Recommendation method and device and electronic equipment

Abstract

The invention provides a recommendation method and device and electronic equipment, and the method comprises the steps: determining a first-class word matrix based on first-class information in user historical behavior data and first-class information of the electronic equipment; determining a second class word matrix based on the second class information in the user historical behavior data and the second class information of the electronic device; determining an equipment information matrix based on the first type of information of the electronic equipment and the second type of information of the electronic equipment; determining a fusion matrix based on the first type of word matrix, the second type of word matrix and the equipment information matrix; inputting question data input by a user and the fusion matrix into a recommendation model to obtain recommendation information for the performance of the electronic equipment; wherein the first type of information and the second type of information represent different attribute information of the electronic equipment.

Description

Technical Field

[0001] The present disclosure relates to the field of large model technology, and in particular to a recommendation method, device, and electronic device. Background Art

[0002] Existing recommendation systems typically rely on linked recommendation methods within big data frameworks when processing structured data. However, with the application of large models, recommendations are often stored and computed in the form of lists. However, when users express their personalized needs through conversations, these systems can only provide a single-dimensional list of recommendations in conversational recommendation scenarios, making it difficult to dynamically link recommendations based on context. Furthermore, these systems rely on mapping user preferences to product ID sequences, which can lead to insufficient or duplicate recommendations for new users or those with diverse needs. These issues limit the intelligence and user experience of recommendation systems, necessitating a more efficient solution. Summary of the Invention

[0003] The present disclosure provides a recommendation method, device, and electronic device to at least solve the above technical problems existing in the prior art.

[0004] According to a first aspect of the present disclosure, a recommendation method is provided, comprising:

[0005] Determining a first word matrix based on the first category of information in the user's historical behavior data and the first category of information of the electronic device;

[0006] determining a second word matrix based on the second type of information in the user's historical behavior data and the second type of information of the electronic device;

[0007] determining a device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device;

[0008] Determine a fusion matrix based on the first category word matrix, the second category word matrix and the device information matrix;

[0009] Inputting the user inputted question data and the fusion matrix into the recommendation model to obtain recommendation information for the performance of the electronic device;

[0010] The first type of information and the second type of information represent different attribute information of the electronic device.

[0011] In the above solution, determining the first-category word matrix based on the first-category information in the user's historical behavior data and the first-category information of the electronic device includes:

[0012] Obtain the first type of information about user tendencies based on user historical behavior data;

[0013] Determine a first-category word embedding matrix based on the first-category information of the user preference and the first-category information of the electronic device;

[0014] Determining a position code for each vector in the first class word embedding matrix, and determining a first class input embedding matrix based on the position code for each vector and the first class word embedding matrix;

[0015] The first-category input embedding matrix is input into the transformer to obtain a first-category word matrix.

[0016] In the above solution, the transformer includes at least one layer of self-attention mechanism and feedforward network, and the first-class input embedding matrix is input into the transformer to obtain the first-class word matrix, including:

[0017] Input the first-class input embedding matrix into any self-attention mechanism and feedforward network to obtain the first-class context vector;

[0018] An average pooling operation is performed on the first-category context vector to obtain the first-category word matrix.

[0019] In the above solution, determining the device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device includes:

[0020] Acquire first category information of the electronic device; the first category information includes at least one of first category installation information, operation information, and update information of the electronic device;

[0021] obtaining second type of information of the electronic device;

[0022] determining a device information embedding matrix based on the first type of information and the second type of information;

[0023] An average pooling operation is performed on the device information embedding matrix to obtain the device information matrix.

[0024] In the above solution, determining the fusion matrix based on the first category word matrix, the second category word matrix and the device information matrix includes:

[0025] The first category word matrix, the second category word matrix and the device information matrix are spliced based on a first order, and the first category word matrix, the second category word matrix and the device information matrix are weighted based on a first weight, a second weight and a third weight to obtain the fusion matrix.

[0026] In the above solution, the method further includes:

[0027] Determining the first weight, the second weight, and the third weight based on the training data specifically includes:

[0028] Obtain the first-category word test matrix, the second-category word test matrix, and the device information test matrix corresponding to each test user;

[0029] The first category word test matrix, the second category word test matrix and the device information test matrix are weighted based on the first test weight, the second test weight and the third test weight respectively, and the first test weight is adjusted based on the second category tendency of each test user so that the result of the weighted processing is consistent with the second category tendency.

[0030] In the above solution, the problem data input by the user and the fusion matrix are input into the recommendation model to obtain recommendation information for the performance of the electronic device, including:

[0031] Inputting the fusion matrix into a recommendation model to obtain a recommendation result output by the recommendation model;

[0032] Determining the user's tendency information based on the question data input by the user;

[0033] It is determined that the recommendation results that are consistent with the user's preference information are recommendation information for the electronic device performance.

[0034] In the above solution, the method further includes training a recommendation model, specifically including:

[0035] Determining a weighted recommendation loss function based on the fourth weight, the fifth weight, a positive fusion test matrix corresponding to the positive sample, and a negative fusion test matrix corresponding to the negative sample;

[0036] Determine a difficult recommendation loss function based on the sixth weight, the difficult fusion test matrix corresponding to the difficult sample, and the easy fusion matrix corresponding to the easy sample;

[0037] Determining a total loss function of the recommendation model based on the weighted recommendation loss function and the difficult recommendation loss function;

[0038] Determine the loss function gradient of each round, and update the weight of the recommendation model based on the loss function gradient;

[0039] The fourth weight is greater than the fifth weight.

[0040] According to a second aspect of the present disclosure, a recommendation device is provided, comprising:

[0041] a first processing unit, configured to determine a first-category word matrix based on the first-category information in the user's historical behavior data and the first-category information of the electronic device;

[0042] a second processing unit, configured to determine a second-category word matrix based on the second-category information in the user's historical behavior data and the second-category information of the electronic device;

[0043] a third processing unit, configured to determine a device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device;

[0044] a fusion unit, configured to determine a fusion matrix based on the first-category word matrix, the second-category word matrix, and the device information matrix;

[0045] The recommendation unit is configured to input the question data input by the user and the fusion matrix into a recommendation model to obtain recommendation information for the performance of the electronic device.

[0046] According to a third aspect of the present disclosure, there is provided an electronic device, including:

[0047] At least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method described in the present disclosure.

[0048] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present disclosure, nor are they intended to limit the scope of the present disclosure. Other features of the present disclosure will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] The above and other objects, features and advantages of the exemplary embodiments of the present disclosure will become readily understood by reading the detailed description below with reference to the accompanying drawings, in which several embodiments of the present disclosure are shown by way of example and not limitation, wherein:

[0050] In the drawings, the same or corresponding reference numerals denote the same or corresponding parts.

[0051] Figure 1 A first optional flow chart of the recommended method provided by an embodiment of the present disclosure is shown;

[0052] Figure 2 A second optional flow chart of the recommended method provided by the embodiment of the present disclosure is shown;

[0053] Figure 3 A third optional flow chart of the recommended method provided by the embodiment of the present disclosure is shown;

[0054] Figure 4 A fourth optional flow chart of the recommended method provided by an embodiment of the present disclosure is shown;

[0055] Figure 5 A fifth optional flow chart of the recommended method provided in the embodiment of the present disclosure is shown;

[0056] Figure 6 A schematic flow chart of a recommendation method provided by an embodiment of the present disclosure is shown;

[0057] Figure 7 A schematic diagram illustrating the processing of a recommendation model provided by an embodiment of the present disclosure is shown;

[0058] Figure 8 shows an optional structural diagram of a recommendation device provided by an embodiment of the present disclosure;

[0059] Figure 9 A schematic diagram of the structure of an electronic device according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0060] To make the purposes, features, and advantages of the present disclosure more apparent and understandable, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present disclosure, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative work shall fall within the scope of protection of the present disclosure.

[0061] In the following description, reference is made to “some embodiments”, which describes a subset of all possible embodiments, but it will be understood that “some embodiments” may be the same subset or different subsets of all possible embodiments and may be combined with each other without conflict.

[0062] In the following description, the terms "first\second" are merely used to distinguish similar objects and do not represent a specific ordering of the objects. It is understandable that "first\second" can be interchanged with a specific order or sequence where permitted, so that the embodiments of the present disclosure described herein can be implemented in an order other than that illustrated or described herein.

[0063] Unless otherwise defined, all technical and scientific terms used in this disclosure have the same meaning as commonly understood by those skilled in the art in the art of this disclosure. The terms used in this disclosure are only for the purpose of describing the embodiments of this disclosure and are not intended to limit this disclosure.

[0064] It should be understood that in the various embodiments of the present disclosure, the size of the serial number of each implementation process does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure.

[0065] Existing recommendation systems, with their large amounts of structured data, can only perform linked recommendations within the big data recommendation framework. With the advent of big models, because recommendations are stored and calculated in the form of sequences, there is a lack of interpretable associations between recommendations, user attributes, and user interests. When users seek recommendations based on their specific needs through conversations, the following situations may occur:

[0066] 1) When making recommendations based on a conversational model for service users, only a single-dimensional recommendation list is provided, lacking relevance to the conversation.

[0067] 2) Relying on the product ID sequence of user interactions to map user preferences, when encountering new users who have not appeared before or the demand for multiple recommended items, problems such as insufficient recommendations and duplicate recommendations may occur.

[0068] 3) Based on enterprise users, there is a lack of recommendation basis based on user organization, work attributes, and complex relational interests. Often after multiple rounds of interaction, there is a lack of personalized recommendation content.

[0069] In response to the defects existing in the relevant technologies, the embodiments of the present disclosure provide a recommendation method. By modeling the association between users, hardware and software, it solves the problem of multiple rounds of interactive recommendations to users based on software and hardware, and introduces richer semantic signals in the representation of recommended items and users compared with the previous representation based on system structured data. Specifically, the method includes: multi-dimensional modeling of the interest relationship between users, hardware and software according to the matching degree of semantic information; construction of association analysis models between users and hardware, users and software, and software and hardware; and utilization of the understanding ability and rich knowledge system contained in the Large Language Model (LLM) to give the recommendation results more interpretability (such as enhanced interactivity and interpretability) to solve some or all of the above technical problems.

[0070] Figure 1 A first optional flow chart of the recommended method provided by the embodiment of the present disclosure is shown, and will be explained according to each part.

[0071] Step S101 : determining a first-category word matrix based on first-category information in the user's historical behavior data and first-category information of the electronic device.

[0072] In some embodiments, the user's historical behavior data may include behavior information generated during the use of the electronic device, such as at least one of click information, reading information, tag information, function information, and attribute information. The click information includes clicks on different types of information, such as the number of clicks, click rankings of different types of information, etc.; reading information includes reading information on different types of information pages, such as reading time and number of reads; tag information includes education, income, marital, and family information; tendency information includes the user's main use of the electronic device, such as programming, drawing, office work, or entertainment, determined based on the user's use of various applications in the electronic device during the user's use of the electronic device; and the attribute information includes the user's organizational relationship, superior-subordinate relationship, and the user's occupation. The different types of information include software type information and hardware type information.

[0073] In specific implementation, the first category of information in the user's historical behavior data may include information related to the software, such as at least one of click information on each software information, reading information and function information on each software page, etc.; optionally, in order to make the recommendation results more accurate, the first category of information may also include tag information and attribute information.

[0074] In some embodiments, the first category of information of the electronic device may be software information in the electronic device; such as at least one of the installation time, running time, running frequency, number of runs and time of the most recent run of each software; optionally, the software information may also include installation information of each software in the electronic device, such as the number of installations of each type of software.

[0075] In some embodiments, a carrier implementing the recommendation method (hereinafter referred to as the carrier) may determine a first-category word matrix based on the first-category information in the user's historical behavior data and the keywords in the first-category information of the electronic device; and may also determine the first-category word matrix based on the first-category information in the user's historical behavior data and the feature words in the first-category information of the electronic device. The feature words are in a feature space.

[0076] In some embodiments, the carrier can be a computer program, electronic circuit, database, mobile application, electronic device, cloud computing platform, distributed system, artificial intelligence framework, mathematical model, automation tool and microcontroller, etc., which can implement software or hardware of algorithm and method process.

[0077] Step S102: determining a second-category word matrix based on the second-category information in the user's historical behavior data and the second-category information of the electronic device.

[0078] In some embodiments, the first and second categories of information represent different attribute information of electronic devices. The second category of information in the user's historical behavior data may include hardware-related information, such as click information on various hardware information, hardware information query information, and function information. The hardware information query information includes information about the hardware operating status viewed through the electronic device's management page, including the number of queries, query objects, and query frequency. Optionally, to make recommendation results more accurate, the second category of information may also include tag information and attribute information.

[0079] In some embodiments, the second category of information of the electronic device includes at least one of brand information, model information, core number information, capacity information, type information, and frequency information of each hardware in the electronic device;

[0080] In some embodiments, the carrier may determine a second-category word matrix based on the second-category information in the user's historical behavior data and keywords in the second-category information of the electronic device; and may also determine a first-category word matrix based on the second-category information in the user's historical behavior data and feature words in the second-category information of the electronic device. The feature words are in a feature space.

[0081] Step S103: determining a device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device.

[0082] In some embodiments, the carrier determines the device information matrix based on keywords of the first category information and the second category information of the electronic device; or determines the device information matrix based on feature words of the first category information and the second category information of the electronic device.

[0083] Step S104: determining a fusion matrix based on the first-category word matrix, the second-category word matrix, and the device information matrix.

[0084] In some embodiments, the carrier may concatenate the first-category word matrix, the second-category word matrix, and the device information matrix to obtain the fusion matrix.

[0085] In other embodiments, the carrier may concatenate the first-category word matrix, the second-category word matrix, and the device information matrix based on different weights to obtain a fusion matrix.

[0086] In some further embodiments, the carrier weights the first category word matrix, the second category word matrix and the device information matrix based on different weights, and the weighted result and the bias term jointly determine the fusion matrix.

[0087] Step S105 : inputting the question data input by the user and the fusion matrix into a recommendation model to obtain recommendation information for the performance of the electronic device.

[0088] In some embodiments, the problem data and the fusion matrix are input into a recommendation model to obtain recommendation information for the performance of the electronic device. The recommendation information is consistent with the user's preference (such as a preference for solving the problem corresponding to the problem data from a software perspective, or a preference for solving the problem corresponding to the problem data from a hardware perspective) and is consistent with the current situation of the electronic device. If the problem is solved from a software perspective, the space required to install the software to solve the problem is less than the available space (or remaining space) of the electronic device; if the problem is solved from a hardware perspective, the hardware corresponding to the recommendation information matches the electronic device.

[0089] In some embodiments, the recommendation model may be a large language model.

[0090] In this way, the recommendation method provided by the embodiment of the present disclosure effectively solves the problem of insufficient accuracy of multi-round recommendations in interactive scenarios in traditional recommendation systems by constructing a multi-dimensional association model of users, hardware and software. Compared with the traditional single representation method based on structured data, this crash embodiment performs multi-dimensional modeling based on the interest relationship and semantic information of users, software and hardware, so that the recommendation model can more accurately capture the complex association relationship between users and hardware, users and software, and between software and hardware, so that the recommendation results have higher adaptability; in addition, based on the knowledge understanding ability of the large language model, the recommendation model can not only generate personalized recommendation plans that meet user needs, but also provide logical explanations, which significantly improves the interactive experience of the recommendation process and the credibility of the results. This multi-dimensional modeling method that integrates semantic understanding enables the recommendation system to avoid the duplication or deviation problems caused by traditional sequential recommendations when dealing with new users, complex needs or multimodal scenarios, and ultimately achieves intelligent, explainable and user-friendly multi-round interactive recommendations.

[0091] Figure 2 A second optional flow chart of the recommended method provided by the embodiment of the present disclosure is shown, which will be explained according to each part.

[0092] Step S201 : determining a first-category word matrix based on the first-category information in the user's historical behavior data and the first-category information of the electronic device.

[0093] In some embodiments, the carrier obtains a first category of user tendency information from the user historical behavior data, wherein the first category of user tendency information, i.e., the first category of information in the user historical behavior data, includes software-related information, such as at least one of click information on each software information, reading information on each software page, and function information; optionally, in order to make the recommendation results more accurate, the first category of information may also include tag information and attribute information.

[0094] In some embodiments, the first category of information of the electronic device may include software information in the electronic device; specifically including at least one of installation information, operation information and update information; the installation information may include the software installation time, installation duration, uninstallation time and the difference between the uninstallation time and the current time; the operation information may include at least one of the software's running duration, running frequency, number of runs and the time of the most recent run; the update information may include the software's update frequency, number of updates, version number of each update, trigger source of each update and update package size of each update (i.e., the memory size occupied by the update package). The trigger source of each update may include website installation package updates, updates according to update instructions within the software and automatic updates. The software information may also include installation information of each software in the electronic device, such as the number of installations of each type of software.

[0095] In some embodiments, the carrier determines the first-category word embedding matrix based on the user's preference for the first category information and the first category information of the electronic device. Specifically, the carrier may determine the first-category word embedding matrix based on keywords or feature words of the user's preference for the first category information and the first category information of the electronic device.

[0096] In some embodiments, the carrier may further determine a position code for each vector in the first-category word embedding matrix, and determine a first-category input embedding matrix based on the position code for each vector in the first-category word embedding matrix and the first-category word embedding matrix. In the first-category word embedding matrix, each vector corresponds to a keyword or feature word that the user prefers for the first-category information and the first-category information of the electronic device.

[0097] Specifically, keywords located in different positions will result in different semantics expressed in the sentence. Adding the position encoding of each vector in the first-category word embedding matrix can make the expression of the first-category input embedding matrix more accurate.

[0098] In some embodiments, the carrier may further process the first type of input embedding matrix based on a transformer to make the keywords in the first type of input embedding matrix more prominent; specifically, the transformer may include at least one layer of self-attention mechanism and a feedforward network.

[0099] During specific implementation, the carrier inputs the first-category input embedding matrix into each attention head of the self-attention mechanism. After the self-attention mechanism, the result is input into the feedforward network, and the output of the last layer of the feedforward network is determined to be the first-category context vector; the first-category context vector is average pooled to obtain the first-category word matrix.

[0100] Step S202 : determining a second-category word matrix based on the second-category information in the user's historical behavior data and the second-category information of the electronic device.

[0101] In some embodiments, the carrier obtains the second category of user preference information from the user's historical behavior data, wherein the second category of user preference information, i.e., the second category of information in the user's historical behavior data, includes hardware-related information, such as click information on each piece of hardware information, hardware information query information, and function information; wherein, the hardware information query information includes information about the hardware operating status viewed through the management page of the electronic device, including the number of queries, query objects, and query frequency. Optionally, in order to make the recommendation results more accurate, the first category of information may also include tag information and attribute information.

[0102] In some embodiments, the first category of information of the electronic device may include hardware information in the electronic device; specifically including at least one of the brand information, model information, core number information, capacity information, type information, frequency information, etc. of each hardware; the frequency information includes the frequency used in work, or the operating frequency.

[0103] In some embodiments, the carrier determines the second-category word embedding matrix based on the user's preference for the second category information and the second category information of the electronic device. Specifically, the carrier may determine the second-category word embedding matrix based on keywords or feature words of the user's preference for the second category information and the second category information of the electronic device.

[0104] In some embodiments, the carrier may further determine a position code for each vector in the second-category word embedding matrix, and determine a second-category input embedding matrix based on the position code for each vector in the second-category word embedding matrix and the second-category word embedding matrix. In the second-category word embedding matrix, each vector corresponds to a keyword or feature word that the user prefers for the second-category information and the second-category information of the electronic device.

[0105] In some embodiments, the carrier may further process the second type of input embedding matrix based on the transformer to make the keywords in the second type of input embedding matrix more prominent.

[0106] During specific implementation, the carrier inputs the second-category input embedding matrix into each attention head of the self-attention mechanism. After the self-attention mechanism, the result is input into the feedforward network, and the output of the last layer of the feedforward network is determined to be the second-category context vector; the second-category context vector is average pooled to obtain the second-category word matrix.

[0107] Step S203: determining a device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device.

[0108] In some embodiments, the carrier obtains first category information of the electronic device; the first category information includes at least one of first category installation information, operation information and update information of the electronic device; the carrier obtains second category information of the electronic device, and the second category information includes at least one of brand information, model information, core number information, capacity information, type information, frequency information, etc. of each hardware; a device information embedding matrix is determined based on the first category information and the second category information; an average pooling operation is performed on the device information embedding matrix to obtain the device information matrix.

[0109] During specific implementation, the carrier may embed the keywords or feature words of the first category of information of the electronic device and the second category of information of the electronic device into a matrix.

[0110] Step S204: splicing the first category word matrix, the second category word matrix, and the device information matrix based on the first order, and weighting the first category word matrix, the second category word matrix, and the device information matrix based on the first weight, the second weight, and the third weight respectively to obtain the fusion matrix.

[0111] In some embodiments, the carrier may concatenate the first category word matrix, the second category word matrix, and the device information matrix based on a first order, and weight the first category word matrix, the second category word matrix, and the device information matrix based on a first weight, a second weight, and a third weight, respectively, and add a bias term to the weighted matrix to obtain the fusion matrix. The first order may be a preset order, such as the first category word matrix, the second category word matrix, and the device information matrix, or any order set according to business requirements.

[0112] The bias term is set based on the expected output range; and the method for determining the first weight, the second weight, and the third weight will be described later.

[0113] Step S205 : inputting the question data input by the user and the fusion matrix into a recommendation model to obtain recommendation information for the performance of the electronic device.

[0114] In some embodiments, the carrier may directly determine prompt words for a recommendation model based on the question data input by the user and the fusion matrix, input the prompt words into the recommendation model, and obtain recommendation information for the performance of the electronic device. The question data includes question information for optimizing the performance of the electronic device.

[0115] In other embodiments, the carrier may determine the user's tendency information based on at least one of the question data input by the user and the user's historical behavior data; the user's tendency information includes software tendency or hardware tendency. Determining the user's tendency information based on at least one of the question data input by the user and the user's historical behavior data may include determining the user's tendency information based on keywords in the question data and the historical behavior data; if the question data includes software type keywords, the user's tendency information is software tendency; if the question data includes hardware type keywords, the user's tendency information is hardware tendency; if the user's historical behavior data includes software type keywords, the user's tendency information is software tendency; if the user's historical behavior data includes hardware type keywords, the user's tendency information is hardware tendency.

[0116] In a specific implementation, in response to the user's tendency information being a hardware tendency, the problem data and the fusion matrix are input into the recommendation model to obtain the recommendation results output by the recommendation model; and the recommendation results that are consistent with the hardware tendency are determined to be recommendation information for the performance of the electronic device.

[0117] Optionally, the carrier can also obtain hardware category information based on the input area, specifically including hardware type (such as processor, memory, fan, etc.) and hardware core description information, such as expected performance, or specific brand information, model information, capacity information, etc.; determine prompt words based on the hardware category information and the problem data and fusion matrix, input the prompt words into the recommendation model, and obtain hardware recommendation information for the performance of the electronic device.

[0118] Alternatively, in a specific implementation, in response to the user's tendency information being a software tendency, the problem data and the fusion matrix are input into a recommendation model to obtain a recommendation result output by the recommendation model; and the recommendation result that is consistent with the software tendency is determined to be recommendation information for the performance of the electronic device.

[0119] Optionally, in response to the user's tendency information being a software tendency, the carrier determines the prompt words of the recommendation model based on the question data input by the user and the fusion matrix, and then adds a description of the software instructions or adds a preferred software recommendation to the prompt words, that is, reflects the software recommendation tendency in the prompt words, and then inputs the prompt words into the recommendation model.

[0120] In a specific implementation, in response to being unable to determine hardware tendency or software tendency based on the user's tendency information, the carrier obtains software category information based on the input area, specifically including software type (such as image processing, garbage cleaning, antivirus, games, audio and video, etc.) and software core description information, such as the expected function, or the specific software name; based on the software category information and the problem data and the fusion matrix, a prompt word is determined, and the prompt word is input into the recommendation model to obtain software recommendation information for the performance of the electronic device.

[0121] Step S206: determining target recommendation information based on the recommendation information.

[0122] In some embodiments, after the carrier determines the recommendation information, it outputs the recommendation information and receives feedback information regarding the recommendation information; the carrier determines the tendency information in the feedback information, and reconfirms the prompt word based on the tendency information and the fusion matrix, and inputs the prompt word into the recommendation model.

[0123] In specific implementation, the specific steps of the carrier reconfirming the prompt word based on the tendency information and the fusion matrix are the same as step S205, and will not be repeated here.

[0124] In this way, the recommendation method provided by the embodiment of the present disclosure effectively solves the problem of insufficient accuracy of multi-round recommendations in interactive scenarios in traditional recommendation systems by constructing a multi-dimensional association model of users, hardware and software. Compared with the traditional single representation method based on structured data, this crash embodiment performs multi-dimensional modeling based on the interest relationship and semantic information of users, software and hardware, so that the recommendation model can more accurately capture the complex association relationship between users and hardware, users and software, and between software and hardware, so that the recommendation results have higher adaptability; in addition, based on the knowledge understanding ability of the large language model, the recommendation model can not only generate personalized recommendation plans that meet user needs, but also provide logical explanations, which significantly improves the interactive experience of the recommendation process and the credibility of the results. This multi-dimensional modeling method that integrates semantic understanding enables the recommendation system to avoid the duplication or deviation problems caused by traditional sequential recommendations when dealing with new users, complex needs or multimodal scenarios, and ultimately achieves intelligent, explainable and user-friendly multi-round interactive recommendations.

[0125] Figure 3 A third optional flow chart of the recommended method provided by the embodiment of the present disclosure is shown, and will be explained according to each step.

[0126] Figure 3 The corresponding method is the method for determining the first weight, the second weight, and the third weight involved in step S204, which specifically includes:

[0127] Step S301: Obtain the first-category word test matrix, the second-category word test matrix, and the device information test matrix corresponding to each test user.

[0128] In some embodiments, the carrier obtains user historical behavior data corresponding to the test user, first category test information of the electronic device, and second category test information of the electronic device; and determines the first category word test matrix, the second category word test matrix, and the device information test matrix based on steps S101 to S103, or steps S201 to S203.

[0129] Step S302: weight the first category word test matrix, the second category word test matrix and the device information test matrix based on the first test weight, the second test weight and the third test weight, and adjust the weighted processing results based on the tendency information of each test user.

[0130] In some embodiments, each test user corresponds to tendency information, i.e., software tendency information or hardware tendency information. For different tendency information, the weights of the first-category word test matrix, the second-category word test matrix, and the device information test matrix are also different. For example, for hardware tendency, the weight of the second-category word test matrix needs to be higher, and for software tendency, the weight of the first-category word test matrix needs to be higher.

[0131] The embodiments of the present disclosure aim to solve the problem that the carrier can ensure that the weights can be adaptively adjusted according to the input data, and the bias term can correctly affect the final output.

[0132] In specific implementation, set the first type of word test matrix to v d , the second type of word test matrix is v r , the device information test matrix is v h , the carrier splices the above three matrices in a preset order to obtain a new vector v concat .

[0133] In some embodiments, a mechanism is designed to assign a dynamic weight to each feature vector. This can be achieved by defining a weight vector W = [w1, w2, w3], where w1 is the first test weight, w2 is the second test weight, and w3 is the third test weight. These weights are learnable parameters that can be adjusted during training to optimize the objective function.

[0134] A fixed bias term b is introduced, which remains unchanged throughout the process and is used to adjust the baseline value of the fusion result. The choice of the bias term should be based on the understanding of the data and the expected output range. Using the three test weights and the fixed bias term mentioned above, the fused feature vector v is calculated. fusion =Wf ·v contat +b f Among them, W f is the weight vector W after training.

[0135] Implement the above algorithm to ensure that the weights can be automatically adjusted according to the input data, and the bias term can correctly affect the final output.

[0136] In specific implementation, the carrier can adjust the results of weighted processing based on the tendency information of each test user, and feed back the adjusted results to the large model, so that the large model adjusts relevant parameters, so that in subsequent applications, the weights and bias items of each matrix can be adaptively determined according to the user's tendency information.

[0137] In some optional embodiments, a prompt word can be set to allow the large language model to learn the weights of the first-category word test matrix, the second-category word test matrix, and the device information test matrix, such as: "In this task, you need to deal with a problem involving the fusion of three-dimensional data matrices. Given three different feature vectors f{v r}, f{v d}, f{v h}, these vectors represent data features of different dimensions. Your goal is to integrate these feature vectors together and achieve three-dimensional matrix fusion by introducing dynamic learning weights and a fixed bias term.

[0138] In this way, through the recommendation method described in the embodiment of the present disclosure, the weights of the first-category word matrix, the second-category word matrix and the device information matrix can be automatically adjusted according to the different input data, and the bias item can correctly affect the final output to ensure the accuracy of the final recommendation information.

[0139] Figure 4 A fourth optional flow chart of the recommended method provided by an embodiment of the present disclosure is shown, and will be explained according to each part.

[0140] In some embodiments, a dynamic loss adjustment mechanism is introduced to enable the recommendation model to better adapt to the training requirements of different stages. By adjusting the loss weights of multiple tasks, the recommendation model can focus on optimizing certain tasks during training.

[0141] Step S401 : determining a weighted recommendation loss function based on the fourth weight, the fifth weight, the positive fusion test matrix corresponding to the positive sample, and the negative fusion test matrix corresponding to the negative sample.

[0142] In some embodiments, positive samples and negative samples are determined, and the positive samples and negative samples are determined based on the first category word test matrix, the second category word test matrix and the device information test matrix in steps S101 to S103, or steps S201 to S203. And weighted summation and bias term processing are performed. The fusion matrix corresponding to the positive sample is v fusion , the fusion matrix corresponding to the negative sample is

[0143] In some embodiments, the carrier sets a fourth weight α re1 , and the fifth weight β re1 ; where α re1 A value greater than 1 is used to amplify the weight of positive samples and improve the recommendation ability of the model. re1 is a value less than 1, thereby reducing the penalty for negative samples. The carrier can determine the weighted recommendation loss function based on the following formula

[0144]

[0145] Step S402 : determining a difficult recommendation loss function based on the sixth weight, the difficult fusion test matrix corresponding to the difficult sample, and the easy fusion matrix corresponding to the easy sample.

[0146] In some embodiments, the carrier guides the model to focus more on difficult samples by increasing the weight of difficult samples (samples that are difficult to predict) and reducing the weight of easy samples (samples that the model has predicted well). The carrier can determine the difficult recommendation loss function based on the following formula

[0147]

[0148] Among them, v fusion is the fusion matrix corresponding to the difficult sample; is the fusion matrix corresponding to the easy sample, and γ is the sixth weight, which controls the model's discrimination between easy and difficult samples.

[0149] Step S403: determining a total loss function of the recommendation model based on the weighted recommendation loss function and the difficult recommendation loss function.

[0150] In some embodiments, the vector may determine the total loss function based on the following formula:

[0151]

[0152] Among them, λ1 and λ2 are weight parameters of the two positive and negative sample tasks (corresponding to step S401) and the difficult and easy sample tasks (corresponding to step S402), which are used to control the contribution of the loss of each task to the total loss.

[0153] In some embodiments, in order to adaptively adjust the loss weight, the carrier introduces the concept of gradient change rate to dynamically adjust the task weight. The gradient change rate reflects the optimization difficulty of the task. A large gradient change means that the task has not been fully optimized and the weight needs to be increased; a small gradient change means that the task has converged and the weight can be reduced. Specifically, it includes:

[0154] Step S404: determine the loss function gradient of each round, and update the weight of the recommendation model based on the loss function gradient.

[0155] Assume that the gradient of the loss function at step t is:

[0156]

[0157] The following formula can be used to update the weight λ according to the gradient change rate:

[0158]

[0159] where ∈ is a small positive number that prevents the denominator from being zero.

[0160] Thus, through the recommendation method described in the embodiments of this disclosure, when the gradient of a task is large, its weight will be reduced; when the gradient is small, the weight will be increased. In this way, the recommendation model can adaptively adjust its focus on different tasks. As training progresses, the loss weight is dynamically adjusted so that the model can balance the optimization between different tasks. In this way, the overall performance of the recommendation model is improved.

[0161] Figure 5 FIG. 5 shows a fifth optional flow chart of the recommended method provided in the embodiment of the present disclosure. Figure 6 A schematic flow chart of the recommended method provided by the embodiment of the present disclosure is shown. Figure 5 and Figure 6 Provide explanation.

[0162] Step S501 : semantically vectorize the user's historical behavior data of the electronic device related information.

[0163] In some embodiments, as Figure 6As shown, the relevant information of the electronic device includes at least the model, attributes (such as silver appearance and large memory), core features (such as processor model i5, launched in 2022, and thin and light laptop), the type of electronic device is a laptop (Notebook), first-category information, and second-category information. User historical behavior data includes at least user attributes (such as organizational relationship, superiors and subordinates, position), user tags (such as income, marital status), the main purpose of the user's use of the electronic device (programming, office, application), and historical records, including click information, English reading information, last use time, etc.

[0164] In some embodiments, the carrier determines a first-category word matrix based on the first-category information of the electronic device and the user's historical behavior data, D=[u1,u2,…,u m ]; Determine the second type of word matrix R = [w1, w2, ..., w based on the second type of information of the electronic device and the user's historical behavior data n ]. In the first and second word matrices, the embedding of each word can be a vector with a dimension of d, w i ,u j ∈R d , position encoding is performed on the vectors in the first-category word matrix and the second-category word matrix respectively, retaining the sequence information:

[0165] R (0) =R+P R ,D (0) =D+P D

[0166] Among them, P R is the position encoding matrix corresponding to the second type of word matrix R, and its size is the same as that of the second type of word matrix; P D It is the position encoding matrix corresponding to the first category word matrix D, and its size is the same as that of the second category word matrix.

[0167] The second word matrix R (0) And the first-category word matrix D (0) Each layer of the transformer includes a self-attention mechanism and a feedforward neural network:

[0168] Self-attention mechanism: For the lth layer, the input

[0169] R (l-1) and D (l-1)

[0170] After multi-head self-attention calculation:

[0171]

[0172] Where Q = W Q R (l-1) , K=W K R (l-1) , V=W V R (l-1) , W Q ,W K ,W V is the projection matrix. The calculation of the first type word matrix D is similar. The output after the self-attention mechanism passes through the feedforward network:

[0173] R (l) =FFN(Attention(Q,K,V)),D (l) =FFN(Attention(Q′,K′,V′))

[0174] Feedforward networks usually consist of two fully connected layers with ReLU activation functions.

[0175] After the above steps, the final second-category word matrix R is obtained (L) And the first-category word matrix D (L) In order to compress the information in the first-category word matrix and the second-category word matrix into a fixed-length vector, an average pooling operation is used on the first-category word matrix and the second-category word matrix:

[0176]

[0177] here, and They are the vector representation of each word in the last layer of the second-category word matrix and the first-category word matrix, respectively.

[0178] In particular, for the device information matrix C=[c1,c2,…,c a ], you can directly use the average pooling operation on the vector based on the large model:

[0179]

[0180] In some embodiments, after obtaining the first-category word matrix, the second-category word matrix, and the device information matrix, an LLM is used to generate a three-dimensional data matrix.

[0181] In order to better capture the interaction between different dimensions, the concatenated vectors can be fused based on the large model:

[0182] v fusion =W f ·[llm_v r,llm_v d ,llm_v h ]+b f

[0183] in, is the learnable weight matrix determined based on step S301 and step S302, b f ∈R d is the bias term. After this transformation, we finally get the d-dimensional vector (i.e., fusion matrix) v fusion ∈R d .

[0184] Step S502: Generate recommendation information based on the user's preference information.

[0185] In some embodiments, as Figure 6 As shown, after determining the first category word matrix, the second category word matrix and the device information matrix, the three matrices are input into the recommendation model, and the recommendation model determines the recommendation information for the electronic device performance.

[0186] Figure 7 A processing diagram of the recommendation model provided by an embodiment of the present disclosure is shown.

[0187] Step S5021: Determine the user's preference information.

[0188] In some embodiments, the recommendation model determines the user's tendency information based on the question data input by the user; or determines the user's tendency information based on the question data input by the user and the user's historical behavior data.

[0189] Optionally, if the recommendation model has completed at least one inference based on the question data input by the user, the recommendation model can also determine the user's tendency information again based on the feedback information, the question data input by the user and at least one of the user's historical behavior data, and execute steps S5021 to S5025.

[0190] In some embodiments, in response to the user's preference information being hardware information, step S5022 is executed; in response to the user's preference information being non-hardware information, step S5023 is executed.

[0191] Step S5022: Determine a first prompt word based on the fusion matrix and the hardware category information.

[0192] In some embodiments, the recommendation model obtains the fusion matrix generated in step S501.

[0193] In some embodiments, the recommendation model obtains hardware category information based on the input area, specifically including hardware type (such as processor, memory, fan, etc.) and hardware core description information, such as expected performance, or specific brand information, model information, capacity information, etc.

[0194] In some embodiments, the recommendation model determines a first prompt word based on the hardware category information, the problem data, and a fusion matrix, and inputs the first prompt word into the recommendation model to obtain hardware recommendation information for the performance of the electronic device.

[0195] Step S5023: Determine whether the user's preference information is software information.

[0196] In some embodiments, in response to the user's tendency information being software information, step S5024 is executed; in response to the user's tendency information being no tendency information, step S5025 is executed.

[0197] Step S5024: Determine a second prompt word based on the fusion matrix and the software category information.

[0198] In some embodiments, the recommendation model obtains the fusion matrix generated in step S501.

[0199] In some embodiments, the recommendation model obtains software category information based on the input area, specifically including software type (such as image processing, junk cleaning, antivirus, games, audio and video, etc.) and software core description information, such as the expected function, or the specific software name; and determines the first sub-prompt word based on the software category information, the problem data, and the fusion matrix.

[0200] In some embodiments, the recommendation model can also generate a second sub-prompt word based on describing software instructions or software recommendations, that is, the second sub-prompt word includes software recommendation tendencies, and then input the first sub-prompt word and the second sub-prompt word as second prompt words into the recommendation model to obtain software recommendation information for the performance of the electronic device.

[0201] Step S5025: Determine a third prompt word based on the software category information.

[0202] In some embodiments, the recommendation model obtains software category information based on the input area; determines a third prompt word based on the software category information, and inputs the third prompt word into the recommendation model to obtain software recommendation information for the performance of the electronic device.

[0203] In this way, the recommendation method provided by the embodiment of the present disclosure effectively solves the problem of insufficient accuracy of multi-round recommendations in interactive scenarios in traditional recommendation systems by constructing a multi-dimensional association model of users, hardware and software. Compared with the traditional single representation method based on structured data, this crash embodiment performs multi-dimensional modeling based on the interest relationship and semantic information of users, software and hardware, so that the recommendation model can more accurately capture the complex association relationship between users and hardware, users and software, and between software and hardware, so that the recommendation results have higher adaptability; in addition, based on the knowledge understanding ability of the large language model, the recommendation model can not only generate personalized recommendation plans that meet user needs, but also provide logical explanations, which significantly improves the interactive experience of the recommendation process and the credibility of the results. This multi-dimensional modeling method that integrates semantic understanding enables the recommendation system to avoid the duplication or deviation problems caused by traditional sequential recommendations when dealing with new users, complex needs or multimodal scenarios, and ultimately achieves intelligent, explainable and user-friendly multi-round interactive recommendations.

[0204] For example, if the problem information includes: "My computer is very slow, are there any recommended junk cleaning programs?", then the user's preference information is software preference. Furthermore, based on the user's historical behavior data, including clicks on software information, such as clicks on different junk cleaning software and page views, the user's preference is determined. Recommended information that matches the user's preference is then filtered out from the recommended information. Alternatively, if the problem information includes: "My computer is very slow, do I need to replace the memory?", then the user's preference information is hardware preference. Based on the user's historical behavior data, including clicks on hardware information, such as clicks on different hardware (e.g., processor clicks, memory clicks), page views, etc., the user's preference is determined (desire for processor recommendations or desire for memory recommendations). Furthermore, based on the second category of electronic device information, such as the model and storage space of the electronic device, hardware information that matches the second category of electronic device information is identified. Then, based on the user's preference, recommended information that meets the user's needs is filtered out from the hardware information. Furthermore, income information or the user's primary use of the electronic device can be combined to recommend hardware that is appropriately priced and meets the performance requirements for the intended use.

[0205] Figure 8 An optional structural diagram of a recommendation device provided in an embodiment of the present disclosure is shown, and will be described according to each part.

[0206] In some embodiments, the recommendation device 700 includes a first processing unit 701 , a second processing unit 702 , a third processing unit 703 , a fusion unit 704 and a recommendation unit 705 .

[0207] The first processing unit 701 is configured to determine a first-category word matrix based on the first-category information in the user's historical behavior data and the first-category information of the electronic device;

[0208] The second processing unit 702 is configured to determine a second-category word matrix based on the second-category information in the user's historical behavior data and the second-category information of the electronic device;

[0209] The third processing unit 703 is configured to determine a device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device;

[0210] The fusion unit 704 is configured to determine a fusion matrix based on the first-category word matrix, the second-category word matrix, and the device information matrix;

[0211] The recommendation unit 705 is configured to input the question data input by the user and the fusion matrix into a recommendation model to obtain recommendation information for the performance of the electronic device.

[0212] The first processing unit 701 is specifically configured to obtain first-category information of user tendencies based on user historical behavior data;

[0213] Determine a first-category word embedding matrix based on the first-category information of the user preference and the first-category information of the electronic device;

[0214] Determining a position code for each vector in the first class word embedding matrix, and determining a first class input embedding matrix based on the position code for each vector and the first class word embedding matrix;

[0215] The first-category input embedding matrix is input into the transformer to obtain a first-category word matrix.

[0216] In some embodiments, the transformer includes at least one layer of self-attention mechanism and feedforward network, and the first processing unit 701 is specifically configured to input the first-type input embedding matrix into any self-attention mechanism and feedforward network to obtain a first-type context vector;

[0217] An average pooling operation is performed on the first-category context vector to obtain the first-category word matrix.

[0218] The third processing unit 703 is specifically configured to obtain first-category information of the electronic device; the first-category information includes at least one of first-category installation information, operation information, and update information of the electronic device;

[0219] obtaining second type of information of the electronic device;

[0220] determining a device information embedding matrix based on the first type of information and the second type of information;

[0221] An average pooling operation is performed on the device information embedding matrix to obtain the device information matrix.

[0222] The fusion unit 704 is specifically configured to concatenate the first-category word matrix, the second-category word matrix, and the device information matrix based on a first order, and perform weighted processing on the first-category word matrix, the second-category word matrix, and the device information matrix based on a first weight, a second weight, and a third weight, respectively, to obtain the fusion matrix.

[0223] The fusion unit 704 is further configured to determine the first weight, the second weight, and the third weight based on the training data, specifically including:

[0224] Obtain the first-category word test matrix, the second-category word test matrix, and the device information test matrix corresponding to each test user;

[0225] The first category word test matrix, the second category word test matrix and the device information test matrix are weighted based on the first test weight, the second test weight and the third test weight respectively, and the result of the weighted processing is adjusted based on the tendency information of each test user so that the result of the weighted processing is consistent with the tendency information of the user.

[0226] The recommendation unit 705 is specifically configured to input the fusion matrix into a recommendation model to obtain a recommendation result output by the recommendation model;

[0227] Determining the user's tendency information based on the question data input by the user;

[0228] It is determined that the recommendation results that are consistent with the user's preference information are recommendation information for the electronic device performance.

[0229] The recommendation unit 705 is specifically configured to determine a weighted recommendation loss function based on the fourth weight, the fifth weight, a positive fusion test matrix corresponding to the positive sample, and a negative fusion test matrix corresponding to the negative sample;

[0230] Determine a difficult recommendation loss function based on the sixth weight, the difficult fusion test matrix corresponding to the difficult sample, and the easy fusion matrix corresponding to the easy sample;

[0231] Determining a total loss function of the recommendation model based on the weighted recommendation loss function and the difficult recommendation loss function;

[0232] Determine the loss function gradient of each round, and update the weight of the recommendation model based on the loss function gradient;

[0233] The fourth weight is greater than the fifth weight.

[0234] According to an embodiment of the present disclosure, the present disclosure also provides an electronic device and a readable storage medium.

[0235] Figure 9 A schematic block diagram of an example electronic device 800 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are provided as examples only and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0236] like Figure 9 As shown, the electronic device 800 includes a computing unit 801, which can perform various appropriate actions and processes according to a computer program stored in a read-only memory (ROM) 802 or a computer program loaded from a storage unit 808 into a random access memory (RAM) 803. In the RAM 803, various programs and data required for the operation of the electronic device 800 can also be stored. The computing unit 801, the ROM 802, and the RAM 803 are connected to each other via a bus 804. An input / output (I / O) interface 805 is also connected to the bus 804.

[0237] Multiple components in the electronic device 800 are connected to the I / O interface 805, including an input unit 806, such as a keyboard, a mouse, etc.; an output unit 807, such as various types of displays, speakers, etc.; a storage unit 808, such as a magnetic disk, an optical disk, etc.; and a communication unit 809, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 809 allows the electronic device 800 to exchange information / data with other devices via a computer network such as the Internet and / or various telecommunication networks.

[0238] The computing unit 801 can be a variety of general-purpose and / or specialized processing components with processing and computing capabilities. Some examples of the computing unit 801 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various dedicated artificial intelligence (AI) computing chips, various computing units that run machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 801 performs the various methods and processes described above, such as the recommendation method. For example, in some embodiments, the recommendation method can be implemented as a computer software program that is tangibly contained in a machine-readable medium, such as the storage unit 808. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 800 via the ROM 802 and / or the communication unit 809. When the computer program is loaded into the RAM 803 and executed by the computing unit 801, one or more steps of the recommendation method described above can be performed. Alternatively, in other embodiments, the computing unit 801 can be configured to perform the recommendation method by any other appropriate means (e.g., by means of firmware).

[0239] Various embodiments of the systems and techniques described herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), system-on-chip systems (SOCs), programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments can include being implemented in one or more computer programs that are executable and / or interpreted on a programmable system that includes at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0240] The program code for implementing the method of the present disclosure can be written in any combination of one or more programming languages. These program codes can be provided to a processor or controller of a general-purpose computer, a special-purpose computer, or other programmable data processing device so that when the program code is executed by the processor or controller, the functions / operations specified in the flow chart and / or block diagram are implemented. The program code can be executed entirely on the machine, partially on the machine, as a stand-alone software package, partially on the machine and partially on a remote machine, or entirely on a remote machine or server.

[0241] In the context of the present disclosure, a machine-readable medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or equipment, or any suitable combination of the foregoing. A more specific example of a machine-readable storage medium can include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0242] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and pointing device (e.g., a mouse or trackball) through which the user can provide input to the computer. Other types of devices can also be used to provide interaction with the user; for example, the feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form (including acoustic input, voice input, or tactile input).

[0243] The systems and techniques described herein can be implemented in a computing system that includes back-end components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes front-end components (e.g., a user computer having a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such back-end components, middleware components, or front-end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0244] A computer system may include a client and a server. The client and server are generally remote from each other and typically interact through a communication network. The client-server relationship arises through computer programs running on the respective computers and having a client-server relationship with each other. The server may be a cloud server, a server in a distributed system, or a server integrated with a blockchain.

[0245] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in this disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in this disclosure can be achieved. This is not a limitation herein.

[0246] The above description is merely a specific embodiment of the present disclosure, but the scope of protection of the present disclosure is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in this disclosure should be included in the scope of protection of the present disclosure. Therefore, the scope of protection of the present disclosure should be based on the scope of protection of the claims.

Claims

1. A recommendation method, comprising: Determining a first-category word matrix based on the first-category information in the user's historical behavior data and the first-category information of the electronic device; Determining a second-category word matrix based on the second-category information in the user's historical behavior data and the second-category information of the electronic device; determining a device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device; Determine a fusion matrix based on the first category word matrix, the second category word matrix and the device information matrix; Inputting the user inputted question data and the fusion matrix into the recommendation model to obtain recommendation information for the performance of the electronic device; The first type of information and the second type of information represent different attribute information of the electronic device.

2. The method according to claim 1, wherein determining the first-category word matrix based on the first-category information in the user's historical behavior data and the first-category information of the electronic device comprises: Obtain the first type of information about user tendencies based on user historical behavior data; Determine a first-category word embedding matrix based on the first-category information of the user preference and the first-category information of the electronic device; Determining a position code for each vector in the first class word embedding matrix, and determining a first class input embedding matrix based on the position code for each vector and the first class word embedding matrix; The first-category input embedding matrix is input into the transformer to obtain a first-category word matrix.

3. The method according to claim 2, wherein the transformer comprises at least one layer of a self-attention mechanism and a feedforward network, and wherein the step of inputting the first-category input embedding matrix into the transformer to obtain the first-category word matrix comprises: Input the first-class input embedding matrix into any self-attention mechanism and feedforward network to obtain the first-class context vector; An average pooling operation is performed on the first-category context vector to obtain the first-category word matrix.

4. The method according to claim 1, wherein determining the device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device comprises: Obtaining the first type of information of electronic devices; The first type of information includes at least one of first type of installation information, operation information and update information of the electronic device; obtaining second type of information of the electronic device; determining a device information embedding matrix based on the first type of information and the second type of information; An average pooling operation is performed on the device information embedding matrix to obtain the device information matrix.

5. The method according to claim 1, wherein determining a fusion matrix based on the first category word matrix, the second category word matrix, and the device information matrix comprises: The first category word matrix, the second category word matrix and the device information matrix are spliced based on a first order, and the first category word matrix, the second category word matrix and the device information matrix are weighted based on a first weight, a second weight and a third weight to obtain the fusion matrix.

6. The method according to claim 5, further comprising: Determining the first weight, the second weight, and the third weight based on the training data specifically includes: Obtain the first-category word test matrix, the second-category word test matrix, and the device information test matrix corresponding to each test user; The first category word test matrix, the second category word test matrix and the device information test matrix are weighted based on the first test weight, the second test weight and the third test weight respectively, and the result of the weighted processing is adjusted based on the tendency information of each test user so that the result of the weighted processing is consistent with the user's tendency information.

7. The method according to claim 1, wherein inputting the user-inputted question data and the fusion matrix into a recommendation model to obtain recommendation information for the electronic device performance comprises: Inputting the fusion matrix into a recommendation model to obtain a recommendation result output by the recommendation model; Determining the user's tendency information based on the question data input by the user; It is determined that the recommendation results that are consistent with the user's preference information are recommendation information for the electronic device performance.

8. The method according to claim 1 or 7, further comprising training a recommendation model, specifically comprising: Determining a weighted recommendation loss function based on the fourth weight, the fifth weight, a positive fusion test matrix corresponding to the positive sample, and a negative fusion test matrix corresponding to the negative sample; Determine a difficult recommendation loss function based on the sixth weight, the difficult fusion test matrix corresponding to the difficult sample, and the easy fusion matrix corresponding to the easy sample; Determining a total loss function of the recommendation model based on the weighted recommendation loss function and the difficult recommendation loss function; Determine the loss function gradient of each round, and update the weight of the recommendation model based on the loss function gradient; The fourth weight is greater than the fifth weight.

9. A recommendation device, comprising: a first processing unit, configured to determine a first-category word matrix based on the first-category information in the user's historical behavior data and the first-category information of the electronic device; a second processing unit, configured to determine a second-category word matrix based on the second-category information in the user's historical behavior data and the second-category information of the electronic device; a third processing unit, configured to determine a device information matrix based on the first type of information of the electronic device and the second type of information of the electronic device; a fusion unit, configured to determine a fusion matrix based on the first-category word matrix, the second-category word matrix, and the device information matrix; The recommendation unit is configured to input the question data input by the user and the fusion matrix into a recommendation model to obtain recommendation information for the performance of the electronic device.

10. An electronic device comprising: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method according to any one of claims 1 to 8.

Images