashtag and a first content vector corresponding to the first content.

Description

Method for recommending control actions using classification information and electronic device for performing the same

[0001] The present disclosure relates to a method and electronic device for recommending control actions using classification information. More specifically, the present disclosure relates to a method and electronic device for generating personalized classification information using hashtag information and recommending applications and control actions from the personalized classification information through a neural network model.

[0002] With the advancement of technology, users have become able to easily access diverse content across various services through electronic devices. While the advantage of being exposed to such a vast amount of content is the ability to encounter a wide variety, the disadvantage is that unless a user has sufficient usage history for an individual service, they may not receive appropriate content that reflects their specific interests. Therefore, methods are being devised to provide personalized experiences by selecting content that matches the user's interests.

[0003] One method to provide a personalized experience to users is to determine content interests using fixed keywords. However, using fixed keywords has limitations, such as a potential lack of specificity for each user, difficulty in dynamically reflecting changes in interests, and the inability to differentiate content for each user even with the same keyword.

[0004] Accordingly, the present invention aims to provide a method for generating personalized classification information from content to reflect the specific interests of a user and dynamically reflect changing interests, and for recommending various control operations within an electronic device based thereon.

[0005] According to one embodiment of the present disclosure, a method is provided for an electronic device to recommend a control action. The method may extract a first hashtag corresponding to a first content. Based on the first hashtag and the first content, the method may generate a first embedding vector corresponding to the first hashtag. The method may determine a first cluster corresponding to the first embedding vector among a plurality of clusters. Based on the first embedding vector and a first centroid vector corresponding to the centroid of the first cluster, the method may generate personalized classification information corresponding to the first hashtag. The method may provide a recommended application and a recommended control action from the personalized classification information using a neural network model. The first embedding vector may include a first hashtag vector corresponding to the first hashtag and a first content vector corresponding to the first content.

[0006] According to one embodiment of the present disclosure, an electronic device for recommending a control operation may include a memory for storing at least one instruction and at least one processor for executing at least one instruction stored in the memory. At least one processor may extract a first hashtag corresponding to a first content. At least one processor may generate a first embedding vector corresponding to the first hashtag based on the first hashtag and the first content. At least one processor may determine a first cluster corresponding to the first embedding vector among a plurality of clusters. At least one processor may generate personalized classification information corresponding to the first hashtag based on the first embedding vector and a first centroid vector corresponding to the centroid of the first cluster. At least one processor may provide a recommendation application and a recommendation control operation from the personalized classification information using a neural network model. The first embedding vector may include a first hashtag vector corresponding to the first hashtag and a first content vector corresponding to the first content.

[0007] According to one embodiment of the present disclosure, a computer-readable recording medium may be provided that records a program for executing the method on a computer.

[0008] FIG. 1 is a diagram showing a plurality of modules for a control operation that recommends a control operation according to one embodiment.

[0009] FIG. 2 is a diagram illustrating a specific operation for generating personalized classification information for recommending control operations according to one embodiment.

[0010] FIG. 3 is a diagram showing a specific operation that recommends a control operation based on personalized classification information according to one embodiment.

[0011] FIG. 4 is a flowchart for explaining a control operation that recommends a control operation according to one embodiment.

[0012] FIG. 5 is a specific flowchart for generating personalized classification information for recommending control actions according to one embodiment.

[0013] FIG. 6 is a diagram illustrating a method for performing clustering to generate personalized classification information according to one embodiment.

[0014] FIG. 7 is an example diagram of an application-control action table for recommending control actions according to one embodiment.

[0015] FIG. 8 is a flowchart illustrating the operation of updating personalized classification information according to one embodiment.

[0016] FIG. 9 is a flowchart illustrating a method for recommending control operations and applications according to one embodiment.

[0017] FIG. 10 is a schematic block diagram of an electronic device according to one embodiment.

[0018] Hereinafter, embodiments of the present disclosure will be described in detail with reference to the attached drawings.

[0019] The present disclosure is capable of various modifications and may have various embodiments, and specific embodiments are illustrated in the drawings and described in detail in the detailed description. However, this is not intended to limit the embodiments of the present disclosure, and it should be understood that the present disclosure includes all modifications, equivalents, and substitutions that fall within the spirit and scope of the various embodiments.

[0020] In describing the embodiments, detailed descriptions of related prior art are omitted if it is determined that such descriptions would unnecessarily obscure the essence of the present disclosure. Furthermore, numbers used in the description of the specification (e.g., first, second, etc.) are merely identifiers to distinguish one component from another.

[0021] The terms used in the embodiments of this specification have been selected to be as widely used as possible, taking into account the functions of this disclosure; however, these terms may vary depending on the intent of those skilled in the art, case law, the emergence of new technologies, etc. Additionally, in specific cases, terms have been arbitrarily selected by the applicant, and in such cases, their meanings will be described in detail in the description section of the relevant embodiments. Therefore, terms used in this disclosure should be defined not merely by their names, but based on their meanings and the content throughout this disclosure.

[0022] The scope of the present disclosure may be defined by the claims set forth below rather than by the detailed description above. Various features mentioned in one claim category of the present disclosure (e.g., in method claims) may also be claimed in other claim categories (e.g., in system claims). Furthermore, an embodiment of the present disclosure may include not only combinations of features specified in the appended claims but also various combinations of individual features within the claims. The scope of the present disclosure should be interpreted as including all modifications or variations derived from the meaning and scope of the claims and their equivalents.

[0023] In addition, components expressed as '~part (unit),' 'module,' etc. in this disclosure may consist of two or more components combined into a single component, or a single component may be divided into two or more components according to more detailed functions. These functions may be implemented in hardware or software, or through a combination of hardware and software. Furthermore, each component described below may additionally perform some or all of the functions performed by other components in addition to the primary function it is responsible for, and it is obvious that some of the primary functions performed by each component may be exclusively performed by other components.

[0024] Singular expressions may include plural expressions unless the context clearly indicates otherwise. Terms used herein, including technical or scientific terms, may have the same meaning as generally understood by those skilled in the art as described in this specification.

[0025] Throughout this disclosure, unless specifically stated otherwise, "or" is inclusive and not exclusive. Accordingly, "A or B" may mean "A, B, or both" unless clearly indicated otherwise in the context. In this disclosure, the phrases "at least one of" or "one or more of" may mean that different combinations of one or more of the listed items may be used, or that only any one of the listed items is required. For example, "at least one of A, B, and C" may include any of the following combinations: A, B, C, A and B, A and C, B and C, or A and B and C.

[0026] It will be understood that each block of the process flow diagrams and combinations of the flow diagrams can be executed by computer program instructions. Since these computer program instructions can be loaded into the processor of a general-purpose computer, a specialized computer, or other programmable data processing equipment, the instructions executed through the processor of the computer or other programmable data processing equipment create means to perform the functions described in the flow diagram block(s). Since these computer program instructions can also be stored in computer-available or computer-readable memory that can be directed toward the computer or other programmable data processing equipment to implement the function in a specific way, the instructions stored in such computer-available or computer-readable memory can also produce a manufactured item containing the means of instruction to perform the function described in the flow diagram block(s). Since computer program instructions can be loaded onto a computer or other programmable data processing equipment, instructions that perform a series of operation steps on the computer or other programmable data processing equipment to create a process executed by the computer can also provide steps for executing the functions described in the flowchart block(s).

[0027] Additionally, each block may represent a module, segment, or part of code containing one or more executable instructions for executing a specified logical function(s). It should also be noted that in some alternative execution examples, the functions mentioned in the blocks may occur out of order. For instance, two blocks described in succession may actually be executed substantially simultaneously, or the blocks may be executed in reverse order according to their corresponding functions.

[0028] When a part of a specification is described as "including" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. Furthermore, terms such as "...part" or "module" as used in the specification refer to a unit that processes at least one function or operation, and this may be implemented in hardware or software, or as a combination of hardware and software.

[0029] Embodiments of the present disclosure are described below in detail with reference to the attached drawings so that those skilled in the art can easily implement them. However, the present disclosure may be embodied in various different forms and is not limited to the embodiments described herein. Furthermore, in order to clearly explain the present disclosure in the drawings, parts unrelated to the explanation have been omitted, and similar parts throughout the specification are denoted by similar reference numerals.

[0030] Functions related to artificial intelligence according to the present disclosure are operated through a processor and memory. The processor may be composed of one or more processors. In this case, the one or more processors may be general-purpose processors such as CPUs, APs, and DSPs (Digital Signal Processors), graphics-dedicated processors such as GPUs and VPUs (Vision Processing Units), or artificial intelligence-dedicated processors such as NPUs. The one or more processors control the processing of input data according to predefined operation rules or artificial intelligence models stored in memory. Alternatively, if the one or more processors are artificial intelligence-dedicated processors, the artificial intelligence-dedicated processors may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0031] The predefined rules of operation or artificial intelligence models are characterized by being created through learning. Here, being created through learning means that a predefined rules of operation or artificial intelligence models configured to perform desired characteristics (or objectives) are created by a basic artificial intelligence model being trained using multiple learning data by a learning algorithm. Such learning may be performed on the device itself where the artificial intelligence according to the present disclosure is executed, or it may be performed through a separate server and / or system. Examples of learning algorithms include supervised learning, unsupervised learning, semi-supervised learning, or reinforcement learning, but are not limited to the examples described above.

[0032] An artificial intelligence model may be composed of multiple neural network layers. Each of the multiple neural network layers has multiple weight values ​​and performs neural network operations through operations between the results of previous layers and the multiple weights. The multiple weights possessed by the multiple neural network layers can be optimized based on the learning results of the artificial intelligence model. For example, the multiple weights may be updated so that the loss value or cost value obtained by the artificial intelligence model during the learning process is reduced or minimized. Artificial neural networks may include deep neural networks (DNNs), such as Convolutional Neural Networks (CNNs), Deep Neural Networks (DNNs), Recurrent Neural Networks (RNNs), Restricted Boltzmann Machines (RBMs), Deep Belief Networks (DBNs), Bidirectional Recurrent Deep Neural Networks (BRDNNs), or Deep Q-Networks, but are not limited to the examples mentioned above.

[0033] The terms used in this disclosure will be briefly explained, and an embodiment of the present invention will be described in detail.

[0034] The terms described below are defined considering their functions in the present invention, and these may vary depending on the intentions or practices of the user or operator. Therefore, their definitions should be based on the content throughout this specification.

[0035] In the present disclosure, 'content' may mean information or material provided to a user digitally by various media.

[0036] In the present disclosure, 'hashtag' may refer to a metadata tag used to classify specific topics or keywords within a service or across multiple services, and may serve to facilitate searching.

[0037] In the present disclosure, 'embedding vector' may refer to a numerical representation of high-dimensional data converted into a vector. For example, 'embedding vector' may refer to data converted into a fixed-dimensional vector for processing complex data such as text, images, and graphs.

[0038] In the present disclosure, 'cluster' may mean a group of data or data pointers having similar characteristics in data analysis.

[0039] In the present disclosure, 'clustering' may mean a process of classifying multiple data into groups having similar characteristics.

[0040] In the present disclosure, 'center point' may refer to a point representing the average position of data points within a cluster. For example, it may refer to data representing the average value of a plurality of data included within a cluster.

[0041] In the present disclosure, 'personalized classification information' may refer to hashtags that reflect user-specific interests regarding content, serve to facilitate searching, and may change dynamically according to changes in user interests.

[0042] In the present disclosure, 'control operation' may mean a series of operations for performing the functions of an application, and multiple applications may be dependent on a single control operation. For example, multiple applications may include the same control operation.

[0043] In the present disclosure, 'decoding' may mean the process of restoring encrypted information or compressed data to its original form.

[0044] In the present disclosure, 'feedback' refers to providing a response to an action or result, and may mean a process of adjusting learning based on output results in a neural network model.

[0045] In the present disclosure, the 'interface' can perform the role of interacting with a user in an application by transmitting and receiving input / output.

[0046] In the present disclosure, 'GNN (Graph Neural Network)' may refer to a neural network model used to learn the relationships between nodes, edges, and graphs as a deep learning model for processing data of a graph structure.

[0047] FIG. 1 is a diagram showing a plurality of modules for a control operation that recommends a control operation according to one embodiment.

[0048] In one embodiment of the present disclosure, an electronic device (100) may generate personalized classification information (120) from content information (110) and provide a recommendation application and a recommendation control operation (130) to a user using the personalized classification information (120). The electronic device (100) may provide a recommendation application and a recommendation control operation (130) using a neural network model (190) from the personalized classification information (120) generated using the content information (110).

[0049] The operation recommended by the electronic device (100) for control operation can be described as a software unit responsible for a specific function or role. The modules (140 to 180) illustrated in FIG. 1 may be classified according to function as operations performed by executing instructions stored in memory in the processor. Therefore, the operations described below as being performed by the modules (140 to 180) illustrated in FIG. 1 can actually be seen as being performed by the processor.

[0050] Personalized classification information (120) may refer to personalized hashtags generated according to the results of clustering and embedding vector operations on hashtags generated from content information (110) through LLM, and may be used interchangeably with personalized summary information, personalized search information, or personalized information. Hereinafter, for convenience of explanation, personalized classification information having dynamic characteristics according to the user's interests is referred to as 'personalized classification information,' and general classification information extracted from content through a model such as LLM is referred to as 'hashtag.' In this disclosure, 'classification information' may include both hashtags and personalized classification information without further explanation.

[0051] Personalized classification information (120) can be continuously updated to reflect the user's content usage patterns as it has dynamic characteristics. The present disclosure discloses a method (130) of providing a user with recommendation control actions and recommendation applications based on the personalized classification information (120).

[0052] In one embodiment of the present disclosure, the electronic device (100) may be an electronic device capable of processing and outputting video or images. The electronic device (100) may be implemented in various forms including a display. For example, the electronic device (100) may be implemented as various electronic devices such as a TV, mobile phone, tablet PC, digital camera, camcorder, laptop computer, desktop, e-book reader, digital broadcasting terminal, PDA (Personal Digital Assistants), PMP (Portable Multimedia Player), navigation, MP3 player, wearable device, etc. The electronic device (100) may include a plurality of modules (140 to 180) and a neural network model (190) to generate user-specific personalized classification information (120) and to provide recommendation control operations and recommendation applications (130) from the personalized classification information.

[0053] The electronic device (100) can generate personalized classification information (120) from content information (110) through the operation of a plurality of modules (140 to 180). The electronic device (100) can provide recommendation applications and recommendation control operations (130) from the personalized classification information (120) through a neural network model (190).

[0054] In one embodiment of the present disclosure, an electronic device (100) may acquire a plurality of contents including a first content. The electronic device (100) may extract a hashtag for the first content and generate a hashtag vector corresponding to the hashtag. A hashtag may refer to a tag for classifying a topic or keyword included in the content. The electronic device (100) may generate a content vector corresponding to the first content. In the present disclosure, an embedding vector corresponding to a hashtag is referred to as a "hashtag vector," and a vector corresponding to the content is referred to as a "content vector."

[0055] In one embodiment, the electronic device (100) may store all or part of a plurality of contents including a first content by configuring them into a knowledge graph. The electronic device (100) may extract a hashtag for the first content based on the knowledge graph.

[0056] In one embodiment, a knowledge graph may be used to provide a personalized service to a user through an electronic device (100) or a neural network model included in the electronic device (100). The knowledge graph may refer to an object processed by a GNN.

[0057] In one embodiment, the electronic device (100) can generate an embedding vector corresponding to a first hashtag, and the embedding vector may include a hashtag vector and a content vector. The electronic device (100) can convert various forms of content into the form of an embedding vector by performing a multimodal-based embedding transformation on the content.

[0058] In one embodiment, the electronic device (100) may store all or part of the hashtags including the first hashtag by configuring them into a knowledge graph. The electronic device (100) may generate an embedding vector corresponding to the first hashtag based on the knowledge graph.

[0059] Hereinafter, without further explanation, the term "embedding vector" in this disclosure may refer to an embedding vector including a hashtag vector and a content vector. That is, as illustrated in FIG. 2, the embedding vector may refer to an embedding vector in the form of a combination of a hashtag vector and a content vector.

[0060] The electronic device (100) can determine a cluster corresponding to an embedding vector and generate personalized classification information (120) by performing calculations on the embedding vector and the cluster center point vector. The cluster center point vector may include the average value of each element of a plurality of embedding vectors included in the cluster. In one embodiment, the cluster center point vector may represent a combined form of a vector composed of the average value of a hashtag vector and a vector composed of the average value of a content vector. In one embodiment, the electronic device (100) can generate personalized classification information (120) by replacing the content vector included in the embedding vector with the content vector included in the cluster center point vector and decoding. A specific method is described below with reference to FIG. 2.

[0061] In one embodiment, the electronic device (100) may store all or part of the embedding vector, the cluster corresponding to the embedding vector, and the personalized classification information (120) that can be obtained by calculating the embedding vector and the center point vector of the cluster, by configuring them into a knowledge graph.

[0062] In one embodiment, the knowledge graph may be newly created based on all or part of the embedding vectors, clusters, and personalized classification information (120). In one embodiment, the knowledge graph may be a modified knowledge graph created by modifying some or all nodes and connections from a base knowledge graph that has already been created using data other than all or part of the embedding vectors, clusters, and personalized classification information (120). For example, the base knowledge graph may be updated by modifying some or all nodes and connections.

[0063] In one embodiment, the knowledge graph may be a modified knowledge graph created by adding additional nodes and / or connections to a base knowledge graph that has already been created using data other than all or part of the embedding vectors, clusters, and personalized classification information (120). In this case, the base knowledge graph or the modified knowledge graph may be a component constituting a platform for personalized services included in an electronic device (100) to provide personalized services to a user.

[0064] The electronic device (100) can determine and provide recommended applications and recommended control actions (130) from personalized classification information (120) through a neural network model (190). In one embodiment, the electronic device (100) can calculate vector similarity between nodes using a Graph Neural Network (GNN) model and determine and provide recommended control actions and recommended applications from personalized classification information. For specific methods below, refer to FIG. 3.

[0065] In one embodiment, the electronic device (100) can extract a hashtag corresponding to the content information (110) from the content information (110) using a hashtag extraction module (140). The hashtag extraction module (140) can extract a hashtag from the content information (110) using LLM. The hashtag extraction module (140) can extract a hashtag by processing high-dimensional data such as text, images, and audio included in the content information (110).

[0066] For example, the hashtag extraction module (140) can use LLM to identify important words and topics in the context of the text and extract hashtags from the text based on learned knowledge. In one embodiment, the hashtag extraction module (140) can process not only text but also unstructured data such as images or audio using multimodal learning techniques. For example, the hashtag extraction module (140) can process image data using a CLIP (Contrastive Language-Image Pretraining) model or a BLIP (Bootstrapping Language-Image Pretraining) model that understands images and text simultaneously, but the model for processing image data is not limited thereto. The hashtag extraction module (140) can process audio data using STT (Speech To Text) that converts audio data into text data, but the method of processing audio data is not limited thereto.

[0067] In one embodiment, the electronic device (100) may generate an embedding vector corresponding to the extracted hashtag using an embedding generation module (150). An embedding vector may refer to data converted into a fixed-dimensional vector for processing high-dimensional data. In the present disclosure, an embedding vector may refer to a combined form of a hashtag vector and a content vector.

[0068] The electronic device (100) can generate an embedding vector including a hashtag vector converted from a hashtag into a vector and a content vector for the content using an embedding generation module (150). For example, the embedding vector may mean a vector in which the hashtag vector and the content vector are concatenated. In one embodiment, the hashtag vector and the content vector may be generated using the same embedding model. In one embodiment, the hashtag vector and the content vector may be vectors of the same dimension as they use the same embedding model.

[0069] In one embodiment, a content vector may include information about the content and information about the use of the content. Information about the content refers to information about the content itself and may include, but is not limited to, the type of content, information about the service on which the content is provided, and the type of data included in the content. Information about the use of the content may refer to information regarding the interaction between the content and the user and may include, but is not limited to, information about the time the user used the content, the number of times the content was clicked, and information about the path through which the user entered the content.

[0070] In one embodiment, information about the content is generated as a vector, and information about the use of the content is calculated as a weight and included in the content vector. For example, if a user is exposed to multiple contents for 100 minutes and content 'A' is exposed for 30 minutes, the electronic device (100) can calculate that content 'A' has a weight of 0.3. The electronic device (100) can generate a content vector that reflects information about the use of content for content 'A' by multiplying each element of the vector corresponding to the information about content 'A' by 0.3.

[0071] In one embodiment, the electronic device (100) can determine a cluster corresponding to an embedding vector generated through an embedding generation module (150) using a clustering module (160). The electronic device (100) can determine a cluster corresponding to an embedding vector by calculating the distance between an embedding vector generated by the embedding generation module (150) and a center point vector corresponding to the center point of a plurality of clusters among a plurality of clusters.

[0072] For example, an electronic device (100) can cluster the 'a' embedding vector into cluster B when the distance between the 'a' embedding vector and the center vectors of multiple clusters (A, B, C) is 0.2, 0.1, and 0.3, respectively.

[0073] In one embodiment, the electronic device (100) may determine that there is no cluster corresponding to the embedding vector if the distance between the embedding vector generated through the embedding module (150) and the center vectors of a plurality of clusters is greater than or equal to a preset threshold value. The electronic device (100) may generate a new cluster containing the embedding vector.

[0074] For example, if the distances from the center vectors of multiple clusters (A, B, C) to the 'b' embedding vector are 0.4, 0.7, and 0.5 respectively, and the preset threshold value is 0.3, the electronic device (100) determines that there is no cluster corresponding to the 'b' embedding vector and can create a new cluster D containing the 'b' embedding vector.

[0075] In one embodiment, the electronic device (100) can generate personalized classification information (120) corresponding to a hashtag based on an embedding vector and a center point vector corresponding to a center point of a cluster using a personalized classification information generation module (170). Hereinafter, for convenience of explanation, the embedding vector is referred to as the 'first embedding vector' and the cluster in which the first embedding vector is clustered is referred to as the 'first cluster'.

[0076] The electronic device (100) can compute a first embedding vector and a center point vector of a first cluster using an embedding computation module (175). The electronic device (100) can generate a second embedding vector by computing the hashtag vector included in the first embedding vector and the content vector included in the center point vector. In one embodiment, the second embedding vector may refer to a vector obtained by concatenating the hashtag vector included in the first embedding vector and the content vector included in the center point vector. Hereinafter, for convenience of explanation, the embedding vector obtained by computing the hashtag vector included in the first embedding vector and the content vector included in the center point vector of the first cluster is referred to as the 'second embedding vector'.

[0077] The electronic device (100) can generate personalized classification information (120) by decoding a second embedding vector generated through computation using a decoding module (180). For a specific method of generating personalized classification information (120), refer to FIGS. 2, 5, and 6.

[0078] In one embodiment, the electronic device (100) can generate personalized classification information (120) corresponding to hashtags extracted from content through a plurality of modules (140 to 180) as new content information is added, and can update the cluster center point vector. As the electronic device (100) updates the cluster center point vector, it can update the embedding vector corresponding to the previously generated personalized classification information.

[0079] For example, the electronic device (100) can generate personalized classification information for a first embedding vector included in a first cluster. The electronic device (100) can acquire and / or generate a new third embedding vector included in the first cluster. Hereinafter, for convenience, the embedding vector newly acquired and / or generated by the electronic device (100) is referred to as the 'third embedding vector'. The electronic device (100) can update the center point vector of the first cluster as the third embedding vector is clustered into the first cluster. As the electronic device (100) updates the center point vector of the first cluster, the second embedding vector and personalized classification information can be updated. For example, as the electronic device (100) updates the center point vector of the first cluster, the content vector included in the center point vector is updated. As the content vector included in the center point vector is updated, the second embedding vector obtained as a result of calculating the hashtag vector of the first embedding vector and the content vector of the center point vector is updated. Since the second embedding vector is updated, the personalized classification information generated by decoding the second embedding vector is also updated. For specific operations to update the personalized classification information below, refer to FIG. 8.

[0080] In one embodiment, the electronic device (100) may obtain recommended applications and recommended control actions (130) from personalized classification information (120) using a neural network model (190) and provide them to the user. In one embodiment, the neural network model (190) may be a Graph Neural Network (GNN) and may process a graph having a plurality of contents, a plurality of hashtags, a plurality of control actions, and a plurality of applications as nodes.

[0081] A neural network model (190) can be trained based on multiple classification information corresponding to multiple contents and control action history information corresponding to multiple classification information. In one embodiment, the multiple classification information may include personalized classification information generated to have hashtags and dynamic characteristics. For example, the neural network model can be trained based on user control action history information for each of the classification information a, b, and c included in content A. If content A includes classification information for "travel" and "seaweed soup," and there is control action history information such as "search for flight tickets" for "travel" and "grocery shopping" for "seaweed soup," the neural network model can be trained based on this classification information and control action history information. Hereinafter, for the training method of the neural network model, refer to FIG. 3.

[0082] In one embodiment, the electronic device (100) can obtain a recommended control action based on similarity between personalized classification information and candidate control actions using a neural network model (190). For example, the electronic device (100) can obtain a recommended control action for "A", which is personalized classification information as a learning result. The electronic device (100) can determine at least one candidate application corresponding to the recommended control action using an application-control action table. Since an application may include multiple control actions, the application-control action table may refer to a table that maps control actions that can be performed per application. The electronic device (100) can generate an application-control action table that indicates a mapping relationship between an application and a control action.

[0083] The electronic device (100) determines at least one candidate application corresponding to a recommended control action based on an application-control action table and can obtain a recommended application based on similarity between the candidate application and personalized classification information through a neural network model. The electronic device (100) can provide the recommended application and the recommended control action (130) obtained through the neural network model (190) to the user. For specific methods below, refer to FIGS. 3, 4, 7 and 10.

[0084] In one embodiment, the neural network model (190) may be updated based on feedback information regarding recommendation applications and recommendation control actions. The electronic device (100) may provide a plurality of recommendation applications and recommendation control actions to the user for one personalized classification information (120). For example, if a recommendation application is installed within the electronic device, an execution-related interface for the recommendation application may be provided. If a recommendation application is not installed within the electronic device, an installation-related interface for the recommendation application may be provided.

[0085] In one embodiment, the electronic device (100) may receive input values ​​from a user based on an interface provided to the user. The electronic device (100) may use the user's input values ​​as feedback information for a plurality of recommendation applications and recommendation control actions. For example, the electronic device may provide the user with actions a1 and a2 for application A and actions a1 and b1 for application B. The electronic device may receive input values ​​from the user for selecting action a1 for application B. The electronic device may update a neural network model using the input values ​​as feedback information. For a specific method, refer to FIG. 3 below.

[0086] In one embodiment, the electronic device (100) may obtain additional information required according to the recommendation control action and the recommendation application from another application. For example, the electronic device (100) may provide a payment action to the user for an application called "A", and when it receives an input signal for the action from the user, it may retrieve payment information from the Samsung Pay application.

[0087] In one embodiment, the electronic device (100) may request additional information from the user according to a recommendation control action and a recommendation application. For example, it may provide a reservation action to the user for an application named "B", and upon receiving an input signal for the action from the user, it may also provide an interface for the user to input personal information.

[0088] In one embodiment, the electronic device (100) may store some or all of the above-described operation history information, recommendation control operation, recommendation application, feedback information regarding recommendation control operation, additional information obtained from other applications, and additional information obtained by requesting from the user by configuring it into a knowledge graph. When the electronic device (100) provides at least one of the recommendation control operation, recommendation application, feedback information regarding recommendation control operation, additional information obtained from other applications, and additional information obtained by requesting from the user to be provided to the user in the future, the electronic device (100) may provide the information based on the contents of the configured knowledge graph. At this time, the knowledge graph may be an element constituting a platform for personalized services included in the electronic device (100) to provide personalized services to the user.

[0089] FIG. 2 is a diagram illustrating a specific operation for generating personalized classification information for recommending control operations according to one embodiment.

[0090] Referring to FIG. 2, the electronic device (100) can generate embedding vectors from content and generate personalized classification information corresponding to hashtags through clustering. The embedding vectors are high-dimensional data such as text, images, and audio obtained from content that has been converted into vectors and can be decoded after undergoing computation to generate personalized classification information.

[0091] In one embodiment of the present disclosure, an electronic device (100) may obtain a plurality of pieces of information from a plurality of pieces of content. The electronic device (100) may generate personalized classification information for a plurality of hashtags included in the plurality of pieces of content in order to provide a recommendation control operation and a recommendation application to the user. Hereinafter, for convenience of explanation, one of the plurality of pieces of content provided through the electronic device (100) is referred to as the 'first piece of content'.

[0092] In one embodiment, the electronic device (100) can extract a hashtag corresponding to the first content from the first content through a Large Language Model (LLM). The electronic device (100) can extract a hashtag from high-dimensional data, such as text, images, and audio, included in the first content using an LLM. A hashtag may refer to metadata used to classify contextually important topics or keywords within the content, and is generated based on predefined keywords.

[0093] In one embodiment, the electronic device (100) can perform preprocessing on various forms of high-dimensional data using LLM. For example, the electronic device (100) can remove unnecessary elements from content and refine data. The electronic device (100) can convert various forms of data, such as images or audio, into a text format. The electronic device (100) can identify the context of the text and select keywords that appear frequently or are deemed contextually important through natural language processing techniques. The electronic device (100) can generate hashtags by transforming the extracted keywords into an appropriate form.

[0094] In one embodiment, the electronic device (100) can generate an embedding vector corresponding to a hashtag based on a hashtag and content. The electronic device (100) can generate a first embedding vector (210) corresponding to a hashtag extracted from the first content. The first embedding vector (210) may include a hashtag vector (220) and a content vector (230) for the hashtag. The electronic device (100) can generate the hashtag vector (220) by converting the extracted hashtag into a vector using an embedding generation model. The electronic device (100) can generate a content vector (230) for the content using the same embedding generation model.

[0095] In one embodiment, the content vector (230) may include information about the content and information about the use of the content. Information about the content refers to information about the content itself and may include, but is not limited to, the type of content, information about the service on which the content is provided, and the type of data included in the content. Information about the use of the content may refer to information regarding the interaction between the content and the user and may include, but is not limited to, information about the time the user used the content, the number of times the content was clicked, and information about the path to the content.

[0096] In one embodiment, the electronic device (100) can convert information about content into a vector using an embedding generation model. For example, the electronic device (100) can convert information about content into a vector using a multimodal-based encoder. The electronic device (100) can generate a content vector (230) by converting information about content usage into weights and multiplying the content information converted into a vector element by element. In one embodiment, the electronic device (100) can generate a first embedding vector (210) by performing operations on the hashtag vector (220) and the content vector (230). For example, the electronic device (100) can generate a first embedding vector (210) by concatenating the hashtag vector (220) and the content vector (230).

[0097] In one embodiment, the electronic device (100) can determine a first cluster (250a) corresponding to a first embedding vector (210) among a plurality of clusters (250a, 250b, 250c). The electronic device (100) can determine a first cluster (250a) corresponding to the first embedding vector (210) by calculating the distance between the first embedding vector (210) and a center point vector (260a, 260b, 260c) corresponding to the center point of the plurality of clusters (250a, 250b, 250c) among a plurality of clusters (250a, 250b, 250c). For example, the electronic device (100) can determine the first cluster (250a) corresponding to the first embedding vector (210) based on the similarity between the center point vector (260a, 260b, 260c) of a plurality of clusters (250a, 250b, 250c) and the first embedding vector (210).

[0098] In one embodiment, the electronic device (100) may determine that there is no cluster corresponding to the first embedding vector (210) if the distance between the first embedding vector (210) and the center point vector (260a, 260b, 260c) corresponding to the center point of a plurality of clusters (250a, 250b, 250c) is less than or equal to a preset threshold. The electronic device (100) may generate a new cluster including the first embedding vector (210).

[0099] In one embodiment, the electronic device (100) can generate personalized classification information corresponding to a hashtag regarding the first embedding vector (210) based on the first embedding vector (210) and the center point vector (260a) of the first cluster (250a). In one embodiment, the electronic device (100) can generate a second embedding vector (280) by concatenating the hashtag vector (220) of the first embedding vector (210) and the content vector (270) contained in the center point vector (260a) of the first cluster. For example, the electronic device (100) can generate a second embedding vector (280) by concatenating the hashtag vector (220) of the first embedding vector (210) and the content vector (270) contained in the center point vector (260a) of the first cluster.

[0100] In one embodiment, the electronic device (100) can decode the second embedding vector (280) to generate personalized classification information (290). The electronic device (100) can decode the second embedding vector (280) to generate personalized classification information (290) corresponding to a hashtag.

[0101] In one embodiment, the electronic device (100) can update the center point vector of the cluster and update personalized classification information as new content information is added. For example, the electronic device (100) can generate a new third embedding vector from the content. The electronic device (100) can cluster the third embedding vector into a first cluster (250a). As the third embedding vector is included in the first cluster (250a), the electronic device (100) can update the center point vector (260a) of the first cluster to include information about the third embedding vector. As the electronic device (100) updates the center point vector (260a) of the first cluster, the content vector (270) included in the center point vector is updated. As the content vector (270) included in the center point vector is updated, the electronic device (100) can update the second embedding vector (280). For example, the electronic device (100) can update the content vector included in the second embedding vector (280) to the content vector (270) included in the center point vector that performed the update. The electronic device (100) can also update the personalized classification information (290) by decoding the updated second embedding vector (280). For specific update operations, refer to FIG. 8 below.

[0102] FIG. 3 is a diagram showing a specific operation that recommends a control operation based on personalized classification information according to one embodiment.

[0103] Referring to FIG. 3, the electronic device (100) can provide recommendation applications and recommendation control actions from personalized classification information using a neural network model. The electronic device (100) can infer recommendation applications and recommendation control actions using a neural network model.

[0104] In one embodiment of the present disclosure, an electronic device (100) can infer recommended applications and recommended control actions from personalized classification information using a learned neural network model and provide them to a user. In one embodiment, the neural network model may be a Graph Neural Network (GNN), and the neural network model may process a graph having a plurality of contents, a plurality of classification information, a plurality of control actions, and a plurality of applications as nodes.

[0105] Multiple classification information may include hashtags and personalized classification information having dynamic features. In one embodiment, the electronic device (100) may use both hashtags and personalized classification information as training data. For example, the electronic device (100) may provide recommendation applications and recommendation control actions based on personalized classification information that is subject to inference. The electronic device (100) may use the personalized classification information as training data based on user input information regarding recommendation applications and recommendation control actions. Hereinafter, classification information obtained from content using LLM is referred to as 'hashtags', and dynamic classification information generated by reflecting the user's interests is referred to as 'personalized classification information'.

[0106] In one embodiment of the present disclosure, a neural network model may be trained to generate a unique embedding vector by utilizing information between neighboring nodes for a graph having nodes corresponding to a plurality of contents, a plurality of classification information, a plurality of control actions, and a plurality of applications. The neural network model may calculate the similarity between nodes using the unique embedding vector of each node. Based on the similarity between nodes, the neural network model may infer recommendation control actions and recommendation applications for personalized classification information.

[0107] In one embodiment of the present disclosure, the electronic device (100) may set initial embedding vectors for nodes included in a graph for learning through a GNN. The electronic device (100) may set the initial embedding vector for classification information as an embedding vector generated during the process of extracting personalized classification information. For example, the electronic device (100) may set a second embedding vector before decoding personalized classification information as the initial embedding vector for the neural network model.

[0108] In one embodiment, the electronic device (100) can set the initial embedding vectors of other nodes of the neural network model using the embedding model used to generate personalized classification information.

[0109] In one embodiment, the electronic device (100) sets an initial embedding vector for classification information as an embedding vector (second embedding vector) for generating personalized classification information, but the initial embedding value for GNN learning can be stored separately. As the electronic device (100) updates the initial embedding value for GNN learning through learning, it can be stored separately from the embedding vector used to generate personalized classification information.

[0110] In one embodiment of the present disclosure, a neural network model may generate connection information between neighboring nodes as an adjacency matrix. In the adjacency matrix, if nodes are connected, the matrix element value may be set to '1', and if they are not connected, the matrix element value may be set to '0'. In one embodiment, if nodes are connected, it may mean that an edge exists between the nodes.

[0111] In one embodiment of the present disclosure, a neural network model can update a unique embedding vector for each node based on an adjacency matrix and weights for each edge. For example, if nodes A and B are connected, the neural network model may set the corresponding element value of the adjacency matrix to '1'. In one embodiment, the neural network model can predict potential connectivity between nodes by performing a Link Prediction Task that predicts whether each edge is connected. The Link Prediction Task is a task that estimates the probability that a given pair of nodes is connected, and the performance of the model can be optimized using a specific loss function. For example, the Link Prediction Task can minimize the difference between the actual connection status and the predicted connection probability by using Binary Cross-Entropy Loss or Margin-based Loss. Through this, the neural network model can learn an embedding vector that reflects the structural characteristics of the network. The neural network model can learn weights for the corresponding edges through learning.

[0112] In one embodiment of the present disclosure, a neural network model can generate edges between content nodes and classification information nodes. The neural network model can generate edges between content nodes and classification information nodes based on a method of extracting hashtags from content or generating personalized classification information from content using a model such as an LLM.

[0113] In one embodiment, the electronic device (100) may use a plurality of learning contents (310a, 310b) for training a neural network model. The electronic device (100) may extract at least one classification information (320a, 320b, 320c, 320d, 320e) from the learning contents. The electronic device (100) may extract hashtags from the learning contents using LLM. The electronic device (100) may generate personalized classification information from the content. Hereinafter, a method for extracting hashtags using LLM is described with reference to FIGS. 1, 2 and 4.

[0114] Multiple classification information may be extracted from a single learning content, and the same classification information may be extracted from multiple contents. For example, first to third classification information (320a, 320b, 320c) may be extracted simultaneously from the first learning content (310a) and the second learning content (310b), and first to fifth classification information (320a, 320b, 320c, 320d, 320e) may be extracted from the second learning content (310b).

[0115] In one embodiment of the present disclosure, a neural network model can generate edges between a control action node and an application node using / based on an application-control model table.

[0116] In one embodiment, the neural network model may use an application-control action table for a plurality of control actions and a plurality of applications. An application-control action table may refer to a table that maps control actions that can be performed per application, as one application may include a plurality of control actions and one control action may be dependent on a plurality of applications. For example, a phone app (340a) may generate edges with control actions such as making a call (330a) and answering a call (330b) based on a pre-generated application-control action table. Hereinafter, a detailed example of an application-control action table is referenced in FIG. 7.

[0117] In one embodiment of the present disclosure, a neural network model can generate edges between a classification information node and a control action node through learning and learn unique embeddings of the classification information node and the control action node. The neural network model can perform learning based on a plurality of classification information and control action history information corresponding to the plurality of classification information.

[0118] In one embodiment, the neural network model may be trained based on control actions performed by a user corresponding to classification information or hashtags. For example, the neural network model may perform training based on receiving an input signal that performs a control action called product search (330c) on the second classification information (320b). When the neural network model receives the input signal, it may update the edge weights between the second classification information (320b) and the product search (330c) through training. The neural network model may perform updates on the embedding vectors of the second classification information (320b) and product search (330b) nodes, respectively, according to the update of the edge weights. In one embodiment, the edge weights between the classification information node and the control action node may have a value between 0 and 1 after normalization.

[0119] In one embodiment, the neural network model may be trained based on an application executed for a control action performed by a user on classification information. For example, the neural network model may receive an input signal executing the Coupang app (340b) while receiving an input signal performing a control action called product search (330c) on the second classification information (320b). Applications capable of performing the control action called product search (330c) may include, in addition to the Coupang app (340b), the Samsung search app (not shown), etc. Based on the user's input signal, the neural network model may train an embedding vector for the second classification information (320b) node, an embedding vector for the product search (330c) node, and an embedding vector for the Coupang app (340b) node.

[0120] For example, the weight of the edge connecting the node of the second classification information (320b) and the product search (330c) node of the Coupang app (340b) is increased, and the weight of the edge connecting the node of the second classification information (320b) and the product search (330c) node of the Samsung search app (not shown) is decreased. Based on the updated weights, the neural network model can update the embedding vector of each node.

[0121] In one embodiment of the present disclosure, a trained neural network model can infer recommendation control actions and recommendation applications from personalized classification information generated from content. The neural network model can infer recommendation control actions and recommendation applications based on similarity calculated based on the embedding vectors of each of the personalized classification information node, control action node, and application node.

[0122] In one embodiment, the neural network model can infer (370) a recommendation control action and a recommendation application for the first personalized classification information (360a) and the second personalized classification information (360b) generated from the first content (350). The neural network model can obtain the recommendation control action that has the highest similarity to the personalized classification information (360a, 360b). In one embodiment, the neural network model can obtain a recommendation control action based on similarity between the personalized classification information and a plurality of candidate control actions. For example, the neural network model can obtain a control action called "Write Review" (330d), which has the highest similarity to the first personalized classification information (360a), as a recommendation control action.

[0123] In one embodiment, a neural network model can obtain a recommended application based on the similarity between at least one candidate application corresponding to a recommendation control action and personalized classification information. The neural network model can determine the application with the highest similarity to the personalized classification information among the candidate applications determined according to the obtained recommendation control action as the recommended application.

[0124] In one embodiment, there is no limitation on whether the neural network model provides a user with one control action and one application, one control action and multiple applications, or multiple control actions and multiple applications for one personalized classification information.

[0125] In one embodiment, the electronic device (100) may obtain additional information required according to the recommendation control action and the recommendation application from another application. For example, the electronic device (100) may provide a payment action to the user for an application called "A", and when it receives an input signal for the action from the user, it may retrieve payment information from the Samsung Pay application.

[0126] In one embodiment, the electronic device (100) may request additional information from the user according to a recommendation control action and a recommendation application. For example, it may provide a reservation action to the user for an application named "B", and upon receiving an input signal for the action from the user, it may also provide an interface for the user to input personal information.

[0127] FIG. 4 is a flowchart for explaining a control operation that recommends a control operation according to one embodiment.

[0128] Referring to FIG. 4, the electronic device (100) can generate personalized classification information corresponding to hashtags extracted from content. The electronic device (100) can provide recommendation control actions and recommendation applications to the user through the personalized classification information.

[0129] In step S410, the electronic device (100) can extract a first hashtag corresponding to the first content. A single piece of content may include multiple hashtags. In one embodiment, the electronic device (100) can extract hashtags from various forms of data included in the content.

[0130] In one embodiment of the present disclosure, the electronic device (100) can extract hashtags from the first content information using a Large Language Model (LLM). The electronic device (100) can process unstructured data, such as audio and images, as well as text, by having multimodal characteristics. For example, the electronic device (100) can process image data using a Contrastive Language-Image Pretraining (CLIP) model or a Bootstrapping Language-Image Pretraining (BLIP) model that understands images and text simultaneously, but the model for processing image data is not limited thereto. The electronic device (100) can process audio data using Speech To Text (STT) that converts audio data into text data, but the method of processing audio data is not limited thereto.

[0131] In one embodiment, the electronic device (100) can perform preprocessing on various unstructured data included in the first content information. For example, the electronic device (100) can remove unnecessary elements and refine the data. The electronic device (100) can identify the context of the content, use natural language processing techniques to select keywords that appear frequently or are deemed contextually important, and convert them into an appropriate form to generate hashtags.

[0132] In step S420, the electronic device (100) can generate a first embedding vector corresponding to the first hashtag based on the first hashtag and the first content. The first embedding vector may include a first hashtag vector corresponding to the first hashtag and a first content vector corresponding to the first content.

[0133] In one embodiment of the present disclosure, the electronic device (100) can generate a first embedding vector corresponding to a hashtag extracted from the first content. The electronic device (100) can generate a hashtag vector by converting the hashtag extracted in step S410 into a vector using an embedding generation model. The electronic device (100) can generate a content vector for the content using the same embedding generation model.

[0134] In one embodiment, a content vector may include information about the content and information about the use of the content. Information about the content refers to information about the content itself and may include, but is not limited to, the type of content, information about the service on which the content is provided, and the type of data included in the content. Information about the use of the content may refer to information regarding the interaction between the content and the user and may include, but is not limited to, information about the time the user used the content, the number of times the content was clicked, and information about the path through which the user entered the content.

[0135] In one embodiment, the electronic device (100) can convert information about content into a vector using an embedding generation model. The electronic device (100) can convert information about content usage into weights and multiply the converted vector element by element to generate a content vector.

[0136] In one embodiment, the electronic device (100) can generate a first embedding vector by performing operations on a hashtag vector and a content vector. For example, the electronic device (100) can generate a first embedding vector by concatenating the hashtag vector and the content vector.

[0137] In step S430, the electronic device (100) can determine a first cluster corresponding to a first embedding vector among a plurality of clusters. The electronic device (100) can perform clustering based on similarity between a plurality of vectors.

[0138] In one embodiment of the present disclosure, the electronic device (100) can determine a first cluster corresponding to the first embedding vector generated in step S420 among a plurality of clusters. The electronic device (100) can determine the cluster corresponding to the first embedding vector by calculating the distance between the first embedding vector and the center point vector corresponding to the center point of the plurality of clusters. For example, the electronic device (100) can cluster the 'a' embedding vector into cluster B when the distance between the center vector of a plurality of clusters (A, B, C) and the 'a' embedding vector is 0.2, 0.1, and 0.3, respectively.

[0139] In one embodiment, the electronic device (100) may determine that there is no cluster corresponding to the first embedding vector if the distance between the first embedding vector and the center point vector corresponding to the center point of a plurality of clusters is greater than or equal to a preset threshold value. The electronic device (100) may create a new cluster including the first embedding vector. For example, if the distance between the center vectors of a plurality of clusters (A, B, C) and the 'b' embedding vector is 0.4, 0.7, and 0.5, respectively, and the preset threshold value is 0.3, the electronic device (100) may determine that there is no cluster corresponding to the 'b' embedding vector and create a new cluster D including the 'b' embedding vector.

[0140] In step S440, the electronic device (100) can generate personalized classification information corresponding to the first hashtag based on the first embedding vector and the first center point vector corresponding to the center point of the first cluster. The personalized classification information may refer to dynamic classification information generated by computation to reflect user-specific interests regarding the first hashtag.

[0141] In one embodiment of the present disclosure, an electronic device (100) can generate personalized classification information corresponding to a first hashtag based on a first embedding vector and a center point vector of a first cluster containing the first embedding vector. The electronic device (100) can compute the first embedding vector and the first center point vector. The electronic device (100) can generate a second embedding vector by computed the hashtag vector included in the first embedding vector and the content vector included in the first center point vector. In one embodiment, the electronic device (100) can generate a second embedding vector by concatenating the first hashtag vector included in the first embedding vector and the content vector included in the first center point vector.

[0142] In one embodiment, the electronic device (100) can generate personalized classification information by decoding a second embedding vector. The electronic device (100) can generate personalized classification information by decoding a second embedding vector generated by computing a content vector of a first center point vector that reflects the user's interests through clustering with an existing hashtag, i.e., a first hashtag.

[0143] In step S450, the electronic device (100) can provide recommendation applications and recommendation control actions from personalized classification information using a neural network model. Since one application may include multiple control actions and one control action may be dependent on multiple applications, the electronic device (100) provides recommendation applications and recommendation control actions.

[0144] In one embodiment of the present disclosure, the neural network model may include a GNN and may process a graph in which a plurality of contents, a plurality of hashtags, a plurality of control actions, and a plurality of applications are each nodes. The neural network model (190) may be trained based on classification information included in the plurality of contents and control action history information corresponding to the plurality of classification information. Hereinafter, the training method and operation method of the neural network model refer to FIG. 3.

[0145] In one embodiment, the electronic device (100) can obtain a recommended control action based on similarity between personalized classification information and candidate control actions using a neural network model. Since one application may include multiple control actions and one electronic device (100) may include multiple applications, multiple candidate control actions may be included for each hashtag. The electronic device (100) can obtain a recommended control action based on similarity among candidate control actions for personalized classification information using a learned neural network model.

[0146] In one embodiment, the electronic device (100) can determine at least one candidate application corresponding to a recommended control action using an application-control action table. The application-control action table refers to a table that maps control actions that can be performed per application, as one application may include multiple control actions, and is described in detail below in FIG. 7. The electronic device (100) can obtain a recommended application based on the similarity between at least one candidate application determined through a neural network model and personalized classification information.

[0147] The electronic device (100) may provide acquired recommended control actions and recommended applications. In one embodiment, if a recommended application is installed within the electronic device (100), the electronic device (100) may provide an execution-related interface for the recommended application. If a recommended application is not installed within the electronic device (100), an installation-related interface for the recommended application may be provided.

[0148] In one embodiment, the electronic device (100) may obtain or load additional information required according to the recommendation control operation and recommendation application from another application. In one embodiment, the electronic device (100) may request additional information required according to the recommendation control operation and recommendation application from the user.

[0149] In one embodiment, the electronic device (100) can update a neural network model based on feedback information regarding a recommendation application and a recommendation control action. The electronic device (100) can receive input values ​​from a user regarding a recommendation application and a recommendation control action based on an interface provided to the user, and can update a neural network model using the input values ​​as feedback information.

[0150] In one embodiment of the present disclosure, as new content is added after step S450, the electronic device (100) can update the center point vector of the cluster and update personalized classification information. Hereinafter, a detailed description of the update operation of personalized classification information is provided with reference to FIG. 8.

[0151] FIG. 5 is a specific flowchart for generating personalized classification information for recommending control actions according to one embodiment.

[0152] Referring to FIG. 5, a specific flowchart is illustrated in which an electronic device (100) generates personalized classification information (590) from a first content (510). Hereinafter, content that overlaps with FIG. 4 is omitted for brevity.

[0153] The electronic device (100) can extract hashtags from content using an LLM (520). The electronic device (100) can extract hashtags from various forms of data included in the content. In one embodiment, the electronic device (100) can extract hashtags from text using an LLM. The electronic device (100) can extract hashtags from unstructured data such as images and audio using a multimodal-based LLM.

[0154] In one embodiment, a single piece of content can extract multiple hashtags. Hereinafter, for convenience of explanation, it is described that an electronic device (100) extracts a first hashtag from a first piece of content (510) using an LLM (520).

[0155] The electronic device (100) can generate a first embedding vector corresponding to a first hashtag using an embedding vector generation model (540). The electronic device (100) can generate a first embedding vector using the embedding vector generation model (540) using the first hashtag and a content vector (530). In one embodiment, the embedding vector generation model (540) can provide a token using a text encoder (550) and obtain a vector corresponding thereto.

[0156] The electronic device (100) can generate a hashtag vector by converting a first hashtag into a vector using an embedding generation model (540). The electronic device (100) can generate a content vector by using the embedding generation model (540) for information (530) about the content.

[0157] In one embodiment, a content vector may include information about the content and information about the use of the content. Information about the content refers to information about the content itself and may include, but is not limited to, the type of content, information about the service on which the content is provided, and the type of data included in the content. Information about the use of the content may refer to information regarding the interaction between the content and the user and may include, but is not limited to, information about the time the user used the content, the number of times the content was clicked, and information about the path through which the user entered the content.

[0158] In one embodiment, the electronic device (100) can generate a first embedding vector by performing operations on a hashtag vector and a content vector. For example, the electronic device (100) can generate the first embedding vector by concatenating the hashtag vector and the content vector using an embedding vector generation model (540).

[0159] The electronic device (100) can determine a first cluster corresponding to a first embedding vector using a clustering module (560). The clustering module (560) can determine a first cluster corresponding to a first embedding vector by calculating the distance between the first embedding vector and a center point vector corresponding to the center point of a plurality of clusters. The clustering module (560) can provide information regarding the first cluster obtained as a clustering result and the first embedding vector to a personalized classification information generator (570). In one embodiment, the information regarding the first cluster may include a center point vector of the first cluster. The center point vector of the first cluster may refer to a vector having the average value of each element, which is the center point of a plurality of vectors included in the first cluster.

[0160] The electronic device (100) can generate personalized classification information (590) using a personalized classification information generator (570). The personalized classification information generator (570) can generate personalized classification information (590) corresponding to a first hashtag based on a first embedding vector and a center point vector of a first cluster. The personalized classification information generator (570) can generate a second embedding vector by performing operations on the hashtag vector included in the first embedding vector and the content vector included in the first center point vector. The personalized classification information generator (570) can generate personalized classification information (590) by decoding the second embedding vector using a text decoder (580).

[0161] FIG. 6 is a diagram illustrating a method for performing clustering to generate personalized classification information according to one embodiment.

[0162] Referring to Fig. 6, a method of performing clustering on embedding vectors to generate personalized classification information is illustrated.

[0163] In one embodiment of the present disclosure, an electronic device (100) may extract hashtags from content and generate corresponding embedding vectors. A specific method for generating embedding vectors is omitted for the sake of brevity in the specification. The electronic device (100) may perform clustering to generate personalized classification information for the generated embedding vectors.

[0164] In one embodiment, the electronic device (100) can generate a first embedding vector corresponding to a first hashtag extracted from the first content. The electronic device (100) can perform clustering on the first embedding vector. The electronic device (100) can determine the cluster corresponding to the first embedding vector by calculating the distance between the first embedding vector and a center point vector (630a, 630b, 630c) corresponding to the center point of a plurality of clusters (620a, 620b, 620c).

[0165] The electronic device (100) can calculate the distance between the first embedding vector and the center point vectors (630a, 630b, 630c) of a plurality of clusters. In one embodiment, when the first embedding vector is at position (a) (610a), the first embedding vector is closest to the center point vector (630a) of the first cluster (620a), so the electronic device (100) can cluster the first embedding vector into the first cluster (620a).

[0166] In one embodiment, when the first embedding vector is at position (b) (610b), the electronic device (100) may determine that the distance between the first embedding vector and the center point vectors (630a, 630b, 630c) of a plurality of clusters is greater than or equal to a preset threshold value. For example, with respect to the first embedding vector located at 'b', the distances to the center vectors (630a, 630b, 630c) of a plurality of clusters (620a, 620b, 620c) are 0.4, 0.5, and 0.7, respectively, and the preset threshold value is 0.3, the electronic device (100) may determine that there is no cluster corresponding to the first embedding vector. The electronic device (100) may create a new cluster including the first embedding vector.

[0167] The electronic device (100) can generate personalized classification information for the first hashtag based on the clustering result for the first embedding vector. For specific methods below, refer to FIGS. 1 to 5.

[0168] FIG. 7 is an example diagram of an application-control action table for recommending control actions according to one embodiment.

[0169] Referring to FIG. 7, the electronic device (100) may use an application-control action table to recommend control actions. The electronic device (100) may use an application-control action table (700) to provide recommended applications and recommended control actions from personalized classification information.

[0170] In one embodiment of the present disclosure, an electronic device (100) may provide recommended applications and recommended control actions from personalized classification information using a neural network model. In one embodiment, the electronic device (100) may obtain recommended control actions based on similarity between personalized classification information and candidate control actions. The electronic device (100) may determine at least one candidate application corresponding to a recommended control action using an application-control action table (700). The electronic device (100) may obtain recommended applications based on similarity between at least one candidate application and personalized classification information. Hereinafter, a method for controlling recommended applications and recommended control actions refers to the description of FIG. 9.

[0171] In one embodiment of the present disclosure, one application may include a plurality of control operations. Alternatively, one control operation may be dependent on a plurality of applications. An electronic device (100) may generate an application-control operation table (700) representing a mapping relationship between an application and a control operation. As one application may include a plurality of control operations, the electronic device (100) may generate an application-control operation table (700) representing control operations that can be performed per application.

[0172] For example, the application-control action table (700) can map control actions such as making a call, answering a call, viewing call history, adding contacts, and checking voicemail to the phone application. The application-control action table (700) can map control actions such as searching for products, viewing product details, adding to cart, placing an order, making a payment, tracking delivery, and writing a review to the shopping application. In addition, the application-control action table (700) may include control actions mapped to each of the following applications: contacts, camera, settings, Samsung Health, social media, messenger, banking / finance, music streaming, travel and transportation, education, games, and productivity applications.

[0173] In one embodiment, a single control action may be included in a plurality of applications. For example, a control action corresponding to "adding a friend" may be included in a social media application and a game application, and a control action corresponding to "writing a review" may be included in a shopping application, an educational application, a game application, etc.

[0174] In one embodiment, the electronic device (100) may obtain a recommended control action based on similarity between personalized classification information and candidate control actions, and then determine at least one candidate application corresponding to the recommended control action using an application-control action table (700). For example, if the electronic device (100) obtains a recommended control action called "Write Review," it may determine a shopping application, an education application, and a game application as candidate applications using the application-control action table (700). The electronic device (100) may obtain a recommended application based on similarity between the candidate application and the personalized classification information. For a specific method of providing a recommended control action and a recommended application, refer to the descriptions in FIGS. 3, 9, and 10.

[0175] FIG. 8 is a flowchart illustrating the operation of updating personalized classification information according to one embodiment.

[0176] FIG. 8 illustrates the operation of updating personalized classification information after step S450 of FIG. 4. In one embodiment, the personalized classification information may be continuously updated to reflect the user's interests as it has dynamic characteristics.

[0177] In step S810, the electronic device (100) can generate a third embedding vector. The third embedding vector may be included in the same first cluster as the first embedding vector.

[0178] In one embodiment of the present disclosure, the electronic device (100) can generate a third embedding vector. In one embodiment, the third embedding vector can be generated based on a second hashtag extracted from the first content and can be generated based on a third hashtag extracted from the second content. Since the third embedding vector has the highest similarity value with the center point vector of the first cluster among a plurality of clusters, it can be clustered into the first cluster.

[0179] In one embodiment, the electronic device (100) can generate personalized classification information through an operation between the third embedding vector and the center point vector of the first cluster. Since the electronic device (100) can generate personalized classification information by performing the same procedure as for the first embedding vector on the third embedding vector, a detailed description is omitted.

[0180] In step S820, the electronic device (100) can update the first center point vector corresponding to the center point of the first cluster based on the third embedding vector. The electronic device (100) can update the first center point vector of the first cluster as the new third embedding vector is included in the first cluster.

[0181] A center vector of a cluster is a vector indicating the average point of multiple vectors included in the cluster, and the center vector can be updated as a new vector is added to the cluster. For example, the existing first cluster contains 9 vectors, and the first center vector, which is the center point of the first cluster, can have values ​​of (0.3, 0.5, ...). If the third embedding vector has values ​​of (0.4, 0.2, ...), the electronic device (100) can update the first center vector to add the characteristics of the third embedding vector included in the first cluster. The updated first center vector can have values ​​of ((0.3*9+0.4) / 10, (0.5*9+0.2) / 10, ...).

[0182] In step S830, the electronic device (100) can update personalized classification information corresponding to the first hashtag based on the updated first center point vector. To dynamically reflect the user's interests in the personalized classification information, the electronic device (100) can continuously update the personalized classification information.

[0183] In one embodiment of the present disclosure, an electronic device (100) may update a second embedding vector used to generate personalized classification information using an updated first center point vector. The second embedding vector is generated through the operation of a hashtag vector of the first embedding vector and a content vector included in the first center point vector of the first cluster. As the first center point vector is updated, the second embedding vector may be updated by operating the hashtag vector of the first embedding vector and the content vector included in the updated first center point vector. The electronic device (100) may update personalized classification information by decoding the updated second embedding vector.

[0184] FIG. 9 is a flowchart illustrating a method for recommending control operations and applications according to one embodiment.

[0185] FIG. 9 is a flowchart illustrating the operation of step S450, which provides the recommended application and recommendation control operation of FIG. 4. Since the electronic device (100) has multiple applications and can perform multiple control operations within the electronic device (100), it can determine and provide the recommended application and recommendation control operation for personalized classification information to the user.

[0186] In step S910, the electronic device (100) can obtain a recommended control action based on the similarity between the personalized classification information and the candidate control action. The electronic device (100) can calculate the similarity between the personalized classification information and the candidate control action based on a learned neural network model.

[0187] In one embodiment of the present disclosure, an electronic device (100) can use a neural network model to recognize personalized classification information as a node of the classification information unit and obtain a recommended control action based on similarity with a plurality of candidate control actions. The neural network model can perform learning based on a plurality of classification information and control action history information corresponding to the plurality of classification information. The electronic device (100) can infer a recommended control action by calculating the similarity between personalized classification information and candidate control actions using the learned neural network model. Hereinafter, specific learning and inference methods refer to FIG. 3.

[0188] In step S920, the electronic device (100) can obtain a recommended application based on the similarity between at least one candidate application corresponding to a recommended control action and personalized classification information. In one embodiment, the electronic device (100) can determine at least one candidate application corresponding to a recommended control action using an application-control action table.

[0189] In one embodiment of the present disclosure, one application can perform a plurality of control operations, and as one control operation is dependent on a plurality of applications, the electronic device (100) can generate an application-control operation table that maps control operations that can be performed per application. Refer to FIG. 7 for a specific example of an application-control operation table.

[0190] In one embodiment, the electronic device (100) can determine at least one candidate application corresponding to a recommended control action based on an application-control action table. For example, if the electronic device (100) obtains a recommended control action of "payment" for a personalized search result, it can determine a shopping application, an education application, a game application, etc., as a candidate application using the application-control action table.

[0191] In one embodiment, the electronic device (100) may obtain a recommended application based on similarity between a candidate application and personalized classification information. The electronic device (100) may provide the obtained recommendation control action and the recommended application to the user. In one embodiment, if the recommended application is installed within the electronic device (100), the electronic device (100) may provide an execution-related interface for the recommended application. If the recommended application is not installed within the electronic device (100), the electronic device (100) may provide an installation-related interface for the recommended application.

[0192] FIG. 10 is a schematic block diagram of an electronic device according to one embodiment.

[0193] Referring to FIG. 10, the electronic device (100) may include memory (1010), a processor (1020), and a communication unit (1030). However, not all of the illustrated components are essential components of the electronic device (100). The electronic device (100) may be implemented by more components than those illustrated, or by fewer components.

[0194] The memory (1010) can store a program for processing and controlling the processor (1020) and can store data that is input to or output from the electronic device (100). The memory (1010) may include at least one type of storage medium among flash memory type, hard disk type, multimedia card micro type, card type memory (e.g., SD or XD memory), RAM (Random Access Memory), SRAM (Static Random Access Memory), ROM (Read-Only Memory), EEPROM (Electrically Erasable Programmable Read-Only Memory), PROM (Programmable Read-Only Memory), magnetic memory, magnetic disk, and optical disk. In one embodiment, the memory (1010) may include a weight memory for storing weights and a feature vector memory for including feature vectors which are intermediate output values ​​of a neural network model, but is not limited thereto.

[0195] In one embodiment of the present disclosure, the memory (1010) may store information related to content obtained by the operation of the processor (1020). The memory (1010) may store hashtags extracted from the content.

[0196] In one embodiment, the memory (1010) may store a hashtag vector corresponding to a hashtag, a content vector, and an embedding vector generated by computation thereof to generate personalized classification information. The memory (1010) may store information regarding a cluster corresponding to an embedding vector obtained as a clustering result for the embedding vector. The memory (1010) may store personalized information corresponding to a hashtag.

[0197] In one embodiment, the memory (1010) may store an application-control action table to train a neural network model. The memory (1010) may store data related to the training of the neural network model. The memory (1010) may store information regarding recommendation control actions and recommendation applications for personalized classification information.

[0198] The processor (1020) controls the overall operation of the electronic device (100). For example, the processor (1020) can perform the functions of the electronic device (100) described in the present disclosure by executing one or more instructions stored in memory (1010). In this case, memory (1010) may store one or more instructions that can be executed by the processor (1020). Additionally, the processor (1020) can store one or more instructions in internally provided memory and control the execution of the one or more instructions stored in the internally provided memory so that the aforementioned operations are performed. That is, the processor (1020) can perform a predetermined operation by executing at least one instruction or program stored in internal memory or memory (1010) provided within the processor (1020).

[0199] The processor (1020) may include one or more processors. In this case, the one or more processors may be general-purpose processors such as CPUs, APs, DSPs (Digital Signal Processors), graphics-dedicated processors such as GPUs, VPUs (Vision Processing Units), or artificial intelligence-dedicated processors such as NPUs. If one or more processors are artificial intelligence-dedicated processors, the artificial intelligence-dedicated processors may be designed with a hardware structure specialized for processing a specific artificial intelligence model.

[0200] In one embodiment of the present disclosure, the processor (1020) can obtain content through the communication unit (1030). The processor (1020) can extract hashtags from the content. The processor (1020) can generate an embedding vector corresponding to the hashtags. The processor (1020) can determine a cluster corresponding to the embedding vector and generate personalized classification information based on the embedding vector and the cluster.

[0201] In one embodiment, the processor (1020) may provide recommended applications and recommended control actions from personalized classification information using a neural network model. The processor (1020) may train the neural network model based on classification information and control action history information corresponding to the classification information. The processor (1020) may obtain recommended control actions based on similarity between personalized classification information and candidate control actions. The processor (1020) may obtain recommended applications based on similarity between at least one candidate application corresponding to the recommended action and personalized classification information. The processor (1020) may determine candidate applications corresponding to the recommended action based on an application-control action table.

[0202] In one embodiment, the processor (1020) may obtain additional information required according to the recommendation control action and the recommendation application from another application or request it from the user. In one embodiment, the processor (1020) may update a neural network model based on feedback information regarding the application and the recommendation control action.

[0203] In one embodiment, the processor (1020) can update the center point vector of a cluster as it generates a new embedding vector. The processor (1020) can update personalized classification information corresponding to the embedding vector included in the cluster as it updates the center point vector of the cluster.

[0204] The communication unit (1030) may include one or more modules that enable wireless communication between the electronic device (100) and a network where an external device (not shown) is located. The communication unit (1030) may transmit or receive data or signals to and from an external device (not shown) through a wired or wireless network. A communication unit (1030) according to one embodiment of the present disclosure includes at least one communication module, such as a short-range communication module, a wired communication module, a mobile communication module, a broadcast reception module, etc. Here, the at least one communication module refers to a communication module capable of transmitting and receiving data through a network that follows a communication standard such as a tuner that performs broadcast reception, Bluetooth, WLAN (Wireless LAN) (Wi-Fi), Wibro (Wireless broadband), Wimax (World Interoperability for Microwave Access), CDMA, WCDMA, etc.

[0205] For example, the communication unit (1030) may include a Wi-Fi module, a Bluetooth module, an infrared communication module and a wireless communication module, a LAN module, an Ethernet module, a wired communication module, etc. At this time, each communication module may be implemented in the form of at least one hardware chip. The Wi-Fi module and the Bluetooth module perform communication in the Wi-Fi method and the Bluetooth method, respectively. When using the Wi-Fi module or the Bluetooth module, various connection information such as SSID and session key is first transmitted and received, and after establishing a communication connection using this, various information can be transmitted and received. The wireless communication module may include at least one communication chip that performs communication according to various wireless communication standards such as Zigbee, 3G (3rd Generation), 3GPP (3rd Generation Partnership Project), LTE (Long Term Evolution), LTE-A (LTE Advanced), 4G (4th Generation), 5G (5th Generation), etc. The communication unit (1030) may include a communication unit that performs communication with electronic devices and Bluetooth, and an interface unit that connects to an external device.

[0206] In one embodiment of the present disclosure, the communication unit (1030) may receive information regarding content. The communication unit (1030) may transmit and display an interface regarding a recommendation control operation and a recommendation application. For example, the communication unit (1030) may transmit and display an execution-related interface for a recommendation application. The communication unit (1030) may transmit and display an installation-related interface for a recommendation application.

[0207] A device-readable storage medium may be provided in the form of a non-transitory storage medium. Here, 'non-transitory storage medium' simply means that it is a tangible device and does not contain a signal (e.g., electromagnetic waves), and the term does not distinguish between cases where data is stored semi-permanently and cases where it is stored temporarily. For example, a 'non-transitory storage medium' may include a buffer in which data is stored temporarily.

[0208] According to one embodiment, the method according to the various embodiments disclosed herein may be provided by being included in a computer program product. The computer program product may be traded between a seller and a buyer as a product. The computer program product may be distributed in the form of a device-readable storage medium (e.g., compact disc read-only memory (CD-ROM)), or distributed online (e.g., download or upload) through an application store or directly between two user devices (e.g., smartphones). In the case of online distribution, at least a portion of the computer program product (e.g., downloadable app) may be temporarily stored or temporarily created on a device-readable storage medium, such as the memory of a manufacturer's server, an application store's server, or a relay server.

[0209] According to one embodiment of the present disclosure, a method is provided for an electronic device to recommend a control action. The method may extract a first hashtag corresponding to a first content. Based on the first hashtag and the first content, the method may generate a first embedding vector corresponding to the first hashtag. The method may determine a first cluster corresponding to the first embedding vector among a plurality of clusters. Based on the first embedding vector and a first centroid vector corresponding to the centroid of the first cluster, the method may generate personalized classification information corresponding to the first hashtag. The method may provide a recommended application and a recommended control action from the personalized classification information using a neural network model. The first embedding vector may include a first hashtag vector corresponding to the first hashtag and a first content vector corresponding to the first content.

[0210] In one embodiment, the method can generate a second embedding vector by combining a first hashtag vector and a content vector included in a first center point vector. The method can generate personalized classification information by decoding the second embedding vector.

[0211] In one embodiment, the first content vector may include information about the first content and information about the use of the first content.

[0212] In one embodiment, the method can generate a second embedding vector included in a first cluster. The method can update a first center point vector corresponding to the center point of the first cluster based on the second embedding vector. The method can update personalized classification information corresponding to a first hashtag based on the updated first center point vector.

[0213] In one embodiment, the method can obtain a recommended control action based on similarity between personalized classification information and a candidate control action. The method can obtain a recommended application based on similarity between at least one candidate application corresponding to the recommended control action and personalized classification information.

[0214] In one embodiment, at least one candidate application can be determined based on candidate control actions using an application-control action table.

[0215] In one embodiment, a neural network model can be trained based on a plurality of classification information corresponding to a plurality of contents and control operation history information corresponding to the plurality of classification information.

[0216] In one embodiment, the neural network model can be updated based on feedback information regarding the recommendation application and recommendation control action.

[0217] In one embodiment, the method may provide an execution-related interface for a recommendation application when the recommendation application is installed within an electronic device. The method may provide an installation-related interface for a recommendation application when the recommendation application is not installed within an electronic device.

[0218] In one embodiment, the neural network model may be a GNN (Graph Neural Network) model.

[0219] According to one embodiment of the present disclosure, an electronic device for recommending a control operation may include a memory for storing at least one instruction and at least one processor for executing at least one instruction stored in the memory. At least one processor may extract a first hashtag corresponding to a first content. At least one processor may generate a first embedding vector corresponding to the first hashtag based on the first hashtag and the first content. At least one processor may determine a first cluster corresponding to the first embedding vector among a plurality of clusters. At least one processor may generate personalized classification information corresponding to the first hashtag based on the first embedding vector and a first centroid vector corresponding to the centroid of the first cluster. At least one processor may provide a recommendation application and a recommendation control operation from the personalized classification information using a neural network model. The first embedding vector may include a first hashtag vector corresponding to the first hashtag and a first content vector corresponding to the first content.

[0220] According to one embodiment of the present disclosure, a computer-readable recording medium may be provided that records a program for executing the method on a computer.

Claims

1. In a method for an electronic device to recommend a control action, A step of extracting a first hashtag corresponding to the first content; A step of generating a first embedding vector corresponding to the first hashtag based on the first hashtag and the first content; A step of determining a first cluster corresponding to the first embedding vector among a plurality of clusters; A step of generating personalized classification information corresponding to the first hashtag based on the first embedding vector and the first center point vector corresponding to the center point of the first cluster; and A method for providing a recommendation application and a recommendation control operation from the personalized classification information using a neural network model; comprising, A method in which the first embedding vector comprises a first hashtag vector corresponding to the first hashtag and a first content vector corresponding to the first content.

2. In Paragraph 1, The step of generating the above personalized classification information is, A step of generating a second embedding vector by performing an operation on the content vector included in the first hashtag vector and the first center point vector; and A method characterized by including the step of decoding the second embedding vector to generate the personalized classification information.

3. In either Paragraph 1 or Paragraph 2, A method characterized in that the first content vector comprises information about the first content and information about the use of the first content.

4. In any one of paragraphs 1 to 3, A step of generating a third embedding vector, wherein the third embedding vector is included in the first cluster; A step of updating the first center point vector corresponding to the center point of the first cluster based on the third embedding vector; and A method characterized by including the step of updating personalized classification information corresponding to the first hashtag based on the updated first center point vector.

5. In any one of paragraphs 1 through 4, The step of providing the above-mentioned recommendation application and recommendation control operation is, A step of obtaining the recommended control action based on the similarity between the personalized classification information and the candidate control action; and The method includes the step of obtaining the recommended application based on the similarity between at least one candidate application corresponding to the recommendation control operation and the personalized classification information; A method characterized in that at least one candidate application is determined based on the candidate control action using an application-control action table.

6. In any one of paragraphs 1 through 5, A method characterized in that the above neural network model is learned based on a plurality of classification information corresponding to a plurality of contents and control operation history information corresponding to the plurality of classification information.

7. In any one of paragraphs 1 through 6, A method characterized in that the neural network model is updated based on feedback information regarding the recommendation application and the recommendation control operation.

8. In any one of paragraphs 1 through 7, The step of providing the above-mentioned recommendation application and recommendation control operation is, When the above recommendation application is installed within the electronic device, the step of providing an execution-related interface for the recommendation application; and A method characterized by including the step of providing an installation-related interface for the recommended application when the recommended application is not installed within the electronic device.

9. In an electronic device (100) for recommending a control operation, Memory (1010) for storing at least one instruction; and It includes at least one processor (1020) that executes at least one instruction stored in the memory, The above at least one processor (1020) executes the above at least one instruction, Extract the first hashtag corresponding to the first content, and Based on the first hashtag and the first content, a first embedding vector corresponding to the first hashtag is generated, and Among a plurality of clusters, determine the first cluster corresponding to the first embedding vector, and Based on the first embedding vector and the first center point vector corresponding to the center point of the first cluster, personalized classification information corresponding to the first hashtag is generated, and Using a neural network model, a recommendation application and a recommendation control operation are provided from the above personalized classification information, An electronic device wherein the first embedding vector comprises a first hashtag vector corresponding to the first hashtag and a first content vector corresponding to the first content.

10. In Paragraph 9, The above at least one processor (1020) executes the above at least one instruction, A second embedding vector is generated by combining the content vector included in the first hashtag vector and the first center point vector, and An electronic device that decodes the second embedding vector to generate the personalized classification information.

11. In either Paragraph 9 or Paragraph 10, An electronic device characterized in that the first content vector includes information about the first content and information about the use of the first content.

12. In any one of paragraphs 9 through 11, The above at least one processor (1020) executes the above at least one instruction, A third embedding vector is generated, wherein the third embedding vector is included in the first cluster, and Based on the above third embedding vector, the first center point vector corresponding to the center point of the first cluster is updated, and An electronic device that updates personalized classification information corresponding to the first hashtag based on the above-mentioned updated first center point vector.

13. In any one of paragraphs 9 through 12, The above at least one processor (1020) executes the above at least one instruction, Based on the similarity between the above personalized classification information and candidate control actions, the above recommended control action is obtained, and Based on the similarity between at least one candidate application corresponding to the above recommendation control operation and the above personalized classification information, the recommendation application is obtained, and An electronic device characterized in that at least one candidate application is determined based on the candidate control action using an application-control action table.

14. In any one of paragraphs 9 through 13, An electronic device characterized in that the above neural network model is learned based on a plurality of classification information corresponding to a plurality of contents and control operation history information corresponding to the plurality of classification information.

15. In any one of paragraphs 9 through 14, An electronic device characterized in that the neural network model is updated based on feedback information regarding the recommendation application and the recommendation control operation.

Images