Online user preference high satisfaction recommendation method integrating product popularity analysis

Abstract

The present invention relates to an online user preference high satisfaction recommendation method integrating commodity popularity analysis, which is used to solve the problems of low accuracy and user satisfaction of object recommendation in the prior art, comprising: obtaining the total amount of interaction between a user and an object and the user's evaluation of the interaction with the object; transforming the total amount of interaction into object popularity through a function and inputting the result into a recommendation model, obtaining high preference user and object representations and obtaining an object recommendation candidate set through calculation; inputting objects divided into several levels into the recommendation model training to obtain high satisfaction user and object representations; respectively fusing high preference user and object representations with high satisfaction user representations and object representations to obtain high preference and high satisfaction user representations and object representations, performing dot product calculations to obtain a prediction score of high preference and high satisfaction of the user for the object, and obtaining a user object recommendation list according to the prediction score, thereby improving the accuracy and satisfaction of recommendations for the user.

Description

Technical Field

[0001] The present invention relates to the field of data analysis, and more specifically, to an online user preference high satisfaction recommendation method integrating commodity popularity analysis. Background Art

[0002] In today's digital age, e-commerce platforms have become one of the main channels for consumers to shop. With the explosive growth of online product categories, users often feel overwhelmed when faced with a huge number of choices. In order to improve user experience and increase conversion rates, e-commerce platforms urgently need accurate recommendation systems to help users discover products they may be interested in. However, although existing recommendation algorithms can make recommendations based on user historical behavior, making recommendations based solely on user historical behavior can easily create an "information cocoon", making it difficult for users to obtain recommendation information other than historical behavior. Therefore, making recommendations based solely on user historical behavior can no longer meet user needs.

[0003] User behavior data covers purchase history, browsing history, search habits and evaluation feedback, and is an important clue to understanding and predicting user preferences. How to make good use of user behavior data to provide users with better recommendation information has become a research focus.

[0004] In addition, although many e-commerce platforms have adopted technologies such as collaborative filtering and content recommendation to provide personalized recommendations, existing systems rarely consider information other than historical behavior as a recommendation dimension, resulting in the current recommendation system being unsatisfactory. Summary of the invention

[0005] The present invention aims to overcome at least one defect (shortcoming) of the above-mentioned prior art and provide a method, apparatus, storage medium and computing device for online user preference high satisfaction recommendation integrating commodity popularity analysis, so as to solve the problems of low accuracy of object recommendation and low user satisfaction in the prior art.

[0006] The technical solution adopted by the present invention is a method for recommending high satisfaction of online user preferences by integrating commodity popularity analysis, the method comprising:

[0007] Obtaining the total amount of interaction between the user and the object and the user's evaluation of the interaction with the object;

[0008] The total amount of interaction behavior is transformed into object popularity through a function and input into a recommendation model to obtain a first user representation with high preference and a first object representation with high preference, and an object recommendation candidate set is obtained based on the first user representation and the first object representation;

[0009] The objects interacted by each user are divided into several levels according to the evaluation, and the objects with a level greater than a preset value are input into the recommendation model training to obtain a second user representation with high satisfaction and a second object representation with high satisfaction;

[0010] The first user representation and the first object representation are respectively fused with the second user representation and the second object representation to obtain a third user representation and a third object representation with high preference and high satisfaction, a dot product operation is performed on the third user representation and the third object representation to obtain a prediction score of the user's high preference and high satisfaction for the object, and an object recommendation list for the user is obtained from the object recommendation candidate set based on the prediction score.

[0011] By obtaining high preference representations through object popularity and optimizing user preferences for satisfaction, we ultimately obtain an object recommendation list. This not only makes the recommendation list consistent with the user's personal preferences, but also has a high user satisfaction, making the recommendation results more accurate.

[0012] Furthermore, the total amount of interaction between the user and the object is obtained by the number of times the user interacts with the object. The expression of the interaction between the user and the object is:

[0013]

[0014] in, Represents the user, Represents an object, Indicates user With Object The interactive relationship, Indicates user With Object Interaction occurs, Indicates user With Object No interaction occurred.

[0015] By defining the interaction behavior expression between the user and the object, it is convenient to obtain the total amount of interaction behavior between the user and the object.

[0016] Furthermore, the total amount of the interactive behavior is transformed into the object popularity through a function, and the function is specifically:

[0017]

[0018] in, Representation Object The popularity of represents the natural base, represents the temperature coefficient, Representing users and objects The total amount of interaction behavior.

[0019] Furthermore, the recommendation method further includes obtaining a user set , object collection The first user representation and the first object representation are obtained by the following steps:

[0020] Form an interaction graph of the interaction relationships between all users and objects , interaction diagram It is expressed as follows:

[0021] ;

[0022] in, Represents a collection of users, Represents a collection of objects;

[0023] Set both user and object as nodes, nodes h The set of neighbor nodes definition

[0024]

[0025] in, is any node in the interaction graph G, For Node A neighbor node around Represents a slave node To neighbor node An edge of Representation Node The set of neighbor nodes of

[0026] The popularity of the object is added to the information transmission and aggregation of the recommendation model, and the activation function is applied to obtain the neighbor node set of the node Polymerization characterization :

[0027]

[0028] in, is the activation function, Representation Node The popularity of is a randomly initialized transformation matrix, the modality representing the object's data, Represents neighbor nodes In modal On the representation, Represents a set of neighbor nodes size;

[0029] Obtain the ID information of the node, fuse the representation of the node with the ID information of the node and apply an activation function to obtain a fused vector representation after information fusion , fusion vector representation The expression is;

[0030]

[0031] in, Representation Node In modal Representation Fusion Node on The fusion vector representation obtained by the ID information is is a randomly initialized transformation matrix, Representation Node In modal On the representation, Representation Node ID information;

[0032] The node Perform multi-stage training to obtain multi-stage representations after training;

[0033] The node In modal The first-order characterization after the first-order connectivity on is:

[0034] in, is a randomly initialized transformation matrix, Representation Node In modal First-order representation on ;

[0035] node In modal on After stage connectivity The order representation is:

[0036]

[0037] in, Representation Node In modal on Level representation, Indicates that in modal on Randomly initialized transformation matrix of order; Representation Node In modal on After the stage connectivity, the ID information of the fused nodes is obtained The fusion vector representation of the order, Order fusion vector representation It is expressed as:

[0038]

[0039] The node Each mode of The data passed Stage training and fusion of each modality The representation of multimodal fusion is obtained for:

[0040]

[0041] in, Representation Node Multimodal fusion representation of Representation Node In modal on order representation;

[0042] Connect the representations obtained after training in each stage to get high preference users First user characterization and high preference objects First object representation .

[0043] By adding the popularity of objects to the information transmission and aggregation of the recommendation model, it is possible to more comprehensively capture user preferences and improve the accuracy of recommendations. By calculating the high-order connectivity of nodes, the recommendation model can more accurately predict user preferences, thereby improving the accuracy of the recommendation model.

[0044] Furthermore, the recommendation model includes an attention mechanism, which adds the popularity of the object to the information transmission and aggregation of the recommendation model, and applies an activation function to obtain the node The set of neighbor nodes Polymerization characterization , specifically:

[0045]

[0046] in, is the attention mechanism, is an adjustable parameter, is a randomly initialized transformation matrix, Represents the mode, Representation Object In modal On the representation, The corresponding features and vector representations are obtained by extracting preprocessing models of different modalities.

[0047] The recommendation model includes an attention mechanism. By adding object popularity to the transmission and aggregation of the recommendation model, the model can dynamically adjust the weight of features according to changes in user interaction behavior, so that the recommendation model can capture changes in user interests more timely and make recommendations that are more in line with user preferences. By introducing the attention mechanism, the recommendation model can better adjust the weight of features for different users and recommendation scenarios, thereby improving the generalization ability of the recommendation model.

[0048] Furthermore, the second user representation and the second object representation are obtained by the following steps:

[0049] Information transmission and aggregation through the recommendation model and application of activation function Get Node The set of neighbor nodes In modal m Aggregate representation on ;

[0050] When the recommendation model includes an attention mechanism, the aggregate representation It is expressed as:

[0051]

[0052] in, For Node The set of neighbor nodes Aggregate representation on modality m, is the attention mechanism, For in modal A randomly initialized transformation matrix, For Node Neighboring nodes In modal Representation of;

[0053] When the recommendation model does not include an attention mechanism, the aggregate representation It is expressed as:

[0054]

[0055] Get Node The ID information of each node Eigenvector Fusion Node After the ID information is obtained, the activation function is applied to obtain the information fusion vector , expressed as:

[0056]

[0057] in, is a randomly initialized transformation matrix, Representation Node ID;

[0058] The node Perform stage training and obtain nodes Each mode of go through The high-level representation after stage training is Higher-order representations behind stage connectivity for:

[0059]

[0060] Fusion modality collection Each mode of The high-order representation of is obtained by multimodal fusion representation for:

[0061]

[0062] in, is the fusion modal set of Level representation, For Node The modal set of data;

[0063] Connect the node representations obtained after training in each stage to get different user Second user representation and different objects The second object representation .

[0064] Furthermore, the third user representation and the third object representation are obtained specifically through the following steps:

[0065] The second user characterization and the first user characterization Perform superposition fusion to represent the second object and the first object characterization Perform superposition fusion to obtain the third user representations with high satisfaction and high preference respectively. and third object representation :

[0066]

[0067]

[0068] in, ∈Cands, Cands is the candidate set recommended for the object.

[0069] Furthermore, the prediction score of the user's high preference and high satisfaction for the object obtained by performing a dot product operation based on the third user representation and the third object representation is specifically:

[0070] The third user characterization With the third object characterization Perform inner product operation to obtain the predicted score of users with high satisfaction and high preference for the object , prediction score The calculation formula is as follows:

[0071] .

[0072] According to another aspect of an embodiment of the present application, there is provided an online user preference high satisfaction recommendation device integrating commodity popularity analysis, the device comprising:

[0073] A user and object information acquisition module, the acquisition module is used to acquire the total amount of interaction between the user and the object and the user's evaluation of the interaction with the object;

[0074] A user preference acquisition module, which is used to transform the total amount of interaction behavior into an influence of object popularity through a function and add it into a recommendation model to obtain a representation of the user's high preference, and obtain a candidate set of object recommendations through the operation of the representation of the user's high preference;

[0075] A satisfaction acquisition module, which is used to divide the objects interacted by each user into objects with high satisfaction and objects with low satisfaction according to the evaluation, and put the objects with high satisfaction into the recommendation model for training to obtain the representation of the high satisfaction of the user;

[0076] The recommendation result acquisition module is used to superimpose and fuse the representations of the user's high preference and high satisfaction to obtain the representations of the user's high preference and high satisfaction, and reorder the candidate set according to the representations of the user's high preference and high satisfaction to obtain the set of items that need to be recommended.

[0077] According to another aspect of an embodiment of the present application, a readable storage medium is provided, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned online user preference high satisfaction recommendation method integrating commodity popularity analysis is implemented.

[0078] Compared with the prior art, the present invention has the following beneficial effects:

[0079] According to the online user preference high-satisfaction recommendation method integrated with commodity popularity analysis provided by the embodiment of the present invention, the user's high-preference recommendation candidate set is obtained by obtaining the historical interaction behavior between the user and the object and combining it with the popularity of the commodity, and the user's high-preference list is optimized for satisfaction to obtain a candidate set with high preference and high satisfaction, so that the recommendation result not only meets the user's personal preferences, but also has a high user satisfaction, thereby improving the accuracy of personalized recommendations for users and the user's satisfaction, and effectively improving the user's usage experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0080] Figure 1 Schematic diagram of an application scenario of an embodiment of the present invention.

[0081] Figure 2 This is a flow chart of the online user preference high satisfaction recommendation method integrating commodity popularity analysis according to Example 1 of the present invention.

[0082] Figure 3 This is a schematic diagram of the structure of an online user preference high satisfaction recommendation device integrating commodity popularity analysis according to Embodiment 3 of the present invention. DETAILED DESCRIPTION

[0083] The drawings of the present invention are only for illustrative purposes and should not be construed as limiting the present invention. In order to better illustrate the following embodiments, some parts of the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product; it is understandable to those skilled in the art that some well-known structures and their descriptions in the drawings may be omitted.

[0084] Example 1

[0085] Product popularity refers to the popularity of a product in the market, which can be measured by multiple indicators such as sales, search frequency, and social media discussion. According to a McKinsey survey, popular products can often attract up to 30% more additional traffic, and the traffic conversion rate of popular products is about 50% higher than that of non-popular products. Therefore, incorporating product popularity into the recommendation system can significantly improve the attractiveness and effectiveness of recommendations.

[0086] User preference and satisfaction are two important considerations in the recommendation system. They are both independent and closely related. User preference is usually based on the user's long-term interests and purchasing habits, while satisfaction reflects more of the user's immediate experience and emotional feedback. In order to provide more accurate and personalized recommendations, a dual-objective recommendation system can consider both aspects at the same time. User satisfaction is crucial to the user stickiness and conversion rate of e-commerce platforms. According to Bain & Company's analysis, a 5% increase in user satisfaction can bring a 1.5-fold increase in revenue. Therefore, improving user satisfaction by optimizing the recommendation system can not only improve the user's shopping experience, but also attract more repeat customers to the e-commerce platform, thereby driving sales growth.

[0087] This embodiment provides a technical solution for solving the above-mentioned problems by taking into account both user preferences and satisfaction. The specific implementation methods of this application are described in detail below in conjunction with the accompanying drawings.

[0088] For example, Figure 1 A schematic diagram of an application scenario of an online user preference high satisfaction recommendation method integrating product popularity analysis provided in an embodiment of the present application, such as Figure 1 As shown, it includes a server 100 and a terminal 200. The server 100 can be connected by wireless communication or by wired communication.

[0089] It is understandable that the server 100 can be an independent electronic device, a cluster or distributed system composed of multiple electronic devices, or a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, network services, middleware services, domain name services, big data and artificial intelligence. The operating systems running on the server include but are not limited to Windows, Linux, iOS, Android, etc.

[0090] The terminal 200 may be a smart phone terminal, a personal computer, a tablet computer, a vehicle-mounted terminal, an augmented reality or virtual reality device, a smart wearable device, etc., but is not limited thereto.

[0091] In some embodiments, the server 100 and the terminal 200 may separately and independently execute the online user preference high satisfaction recommendation method integrating product popularity analysis provided in the embodiment of the present application, or, optionally, the online user preference high satisfaction recommendation method integrating product popularity analysis provided in the embodiment of the present application is partially executed in the server 100 and partially executed in the terminal 200. This embodiment does not limit the specific application environment.

[0092] Figure 2 A flowchart of a method for recommending high satisfaction of online user preferences by integrating commodity popularity analysis is provided for this embodiment. Figure 2As shown, the online user preference high satisfaction recommendation method integrating commodity popularity analysis includes the following steps:

[0093] S1. Obtain the total amount of interaction between the user and the object and the user's evaluation of the interaction with the object.

[0094] The online user preference high satisfaction recommendation method integrated with commodity popularity analysis provided by the present invention is applicable to recommendation scenarios of various types of items or content, for example, commodity recommendations in scenarios such as e-commerce platforms or shopping websites, song recommendations in services such as music platforms or online music players, and video recommendations in video website scenarios. In this embodiment, the interaction between users and objects includes but is not limited to users purchasing commodities, users playing videos or music, etc. Taking the e-commerce platform as an example, the total amount of interaction between users and objects is the total transaction volume of users purchasing commodities. The evaluation of users' interaction with the objects can be calculated by obtaining the user's rating of the commodities they purchased. In actual operation, it can be calculated according to the evaluation criteria of commodities on different e-commerce platforms.

[0095] Exemplarily, this embodiment takes an e-commerce platform scenario as an example to illustrate the online user preference high satisfaction recommendation method integrated with product popularity analysis provided by the present invention.

[0096] Step S1 also includes obtaining a user set , object collection , the set of interaction behaviors between each user and the object and the evaluation criteria for the corresponding objects .

[0097] The interaction relationship between the user and the object is set to , when the user interacts with the object, =1, otherwise =0, the interaction between the user and the object can be expressed as follows:

[0098]

[0099] in, Represents the user, Represents an object, Indicates user With Object The interactive relationship, Indicates user With Object Interaction occurs, Indicates user With Object No interaction occurred.

[0100] With each user A collection of objects that interact It is expressed as: ,in Represents an object, Represents a collection of objects.

[0101] In this embodiment, an interaction graph is used to indicate which users have interacted with which objects. The interaction graph contains the connection relationship between user nodes and object nodes. If a user interacts with an object, the user node is connected to the object node. Let G be the interaction graph between users and objects. In the interaction graph G, users and interacting objects form edges. Each edge represents an interaction between a user and an object. All users and objects are nodes, and all interactions between users and objects form an interaction graph G. The interaction graph G is represented as follows:

[0102]

[0103] in, Represents a collection of users, Represents a collection of objects.

[0104] In practical applications, objects in the real world usually contain data in multiple modes, such as audio, text, images, etc. These different modes of data each provide information about different aspects of the object. Integrating data from different modes together can provide richer information and help to more accurately characterize the characteristics of the object. In this embodiment, Represents the data of any modality of the object. The modal set composed of all modalities of the object's data is .

[0105] The following will introduce the required variables in the recommended model so that those skilled in the art can better understand the subsequent related content. The required variables in the recommended model are:

[0106]

[0107] in, Indicates user ID, Representation Object ID, Indicates user In modal On the representation, Representation Object In modal On the representation, Indicates user To the object Evaluation, For all users and objects The total amount of interactive behaviors. The corresponding features and vector representations can be extracted through preprocessing models of different modalities.

[0108] S2. The total amount of interaction behaviors is transformed into object popularity through a function and input into a recommendation model to obtain a first user representation with high preference and a first object representation with high preference, and an object recommendation candidate set is obtained based on the first user representation and the first object representation.

[0109] In this embodiment, the total amount of interaction behavior of each object is an explicit expression of its popularity. The popularity characteristics of an object include the potential attractiveness of the object and the influence of the herd factor on users. Considering the marginal effect and the smooth line of the function, in this embodiment, the popularity of an object is represented by the following function:

[0110]

[0111] in, Representation Object The popularity of Represents temperature coefficient.

[0112] Temperature coefficient is a hyperparameter used to add a small positive offset when calculating the logarithmic function by adding a temperature coefficient to the logarithmic function To adjust the shape of the logarithmic function curve to change the function's response to input data Sensitivity to change.

[0113] In the specific implementation process, all users and objects can be The total amount of interaction The function is appropriately smoothed or scaled according to the size and needs of different models.

[0114] After calculating the popularity of each object, the popularity is input into the recommendation system. h The set of neighbor nodes Defined as:

[0115]

[0116] in, is any node in the interaction graph G, For Node A neighbor node around Represents a slave node To neighbor node An edge of Representation Node The set of neighbor nodes.

[0117] The popularity of the object in this embodiment can be introduced into the recommendation model including the attention mechanism or the recommendation model without the attention mechanism. For the recommendation model without the attention mechanism, the popularity of the object is input into the information transmission and aggregation of the model and the activation function is added. , get the neighbor node set In modal m Aggregate representation on :

[0118]

[0119] in, Representation Node The popularity of is a randomly initialized transformation matrix, the modality representing the object's data, Represents neighbor nodes In modal On the representation, Represents a set of neighbor nodes size.

[0120] For the recommendation model that introduces the attention mechanism, the weight ratio of popularity is determined based on the attention mechanism, and the popularity is combined with the determined weight ratio to input the information transmission and aggregation of the model and add an activation function , get the node's neighbor node set Polymerization characterization :

[0121]

[0122] in, is the attention mechanism, is an adjustable parameter, is a randomly initialized transformation matrix, The modality of the data representing the node, Representation Object In modal In the actual implementation process, The corresponding features and vector representations can be extracted through preprocessing models of different modalities. The attention mechanisms used by different models can be different. For different attention mechanisms, the parameters can be adjusted The value of is used to find the most appropriate weight ratio.

[0123] Any node In modal The representation of , the node Characterization With Node ID information Fusion and application of activation function to obtain the fusion vector representation after information fusion , fusion vector representation The expression is:

[0124]

[0125] in, is a randomly initialized transformation matrix, Representation Node In modal On the representation, Representation Node ID information, Representation Node In modal Representation Fusion Node on Regardless of whether the recommendation model introduces the attention mechanism, the introduction of popularity will not change each node information fusion.

[0126] Higher-order connectivity means that any two nodes can be connected through multiple intermediate nodes. For example, first-order connectivity means that a node is directly connected to its neighbors, second-order connectivity means that any two nodes are connected through one intermediate node, and so on. m-order connectivity means that any two nodes are connected through m-1 intermediate nodes.

[0127] The node Training is performed after each order of connectivity to obtain a multi-stage representation after training.

[0128] node In modal The first-order characterization after the first-order connectivity on is:

[0129]

[0130] in, is a randomly initialized transformation matrix.

[0131] Is a node The initial vector representation is: .

[0132] Similarly, nodes In modal on After stage connectivity The order representation is:

[0133]

[0134] in, Representation Node In modal on Level representation, Indicates that in modal on The randomly initialized transformation matrix of each order can be obtained by random initialization. Representation Node In modal on After the stage connectivity, the ID information of the fused nodes is obtained Order fusion vector representation.

[0135] Order fusion vector representation It is expressed as:

[0136]

[0137] in, is the initial fusion vector representation ,Right now: .

[0138] In this embodiment, each mode Up-conversion matrix , , The value of is randomly initialized each time, so the transformation matrix , , The value of may be different, and the connectivity in different stages The values ​​of the transformation matrix may also be different.

[0139] The node Each mode of The data have been After the stage training, the multimodal fusion of the final node obtains the node The characterization is:

[0140]

[0141] After the above model transformation, the node representations obtained after each stage of training are connected as the final node representation to obtain high preference users. First user characterization and high preference objects First object representation :

[0142]

[0143]

[0144] The first user and the first object Get the user through dot product calculation To the object Preference score :

[0145]

[0146] Get preference score Finally, the model parameters of the object recommendation model are adjusted through the loss function. The loss function used by the model is:

[0147]

[0148] in, , Indicates user With Object Interaction occurs, Indicates user Not with the object Interaction occurs, represents the Sigmoid function, is the attenuation factor in the model, are the trainable parameters used in the model.

[0149] In actual application, according to different recommendation requirements, the user's preference score for the object can be obtained through the representation of different users and different objects, and a candidate set of object recommendation lists that meet the requirements can be obtained based on the preference score. For example, if the model recommends K objects, the user's preference score for the object can be obtained through the inner product operation of the user representation and the representation of different objects. Based on these preference scores, the corresponding preference score can be obtained. The object is taken as the candidate set, and the obtained candidate set is expressed as follows:

[0150]

[0151] in, Before scoring preference object, express Candidate sets.

[0152] S3, classifying each user-interacted object into several levels according to the evaluation, inputting the objects with a level greater than a preset value into the recommendation model training, and obtaining a second user representation with high satisfaction and a second object representation with high satisfaction.

[0153] The model used for training the representation of satisfaction is the same as the model used by the recommendation system for training the representation of high preference, but the training objectives and data are different. Specifically, the evaluations of users and objects are normalized, and the objects are divided according to the normalized evaluations. Objects with a rating higher than a certain preset value are regarded as objects with high user satisfaction, so as to perform the representation training of user satisfaction.

[0154] For example, assuming that the object evaluation score is , the preset evaluation is greater than or equal to The object is a high satisfaction object, then the high satisfaction object The representation is as follows:

[0155]

[0156] in, For the object, For users For objects 's evaluation.

[0157] Highly satisfied objects The input does not consider the popularity of the recommendation model for training, and the model uses The loss function is as follows:

[0158]

[0159] in, .

[0160] The difference between the representation training of satisfaction and the representation training of user preference is that the model's information transmission and aggregation, information fusion, stage fusion, and modal fusion processes do not input the popularity of the object for training.

[0161] For recommendation models that do not introduce attention mechanisms, the activation function is applied through information transmission and aggregation of the model , get the node The set of neighbor nodes In modal Aggregate representation on :

[0162]

[0163] in, For Node The set of neighbor nodes In modal Aggregate representation on is a randomly initialized transformation matrix, For Node Neighboring nodes In modal Due to the user's preference The loss function is different from the data of the two, and the transformation matrix in training The value and transformation matrix Also different.

[0164] For the recommendation model that introduces the attention mechanism, the information transmission and aggregation of the model and the application of the activation function , get the node The set of neighbor nodes In modal Aggregate representation on :

[0165]

[0166] in, For the attention mechanism.

[0167] Get Node The ID information of each node Eigenvector Fusion Node After the ID information is obtained, the activation function is applied to obtain the information fusion vector , the information fusion vector It is expressed as:

[0168]

[0169] in, is a randomly initialized transformation matrix, Representation Node In modal On the representation, Representation Node ID information.

[0170] The node Perform stage training and obtain nodes Each mode of go through The high-level representation after stage training is Higher-order representations behind stage connectivity for:

[0171]

[0172] node Each mode of The data passed After stage training, the final fusion modality set Each mode of The high-order representation of is obtained by multimodal fusion representation for:

[0173]

[0174] in, is the fusion modal set of Level representation, For Node The modal collection of data.

[0175] Through the above model conversion, the nodes trained in each stage are connected as the final node representation to obtain the second user representation of users with high satisfaction. and the corresponding second object representation of the highly satisfactory object under satisfaction supervision :

[0176]

[0177]

[0178] S4. The first user representation and the first object representation are respectively merged with the second user representation and the second object representation to obtain a third user representation and a third object representation with high preference and high satisfaction, a dot product operation is performed on the third user representation and the third object representation to obtain a prediction score of the user's high preference and high satisfaction for the object, and an object recommendation list for the user is obtained from the object recommendation candidate set based on the prediction score.

[0179] To make high satisfaction recommendations in the user's highly preferred object candidate set, it is necessary to superimpose and fuse the high satisfaction user representation with the high preference user representation. and the first user characterization Perform superposition fusion to represent the second object and the first object characterization Perform superposition fusion to obtain the third user representations with high satisfaction and high preference respectively. and third object representation :

[0180]

[0181]

[0182] in, ∈Cands, Cands is the candidate set recommended for the object.

[0183] Get the third user representation Then, the third user characterization With the third object characterization Perform inner product operation to obtain the predicted score of users with high satisfaction and high preference for the object , prediction score The calculation formula is as follows:

[0184]

[0185] For the recommended candidate set with high satisfaction The recommended objects in can be based on the predicted scores Reorder the corresponding objects by their height:

[0186]

[0187] in, is the sorting function, represents the civil engineering candidate set, For the objects.

[0188] After obtaining the object list by sorting the predicted scores, let the object to be recommended to the user be If The objects are used as the recommended object list, and the ranking is the top The object list is a recommended list of objects with high preference and high satisfaction.

[0189]

[0190] in, Recommend a list for an object, To get the top ranking A function of an object.

[0191] Example 2

[0192] This embodiment provides a computer-readable storage medium on which a computer program is stored. When the computer program is executed, a 3D gesture real-time recognition method as described in Embodiment 1 is implemented.

[0193] Optionally, the computer-readable storage medium may be a random access memory, a read-only memory, a programmable read-only memory, an erasable programmable read-only memory, an electrically erasable programmable read-only memory, or the like.

[0194] Example 3

[0195] This embodiment provides an object recommendation device 300, which includes a user and object information acquisition module 301, a user preference acquisition module 302, a satisfaction acquisition module 303, and a recommendation result acquisition module 304. Among them:

[0196] The user and object information acquisition module 301 is used to acquire the total amount of interaction between the user and the object and the user's evaluation of the interaction with the object;

[0197] The user and object information acquisition module 301 is used to execute Figure 2 As shown in step S1 , for a detailed description of the user and object information acquisition module 301 , reference may be made to the description of step S1 .

[0198] The user preference acquisition module 302 is used to transform the total amount of interaction behavior into object popularity through a function and input it into the recommendation model, obtain a high-preference first user representation and a high-preference first object representation, and obtain an object recommendation candidate set based on the first user representation and the first object representation.

[0199] The user preference acquisition module 302 can be used to perform Figure 2 Regarding the step S2, the specific description of the user preference acquisition module 302 can refer to the description of the step S2;

[0200] The satisfaction acquisition module 303 is used to classify each user-interacted object into several levels according to evaluation, and input objects with a level greater than a preset value into the recommendation model training to obtain a second user representation with high satisfaction and a second object representation with high satisfaction.

[0201] The satisfaction acquisition module 303 can be used to perform Figure 2 Regarding the step S3, the specific description of the satisfaction acquisition module 303 can refer to the description of the step S3;

[0202] The recommendation result acquisition module 304 is used to merge the first user representation and the first object representation with the second user representation and the second object representation respectively to obtain a third user representation and a third object representation with high preference and high satisfaction, perform a dot product operation on the third user representation and the third object representation to obtain a prediction score of the user's high preference and high satisfaction for the object, and obtain a user's object recommendation list from the object recommendation candidate set based on the prediction score. The recommendation result acquisition module 304 can be used to perform Figure 2 Regarding step S4, the specific description of the recommendation result acquisition module 304 may refer to the description of step S4.

[0203] It can be understood that the above-mentioned device embodiments and the above-mentioned method embodiments can correspond to each other, and similar descriptions of the device embodiments can refer to the method embodiments. To avoid repetition, it will not be repeated here. An object recommendation device provided in an embodiment of the present application can execute an online user preference high satisfaction recommendation method that integrates commodity popularity analysis provided in any embodiment of the present application, and has functional modules and beneficial effects corresponding to the execution method. The functional modules of the object recommendation device can be implemented in the form of hardware, can be implemented by instructions in the form of software, and can also be implemented by a combination of hardware and software modules.

[0204] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the technical solution of the present invention, and are not intended to limit the specific implementation methods of the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the claims of the present invention shall be included in the protection scope of the claims of the present invention.

Claims

1. A method for online user preference recommendation with high satisfaction based on commodity popularity analysis, characterized in that: The method comprises: Obtaining the total amount of interaction between the user and the object and the user's evaluation of the interaction with the object; The total amount of interaction behavior is transformed into object popularity through a function and input into a recommendation model to obtain a first user representation with high preference and a first object representation with high preference, and an object recommendation candidate set is obtained based on the first user representation and the first object representation; The objects interacted by each user are divided into several levels according to the evaluation, and the objects with a level greater than a preset value are input into the recommendation model training to obtain a second user representation with high satisfaction and a second object representation with high satisfaction; fusing the first user representation and the first object representation with the second user representation and the second object representation respectively to obtain a third user representation and a third object representation with high preference and high satisfaction, performing a dot product operation on the third user representation and the third object representation to obtain a prediction score of the user's high preference and high satisfaction for the object, and obtaining an object recommendation list for the user from the object recommendation candidate set based on the prediction score; The total amount of interaction between the user and the object is obtained by the number of times the user and the object interact. The expression of the interaction between the user and the object is: in, Represents the user, Represents an object, Indicates user With Object The interactive relationship, Indicates user With Object Interaction occurs, Indicates user With Object No interactive behavior occurred; The total amount of the interactive behavior is transformed into the object popularity through a function, and the function is specifically: in, Represents any object The popularity of represents the natural base, represents the temperature coefficient, Representing users and objects The total amount of interaction behavior; The recommendation method also includes obtaining a user set , object collection The first user representation and the first object representation are obtained by the following steps: Form an interaction graph of the interaction relationships between all users and objects , interaction diagram It is expressed as follows: ; in, Represents a collection of users, Represents a collection of objects; Set users and objects as nodes, and the neighbor node set of node h definition in, is any node in the interaction graph G, For Node A neighbor node around Represents a slave node To neighbor node An edge of Representation Node The set of neighbor nodes of The popularity of the object is added to the information transmission and aggregation of the recommendation model, and the activation function is applied to obtain the neighbor node set of the node Polymerization characterization : in, is the activation function, Representation Node The popularity of is a randomly initialized transformation matrix, the modality representing the object's data, Represents neighbor nodes In modal On the representation, Represents a set of neighbor nodes size; Obtain the ID information of the node, fuse the representation of the node with the ID information of the node and apply an activation function to obtain a fused vector representation after information fusion , fusion vector representation The expression is; in, Representation Node In modal Representation Fusion Node on The fusion vector representation obtained by the ID information is is a randomly initialized transformation matrix, Representation Node In modal On the representation, Representation Node ID information; The node Perform multi-stage training to obtain multi-stage representations after training; The node In modal The first-order characterization after the first-order connectivity on is: in, is a randomly initialized transformation matrix, Representation Node In modal First-order representation on ; node In modal on After stage connectivity The order representation is: in, Representation Node In modal on Level representation, Indicates that in modal on Randomly initialized transformation matrix of order; Representation Node In modal on After the stage connectivity, the ID information of the fused nodes is obtained Order fusion vector representation, Order fusion vector representation It is expressed as: The node Each mode of The data passed Stage training and fusion of each modality The representation of multimodal fusion is obtained for: in, Representation Node Multimodal fusion representation of Representation Node In modal on Level representation, a modal collection of data representing said object; Connect the representations obtained after training in each stage to get high preference users First user characterization and high preference objects First object representation .

2. According to claim 1, a method for recommending online user preferences with high satisfaction by integrating commodity popularity analysis is characterized in that: The recommendation model includes an attention mechanism, which adds the popularity of the object to the information transmission and aggregation of the recommendation model, and applies an activation function to obtain the node The set of neighbor nodes Polymerization characterization , specifically: in, is the attention mechanism, is an adjustable parameter, is a randomly initialized transformation matrix, Represents the mode, Representation Object In modal On the representation, The corresponding features and vector representations are obtained by extracting preprocessing models of different modalities.

3. The online user preference high satisfaction recommendation method integrating commodity popularity analysis according to claim 1 or 2, characterized in that: The second user representation and the second object representation are obtained by the following steps: Information transmission and aggregation through the recommendation model and application of activation function Get Node The set of neighbor nodes In modal m Aggregate representation on ; When the recommendation model includes an attention mechanism, the aggregate representation It is expressed as: in, For Node The set of neighbor nodes Aggregate representation on modality m, is the attention mechanism, For in modal A randomly initialized transformation matrix, For Node Neighboring nodes In modal Representation of; When the recommendation model does not include an attention mechanism, the aggregate representation It is expressed as: Get Node The ID information of each node Eigenvector Fusion Node After the ID information is obtained, the activation function is applied to obtain the information fusion vector , expressed as: in, is a randomly initialized transformation matrix, Representation Node ID; The node Perform stage training and obtain nodes Each mode of go through The high-level representation after stage training is Higher-order representations behind stage connectivity for: Fusion modality collection Each mode of The high-order representation of is obtained by multimodal fusion representation for: in, is the fusion modality set of Level representation, For Node The modal set of data; Connect the node representations obtained after training in each stage to get different user Second user representation and different objects The second object representation .

4. According to claim 3, a method for recommending online user preferences with high satisfaction by integrating commodity popularity analysis is characterized in that: The third user representation and the third object representation are specifically obtained by the following steps: The second user characterization and the first user characterization Perform superposition fusion to represent the second object and the first object characterization Perform superposition fusion to obtain the third user representations with high satisfaction and high preference respectively. and third object representation : in, ∈Cands, Cands is the candidate set recommended for the object.

5. According to claim 4, a method for recommending online user preferences with high satisfaction by integrating commodity popularity analysis is characterized in that: The prediction score of the user's high preference and high satisfaction for the object obtained by performing a dot product operation based on the third user representation and the third object representation is specifically: The third user characterization With the third object characterization Perform inner product operation to obtain the predicted score of users with high satisfaction and high preference for the object , prediction score The calculation formula is as follows: 。 6. An online user preference high satisfaction recommendation device integrating commodity popularity analysis, characterized in that: The device comprises: A user and object information acquisition module, the acquisition module is used to acquire the total amount of interaction between the user and the object and the user's evaluation of the interaction with the object; A user preference acquisition module, the user preference acquisition module is used to transform the total amount of interaction behavior into object popularity through a function and input it into a recommendation model, obtain a high-preference first user representation and a high-preference first object representation, and obtain an object recommendation candidate set based on the first user representation and the first object representation; A satisfaction acquisition module, the satisfaction acquisition module is used to classify the objects interacted by each user into several levels according to the evaluation, input the objects with a level greater than a preset value into the recommendation model training, and obtain a second user representation with high satisfaction and a second object representation with high satisfaction; a recommendation result acquisition module, the recommendation result acquisition module being used to fuse the first user representation and the first object representation with the second user representation and the second object representation respectively, to obtain a third user representation and a third object representation with high preference and high satisfaction, to perform a dot product operation on the third user representation and the third object representation to obtain a prediction score of the user's high preference and high satisfaction for the object, and to obtain an object recommendation list for the user from the object recommendation candidate set based on the prediction score; The total amount of interaction between the user and the object is obtained by the number of times the user and the object interact. The expression of the interaction between the user and the object is: in, Represents the user, Represents an object, Indicates user With Object The interactive relationship, Indicates user With Object Interaction occurs, Indicates user With Object No interactive behavior occurred; The total amount of the interactive behavior is transformed into the object popularity through a function, and the function is specifically: in, Represents any object The popularity of represents the natural base, represents the temperature coefficient, Representing users and objects The total amount of interaction behavior; The first user representation and the first object representation are obtained by the following steps: Form an interaction graph of the interaction relationships between all users and objects , interaction diagram It is expressed as follows: ; in, Represents a collection of users, Represents a collection of objects; Set users and objects as nodes, and the neighbor node set of node h definition in, is any node in the interaction graph G, For Node A neighbor node around Represents a slave node To neighbor node An edge of Representation Node The set of neighbor nodes of The popularity of the object is added to the information transmission and aggregation of the recommendation model, and the activation function is applied to obtain the neighbor node set of the node Polymerization characterization : in, is the activation function, Representation Node The popularity of is a randomly initialized transformation matrix, the modality representing the object's data, Represents neighbor nodes In modal On the representation, Represents a set of neighbor nodes size; Obtain the ID information of the node, fuse the representation of the node with the ID information of the node and apply an activation function to obtain a fused vector representation after information fusion , fusion vector representation The expression is; in, Representation Node In modal Representation Fusion Node on The fusion vector representation obtained by the ID information is is a randomly initialized transformation matrix, Representation Node In modal On the representation, Representation Node ID information; The node Perform multi-stage training to obtain multi-stage representations after training; The node In modal The first-order characterization after the first-order connectivity on is: in, is a randomly initialized transformation matrix, Representation Node In modal First-order representation on ; node In modal on After stage connectivity The order representation is: in, Representation Node In modal on Level representation, Indicates that in modal on Randomly initialized transformation matrix of order; Representation Node In modal on After the stage connectivity, the ID information of the fused nodes is obtained Order fusion vector representation, Order fusion vector representation It is expressed as: The node Each mode of The data passed Stage training and fusion of each modality The representation of multimodal fusion is obtained for: in, Representation Node Multimodal fusion representation of Representation Node In modal on Level representation, a modal collection of data representing said object; Connect the representations obtained after training in each stage to get high preference users First user characterization and high preference objects First object representation .

7. A readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the online user preference high satisfaction recommendation method integrating commodity popularity analysis according to any one of claims 1 to 5 is implemented.

Images