Object recommendation method and device and storage medium
By calculating the user's user style vector and entering the user style prediction model, the problem of inaccurate courseware template recommendations in the existing technology is solved, and more accurate courseware recommendations and improved user experience is achieved.
Patent Information
- Application Number
- CN202311452351.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2023-11-02
- Publication Date
- 2025-05-06
AI Technical Summary
When recommending courseware templates, it is difficult for existing courseware resource platforms to accurately cover and describe the courseware styles that users prefer, resulting in inaccurate recommendations and poor user experience.
By obtaining the user's history courseware abbreviation information, the user's user style vector is calculated, and inputting it into the user style prediction model to output more accurate user style tags, and then recommending matching courseware templates.
It improves the accuracy of courseware recommendations, can better meet users' current preferences and improve user experience.
Smart Images
Figure CN119939005A_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of deep learning, and in particular to an object recommendation method, apparatus, device and storage medium. Background Art
[0002] With the development and popularization of computer technology and multimedia technology, it has become a trend for users to play courseware on a projection screen or display screen to accompany their speeches in teaching, conferences and other scenarios. Therefore, making courseware has become a basic skill for teachers and other users.
[0003] In order to reduce the difficulty of courseware production and improve production efficiency, some courseware resource platforms provide various courseware templates for users to use. Users can search and obtain suitable courseware templates from the courseware resource platform according to their preferences, or directly obtain courseware templates recommended by the courseware resource platform, and then produce courseware based on the obtained courseware templates.
[0004] Existing courseware resource platforms usually first infer the style labels of user-preferred courseware based on courseware style prediction models, and then recommend courseware templates corresponding to the style labels to users. Once the label system cannot cover or correctly describe the user's preferred courseware style, it is difficult to recommend a courseware template that satisfies the user, resulting in a poor user experience. Summary of the invention
[0005] The embodiments of the present application provide an object recommendation method, apparatus, device and storage medium, which can solve the problems of inaccurate recommendation of existing courseware templates and poor user experience. To solve the above problems, the technical solutions provided by the embodiments of the present application are as follows:
[0006] In a first aspect, an embodiment of the present application provides an object recommendation method, comprising:
[0007] Acquire thumbnail information of a history object of a first user, and acquire an object style vector of the history object based on the thumbnail information;
[0008] Calculating a user style vector of the first user according to the multiple object style vectors of the first user, and inputting the user style vector into a user style prediction model;
[0009] A user style label output by the user style prediction model is obtained to recommend candidate objects corresponding to the user style label to the first user.
[0010] In a second aspect, an embodiment of the present application provides an object recommendation device, including:
[0011] An acquisition module, configured to acquire thumbnail information of a history object of a first user, and acquire an object style vector of the history object based on the thumbnail information;
[0012] a calculation module, configured to calculate a user style vector of the first user according to the plurality of object style vectors of the first user, and input the user style vector into a user style prediction model;
[0013] A recommendation module is used to obtain the user style label output by the user style prediction model to recommend candidate objects corresponding to the user style label to the first user.
[0014] In a fourth aspect, an embodiment of the present application provides an object recommendation device, comprising: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method described in the first aspect.
[0015] In a fifth aspect, an embodiment of the present application provides a computer-readable storage medium, which stores a plurality of instructions, wherein the instructions are suitable for being loaded by a processor and executing the method as described in the first aspect.
[0016] Taking the above-mentioned object as a courseware as an example, the thumbnail information of the historical object can be a thumbnail of the history courseware, and the object style vector can be a courseware style vector. In an embodiment of the present application, a thumbnail of a history courseware belonging to a certain user can be input into a courseware style prediction model to perform feature extraction and / or classification decisions, and the output of the last layer of the model is taken to obtain a one-dimensional vector: a courseware style vector. Then, the user style vector of the user is obtained using multiple courseware style vectors of the same user. The user style vector integrates multiple coursewares produced or used by the user, and can more accurately reflect the user's style. Then, after the user style vector is input into the user style prediction model, a more accurate user style label can be output to reflect the user's style. Furthermore, the alternative courseware recommended to the user based on the user style label can more accurately meet the user's current preferences, thereby improving the accuracy of the courseware recommendation. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.
[0018] Figure 1 A schematic diagram of the principle of a courseware recommendation method provided in an embodiment of the present application;
[0019] Figure 2 A schematic diagram of another method for recommending courseware provided in an embodiment of the present application;
[0020] Figure 3 A schematic diagram of a training process of a user style prediction model provided in an embodiment of the present application;
[0021] Figure 4 A schematic diagram of a method for obtaining input training data for a user style prediction model provided in an embodiment of the present application;
[0022] Figure 5 A schematic diagram of a method for obtaining output training data of a user style prediction model provided in an embodiment of the present application;
[0023] Figure 6 A flowchart of an object recommendation method provided in an embodiment of the present application;
[0024] Figure 7 A schematic diagram of the structure of an object recommendation device provided in an embodiment of the present application;
[0025] Figure 8 A schematic diagram of the structure of an object recommendation device provided in an embodiment of the present application. DETAILED DESCRIPTION
[0026] With the development and popularization of computer technology and multimedia technology, it has become a trend for users to play courseware on a projection screen or display screen to accompany their speeches in teaching, conferences and other scenarios. Therefore, making courseware has become a basic skill for teachers and other users.
[0027] Each user has his or her own style when making courseware. Different users may make different styles of courseware for the same content. In order to improve the efficiency of courseware production, users can obtain courseware templates from the courseware resource platform and make courseware based on the courseware templates. Users can actively select courseware templates according to their preferred style, or they can select courseware templates based on the recommendations of the courseware resource platform.
[0028] The current courseware resource platform mainly recommends courseware based on courseware labels. Specifically, the process of courseware resource platform recommending courseware may include: firstly inferring the style label of the courseware preferred by the user based on the courseware style prediction model, and then recommending the courseware template corresponding to the style label to the user. For example, see Figure 1In a courseware recommendation method provided in the present application, the courseware previously made and / or used by the user (which may be referred to as the user's historical courseware) can be obtained; each user's historical courseware is used as the input data source of the courseware style prediction model, and the courseware style prediction model is used to infer the courseware style label of each user's historical courseware; then these courseware style labels are simply weighted and averaged; the weighted average is used as the user's style label, and then the courseware template corresponding to the user's style label is recommended to the user. Whether the above method can recommend user-satisfactory courseware depends on the accuracy of the courseware style prediction and the integrity of the label system. Once the label system cannot cover or cannot correctly describe the user's preferred courseware style, it is difficult to recommend a user-satisfactory courseware template. This leads to inaccurate courseware template recommendations and poor user experience.
[0029] It is worth mentioning that Figure 1 The illustrated implementation method was obtained by the inventor of the present application during the process of developing a courseware recommendation method, and it was not a method known to the public at home and abroad before the present application was proposed.
[0030] Based on this, the present application also provides another object recommendation method, which can be applied to scenarios such as courseware recommendation, home style recommendation, and outfit recommendation. The following will take the courseware recommendation method as an example and further describe the various implementation methods of the present application in detail in conjunction with the accompanying drawings.
[0031] See also Figure 2 This application builds a user style prediction model based on the courseware style prediction model to predict the user style label, so as to recommend courseware to the user based on the user style label. In addition, this application can also combine the user recommendation feedback module to fine-tune the model based on the courseware template selected by the user. In this way, the object recommendation method provided by this application can more accurately describe the user style, and then recommend courseware with a style that satisfies the user.
[0032] It should be clear that the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in the field without creative work are within the scope of protection of the present application.
[0033] In the description of the present application, it should be understood that the terms "first", "second", etc. are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances. In addition, in the description of the present application, unless otherwise specified, "multiple" refers to two or more. "And / or" describes the association relationship of associated objects, indicating that three relationships may exist. For example, A and / or B can represent: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the objects associated before and after are in an "or" relationship.
[0034] In one embodiment, before using the user style prediction model to recommend candidate objects to the first user, the present application may pre-train the user style prediction model. The first user may be any user who needs the courseware recommendation service. Figure 3 , the training steps may include:
[0035] S302, constructing a second training data set for a user style prediction model based on the first training data set for the object style prediction model;
[0036] In implementation, for ease of description, the training data set of the object style prediction model can be referred to as the first training data set, and the training data set of the user style prediction model can be referred to as the second training data set. The first training data set can be provided by the object production platform. For example, the courseware production platform has a certain accumulation of courseware produced and shared by each user, and these courseware can be manually annotated in advance, and the courseware can be labeled with corresponding courseware style labels, such as ancient style, minimalism, etc. Correspondingly, the training data of the first training data set can include multiple historical courseware with labels for multiple users. Furthermore, the second training data set can be constructed based on the first training data set.
[0037] S304: Use the second training data set to train the user style prediction model.
[0038] By adopting this embodiment, a user style prediction model can be trained based on multiple history courses of the user, comprehensively considering the characteristics of multiple history courses of each user, thereby expanding the coverage of the label system, correctly describing the user style, and recommending a courseware template that satisfies the user.
[0039] In one embodiment, step S302 may specifically include: obtaining a first training data set; wherein the input training data of the first training data set is thumbnail information of the second user's historical objects, and the output training data of the first training data set is an object style label of the second user's historical objects; inputting the input training data of the first training data set into an object style prediction model to obtain an object style vector output by a target level of the object style prediction model; calculating a user style vector of the second user based on multiple object style vectors of the second user, and calculating a user style label of the second user based on multiple object style labels of the second user; and using the user style vector and the user style label of the second user as the input training data and output training data of the second training data set, respectively.
[0040] In implementation, the second user may be any user to whom the training data used by the user style prediction model in the training phase belongs. A single user often has multiple coursewares, and its style may be composed of the coursewares produced and / or used by it. The input training data of the courseware style prediction model are thumbnails of each history courseware of the second user, and the output training data of the courseware style prediction model are courseware style labels corresponding to each history courseware of the user. In addition, the output vector of the target layer of the courseware style prediction model can be taken as the courseware style vector to describe the courseware style. Among them, the target layer can be the last layer or multiple layers of the object style prediction model.
[0041] See also Figure 4 , the user style vector can be calculated based on multiple courseware style vectors, and the user style vector can be used as input training data for the user style prediction model. Figure 5 , the user style label can be calculated based on multiple courseware style labels of a single user, and the user style label can be used as the output training data of the user style prediction model.
[0042] It can be understood that the output data of the last layer or layers of the courseware style prediction model has undergone multiple layers of feature selection and extraction processes, and these features have a certain reference and guidance role in the training of the user style prediction model. In this way, the user style prediction model can use these processed features for learning during the training phase, which can improve the efficiency and accuracy of the user style prediction model.
[0043] In one embodiment, the courseware style vector may be a one-dimensional array containing n elements, denoted as [x1, x2, x3, ... x n ], accordingly, the courseware style label can be recorded as [l1,l2,l3,...l n ]; the user style vector can be recorded as [xt1,xt2,xt3,...xt n ], user style tags can be recorded as [lt1, lt2, lt3, ... ltn ].
[0044] In one embodiment, the weighted average vector xt of multiple courseware style vectors of a single user can be i The training vector used for the user style prediction model is the user style vector. The calculation formula of the user style vector can be:
[0045]
[0046] Among them, xt i represents the value of the user style vector in the i-th dimension; N is the number of courseware created by the user, x i,k refers to the value of the kth courseware style vector on the i-th dimension; w k It is the weight of the kth courseware, which is related to the production time of the courseware.
[0047] It is understandable that user-made courseware tends to have a short-term tendency, and the styles of courseware made in the same period will be more similar. In other words, the closer the date of completion of the courseware is to the current date, the more the style of the courseware can reflect the user's preferences. Therefore, if the courseware that matches the user in the near future is recommended, the user will be more likely to read and use it.
[0048] In one embodiment, the weighted average vector of multiple courseware style labels of a single user can be used as a training label of the user style prediction model, that is, the user style label. The calculation formula of the user style label can be:
[0049]
[0050] Among them, lt i represents the value of the user style label in the i-th dimension; N is the number of courseware created by the user; l i,k refers to the value of the i-th dimension of the k-th courseware style label vector; w k It is the weight of the kth courseware, which is related to the production time of the courseware.
[0051] In one embodiment, the calculation formula of the weight w of each courseware may be:
[0052] W = (maxtime-time) / maxtime
[0053] Among them, time refers to the number of days a certain courseware has been produced, that is, the number of days from the date the courseware was completed to the current date. Maxtime refers to the maximum number of days, that is, a period before the current date. Maxtime can be a preset constant. In this way, the courseware whose completion date is closer to the current date will be given a greater weight.
[0054] In one embodiment, maxtime may be set to 365. When recommending courseware, courseware templates may be recommended to the first user based only on the history courseware produced by the first user within 365 days.
[0055] Based on the above method, the user style vector of the second user [xt1, xt2, xt3, ... xt n ] and user style tags [lt1,lt2,lt3,...lt n ], and the user style vector of the second user [xt1,xt2,xt3,...xt n ] and user style tags [lt1,lt2,lt3,...lt n ] are respectively used as the input training data and output training data of the second training data set.
[0056] In one embodiment, the user style prediction model may be a classification model constructed using a multi-layer fully connected network.
[0057] In implementation, in order to enable the user style prediction model to fully learn the features and rules in the training data, thereby improving the generalization ability and prediction accuracy of the user style prediction model, the second training data set can be used to train the user style prediction model for N epochs, that is, the second training data set is traversed N times.
[0058] Specifically, when training the user style prediction model, each epoch will pass all the training data through the model once and update the model's parameters to minimize the loss function. After an epoch is completed, the model's learning of the training data will be deepened, but there may be problems of overfitting or underfitting. At this time, through training of multiple epochs, the model can gradually adapt to the training data and gradually improve the generalization ability of the model, thereby avoiding the problem of overfitting or underfitting and obtaining a more stable and reliable model.
[0059] It is worth mentioning that the number of epochs can be determined based on actual conditions and experience, and this application does not limit the value of N epoch.
[0060] In one embodiment, see Figure 6 The object recommendation method provided in this application may include the following steps:
[0061] S602: Acquire thumbnail information of a history object of a first user, and acquire an object style vector of the history object based on the thumbnail information.
[0062] In implementation, the first user may be any user who needs courseware recommendation service, the first user's historical object may be any courseware produced or used by the user, the thumbnail information of the historical object may be a thumbnail of the historical courseware, and the object style vector of the historical object may be a courseware style vector of the historical courseware.
[0063] In one embodiment, the history object is a history courseware of the first user, and one history courseware may include multiple pages; accordingly, the process of obtaining thumbnail information of the history object of the first user in step S602 specifically includes: combining multiple pages of one history courseware into an image, obtaining a thumbnail of the image, and using the thumbnail of the image as thumbnail information of the history courseware.
[0064] In practice, a courseware usually includes multiple pages. In this case, multiple pages of the courseware can be combined into an image, and the thumbnail of the image is used as the thumbnail of the courseware. By combining multiple pages of the same courseware, a more comprehensive and accurate courseware style vector can be obtained. The thumbnail of the courseware can be a matrix of corresponding pixels of the courseware.
[0065] In one embodiment, the process of obtaining the object style vector of the historical object based on the thumbnail information in step S602 specifically includes: inputting the thumbnail information into the object style prediction model, obtaining the output vector of the target level of the object style prediction model, and using the output vector of the target level as the object style vector.
[0066] In implementation, the thumbnail of a courseware can be a matrix of pixels corresponding to the courseware. Accordingly, the pixel matrix of each courseware is input into the classification model of the courseware style prediction model to predict the courseware style label of the courseware. In addition, the output data of the target level of the model can be taken to obtain the courseware style vector of the history courseware.
[0067] By using this embodiment, the output data of the last layer of the courseware style prediction model has undergone multiple layers of feature selection and extraction processes, and the user style prediction model can use these processed features to predict the user style, thereby improving the efficiency and accuracy of the user style prediction model.
[0068] In one embodiment, the object style prediction model can be built using a multi-layer neural convolutional network, and the target layer is a fully connected layer of the multi-layer neural convolutional network.
[0069] In implementation, the courseware style prediction model can be built using a multi-layer neural convolutional network. Similar to the traditional image classification model, the courseware style prediction model can include a convolutional layer, a pooling layer, and a fully connected layer. Correspondingly, the last layer of the multi-layer neural convolutional network classification model can be a fully connected layer, and its output is usually a one-dimensional vector. The length of this vector can represent the number of categories, and each element can represent the probability that the input data belongs to the corresponding category. Of course, different model structures and data sets may result in different specific meanings and lengths of the output vector of the last layer of the model. The meaning and length of the one-dimensional vector output of the last layer of the model can be determined based on the actual model structure and classification task used.
[0070] S604: Calculate a user style vector of the first user according to the multiple object style vectors of the first user, and input the user style vector into a user style prediction model.
[0071] In implementation, multiple history coursewares of the first user may be loaded, the style vectors of each history courseware may be obtained, and the weighted average of the multiple courseware style vectors may be used to obtain the user style vector of the first user. Figure 2 , multiple courseware style vectors of the first user (3 in the figure) can be weighted and averaged to obtain the user style vector of the first user. User style vector [xt1, xt2, xt3, ... xt n ] can be calculated by referring to the above-mentioned embodiment of the user style prediction model in the training stage, which will not be elaborated in this application.
[0072] In one embodiment, the weight of the history courseware may be associated with the production time, and then a weighted average of multiple object style vectors may be calculated, and the weighted average may be used as the user style vector of the first user. Accordingly, the process of calculating the user style vector of the first user based on the multiple object style vectors of the first user in step S604 may specifically include: calculating the weight of each corresponding object style vector based on the production time of each historical object of the first user; calculating the weighted average of the multiple object style vectors of the first user, and using the weighted average as the user style vector.
[0073] In practice, since the courseware produced by users often has a short-term tendency, the styles of courseware in the same period will be more similar. Therefore, the weight of historical courseware is associated with the production time, and the obtained user style vector can more accurately reflect the user's recent preferences. The calculation method of the weight w of each courseware can refer to the embodiment of the user style prediction model in the training stage, and this application will not be repeated here.
[0074] S606: Obtain a user style label output by the user style prediction model to recommend candidate objects corresponding to the user style label to the first user.
[0075] In implementation, after the user style vector of the first user is input into the user style prediction model, the user style prediction model may output a user style label. Thereafter, the corresponding candidate object may be recommended to the user based on the user style label.
[0076] Taking the recommended object as a courseware template as an example, the thumbnail information of the historical object can be a thumbnail of the history courseware, and the object style vector can be a courseware style vector. In an embodiment of the present application, the thumbnail of the history courseware belonging to a certain user can be input into the courseware style prediction model, feature extraction and / or classification decision-making can be performed, and the output of the last layer of the model is taken to obtain the courseware style vector. Then, the user style vector of the user is obtained using multiple courseware style vectors of the same user. The user style vector integrates multiple coursewares produced and / or used by the user, and can more accurately reflect the user's style. Then, after the user style vector is input into the user style prediction model, a more accurate user style label can be output. Furthermore, the alternative courseware recommended to the user based on the user style label can better meet the user's current preferences and improve the accuracy of courseware recommendations.
[0077] In one embodiment, recommending candidate objects corresponding to the user style label to the first user specifically includes: obtaining multiple candidate objects and calculating the similarity between the object style label of each candidate object and the user style label; and determining a preset number of candidate objects with the highest similarity as candidate objects.
[0078] In the implementation, the courseware style labels [l1,l2,l3,...l n ], and calculate the courseware style label and user style label [lt1, lt2, lt3, ... lt n ] similarity, the higher the similarity, the closer the style is, and the higher the recommendation priority is. For example, the cosine similarity can be calculated, and the closer the cosine similarity is to 1, the higher the recommendation priority is.
[0079] In one embodiment, after step S606, the object recommendation method provided by the present application may also include: obtaining the object style vector of the target candidate object selected by the first user, calculating the user style vector of the first user based on multiple object style vectors of the first user, and using the user style vector for the next object recommendation process.
[0080] In implementation, for ease of description, any candidate object selected by the first user from multiple candidate objects can be referred to as a target candidate object. Correspondingly, when the candidate object is a candidate courseware, any candidate courseware selected by the first user from multiple candidate courseware can be referred to as a target candidate courseware. This embodiment can use the target candidate courseware selected by the first user to retrain the user style prediction model. Specifically, a thumbnail of the target candidate courseware can be input into the courseware style prediction model to obtain a new courseware style vector. Afterwards, the user style vector can be calculated based on multiple courseware style vectors including the new courseware style vector, and the user style prediction model can be fine-tuned to obtain a new user style prediction model. Afterwards, when a new user performs style prediction, or an old user performs style prediction again, the new user style prediction model will more accurately describe the user style.
[0081] It is worth mentioning that the history courseware of the first user can be obtained from the courseware production platform. If the first user is a new user who has not produced courseware on the courseware production platform, a plurality of courseware templates of different styles can be provided to the new user for selection at the beginning, and one or more courseware templates initially selected by the new user are used as the history courseware of the new user, and then suitable alternative courseware is recommended to the new user based on these history coursewares.
[0082] In one embodiment, when calculating the user style vector of the first user, the weight of the object style vector of the target candidate object selected by the first user is not higher than a preset weight value.
[0083] In this embodiment, the courseware template selected by the first user will be given a smaller weight before being fed back to the user style prediction model. In this way, user feedback can be used to optimize the user style prediction model, and the adverse effects of accidental user operations on the user style prediction model can also be weakened.
[0084] It is worth mentioning that if two different users choose the same courseware template, when predicting the user style of these two users, since the time for completing the production of multiple history coursewares of different users is usually not exactly the same, even if different users have produced similar courseware or selected the same courseware template, the weighted average of the multiple courseware style vectors of different users will result in differences in the calculated user style vectors. Furthermore, the user style prediction model can predict different user style labels. In this way, the coverage of the label system is further expanded, and the user style can be described more accurately, so that the courseware template that satisfies the user can be recommended.
[0085] It should be noted that due to space limitations, not all implementation methods are listed in this application. As long as the features are not contradictory, they can be freely combined to become optional implementation methods of this application.
[0086] Based on the same technical concept, the present application also provides an object recommendation device, see Figure 7 , the object recommendation device may include:
[0087] An acquisition module, configured to acquire thumbnail information of a history object of a first user, and acquire an object style vector of the history object based on the thumbnail information;
[0088] A calculation module, configured to calculate a user style vector of the first user according to the plurality of object style vectors of the first user, and input the user style vector into a user style prediction model;
[0089] The recommendation module is used to obtain the user style label output by the user style prediction model to recommend candidate objects corresponding to the user style label to the first user.
[0090] Furthermore, the computing module is specifically used for:
[0091] Based on the production time of each historical object of the first user, calculate the weight of each corresponding object style vector;
[0092] A weighted average of the plurality of object style vectors of the first user is calculated, and the weighted average is used as the user style vector.
[0093] Furthermore, the object recommendation device also includes a training module, which is used to:
[0094] Based on the first training data set of the object style prediction model, construct a second training data set of the user style prediction model;
[0095] The user style prediction model is trained using the second training data set.
[0096] Furthermore, the training module is specifically used to:
[0097] Acquire a first training data set; wherein the input training data of the first training data set is the abbreviated information of the historical objects of the second user, and the output training data of the first training data set is the object style label of the historical objects of the second user;
[0098] Inputting the input training data of the first training data set into the object style prediction model to obtain an object style vector output by a target level of the object style prediction model;
[0099] Calculating a user style vector of the second user according to the multiple object style vectors of the second user, and calculating a user style label of the second user according to the multiple object style labels of the second user;
[0100] The user style vector and the user style label of the second user are respectively used as input training data and output training data of the second training data set.
[0101] Furthermore, the object recommendation device further includes a feedback module, which is used to:
[0102] The object style vector of the target candidate object selected by the first user is obtained, a user style vector of the first user is calculated according to the multiple object style vectors of the first user, and the user style vector is used for the next object recommendation process.
[0103] Further, when calculating the user style vector of the first user, the weight of the object style vector of the target candidate object selected by the first user is not higher than a preset weight value.
[0104] Furthermore, the history object is a history courseware of the first user, and the history courseware includes multiple pages; the acquisition module is specifically used for:
[0105] Combine multiple pages into one image, obtain thumbnails of the image, and use the thumbnails of the image as thumbnail information of history courseware.
[0106] Furthermore, the acquisition module is also specifically used for:
[0107] The abbreviated information is input into the object style prediction model, an output vector of a target level of the object style prediction model is obtained, and the output vector of the target level is used as the object style vector.
[0108] Furthermore, the object style prediction model is built using a multi-layer neural convolutional network, and the target layer is the fully connected layer of the multi-layer neural convolutional network.
[0109] Furthermore, the user style prediction model is built using a multi-layer fully connected network.
[0110] Furthermore, the recommended module is specifically used for:
[0111] Obtain multiple objects to be selected, and calculate the similarity between the object style label of each object to be selected and the user style label;
[0112] A preset number of candidate objects with the highest similarity are determined as candidate objects.
[0113] It should be noted that: the object recommendation device provided in the above embodiment only uses the division of the above functional modules as an example when performing object recommendation. In actual applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the object recommendation device is divided into different functional modules to complete all or part of the functions described above. In addition, the object recommendation device provided in the above embodiment and the object recommendation method embodiment belong to the same concept, and the specific implementation process can be found in the method embodiment, which will not be repeated here.
[0114] Based on the same technical concept, the present application embodiment also provides an object recommendation device. Figure 8 The object recommendation device may include a processor and a memory, wherein the memory stores at least one instruction, at least one program, a code set or an instruction set, and the at least one instruction, the at least one program, the code set or the instruction set is loaded and executed by the processor to implement the object recommendation method of any of the above embodiments.
[0115] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course can also be implemented by hardware. Based on this understanding, the above technical solution can essentially or contribute to the prior art in the form of a software product, and the software product of the object recommendation method can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including storing a number of instructions to enable an electronic device to execute the methods described in each embodiment or some parts of the embodiments.
[0116] The above description is only a preferred embodiment of the present application and is not intended to limit the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present application should be included in the protection scope of the present application.
Claims
1. An object recommendation method, characterized in that: The method comprises: Acquire thumbnail information of a history object of a first user, and acquire an object style vector of the history object based on the thumbnail information; Calculating a user style vector of the first user according to the multiple object style vectors of the first user, and inputting the user style vector into a user style prediction model; A user style label output by the user style prediction model is obtained to recommend candidate objects corresponding to the user style label to the first user.
2. The object recommendation method according to claim 1, characterized in that: The calculating the user style vector of the first user according to the multiple object style vectors of the first user specifically includes: Calculating the weight of each corresponding style vector of the object based on the creation time of each of the historical objects of the first user; A weighted average of a plurality of object style vectors of the first user is calculated, and the weighted average is used as the user style vector.
3. The object recommendation method according to claim 1, characterized in that: Before using the user style prediction model to recommend the candidate object to the first user, the method further includes: Constructing a second training data set for the user style prediction model based on the first training data set for the object style prediction model; The user style prediction model is trained using the second training data set.
4. The object recommendation method according to claim 3, characterized in that: Constructing a second training data set for the user style prediction model based on the first training data set for the object style prediction model specifically includes: Acquire the first training data set; wherein the input training data of the first training data set is the abbreviated information of the historical objects of the second user, and the output training data of the first training data set is the object style label of the historical objects of the second user; Inputting the input training data of the first training data set into the object style prediction model to obtain an object style vector output by a target layer of the object style prediction model; Calculating a user style vector of the second user according to the multiple object style vectors of the second user, and calculating a user style label of the second user according to the multiple object style labels of the second user; The user style vector and the user style label of the second user are used as input training data and output training data of the second training data set, respectively.
5. The object recommendation method according to claim 1, characterized in that: After recommending candidate objects corresponding to the user style label to the first user, the method further includes: The object style vector of the target candidate object selected by the first user is obtained, a user style vector of the first user is calculated according to the multiple object style vectors of the first user, and the user style vector is used for the next object recommendation process.
6. The object recommendation method according to claim 5, characterized in that: When calculating the user style vector of the first user, the weight of the object style vector of the target candidate object selected by the first user is not higher than a preset weight value.
7. The object recommendation method according to claim 1, characterized in that: The history object is the history courseware of the first user, and the history courseware includes a plurality of pages; The obtaining of the abbreviated information of the historical objects of the first user specifically includes: The plurality of pages are combined into one image, a thumbnail of the image is obtained, and the thumbnail of the image is used as thumbnail information of the history courseware.
8. The object recommendation method according to claim 1, characterized in that: The acquiring the object style vector of the historical object based on the thumbnail information specifically includes: The abbreviated information is input into an object style prediction model, an output vector of a target level of the object style prediction model is obtained, and the output vector of the target level is used as the object style vector.
9. The object recommendation method according to claim 8, characterized in that: The object style prediction model is built using a multi-layer neural convolutional network, and the target layer is a fully connected layer of the multi-layer neural convolutional network.
10. The object recommendation method according to claim 1, wherein: The user style prediction model is built using a multi-layer fully connected network.
11. The object recommendation method according to claim 1, wherein: The recommending to the first user a candidate object corresponding to the user style label specifically includes: Acquire multiple objects to be selected, and calculate the similarity between the object style label of each of the objects to be selected and the user style label; A preset number of the candidate objects with the highest similarity are determined as the candidate objects.
12. An object recommendation device, characterized in that: include: An acquisition module, configured to acquire thumbnail information of a history object of a first user, and acquire an object style vector of the history object based on the thumbnail information; a calculation module, configured to calculate a user style vector of the first user according to the plurality of object style vectors of the first user, and input the user style vector into a user style prediction model; A recommendation module is used to obtain the user style label output by the user style prediction model to recommend candidate objects corresponding to the user style label to the first user.
13. An object recommendation device, characterized in that: The object recommendation device comprises: a processor and a memory; wherein the memory stores a computer program, and the computer program is suitable for being loaded by the processor and executing the method according to any one of claims 1-11.
14. A computer-readable storage medium, characterized in that: The computer storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor and executing the method according to any one of claims 1-11.