Object recommendation method, device, electronic device and readable storage medium

Abstract

The present disclosure relates to the technical field of recommendation system algorithms, and provides an object recommendation method, device, electronic device, and readable storage medium. The method includes: performing multimodal vector representation on the feature data of each candidate object and the feature data of each interactive object to obtain a representation feature vector of each candidate object and a representation feature vector of each interactive object; performing nonlinear transformation on the representation feature vector of each candidate object and the representation feature vector of each interactive object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object and an enhanced feature vector of each interactive object; performing probability prediction on each candidate object based on the enhanced feature vector of each candidate object and the enhanced feature vector of each interactive object to obtain a predicted probability of each candidate object and select a target object to recommend to the user. This method solves the problem in the prior art of inaccurate recommendation results of the recommendation system based on identification features, and improves the accuracy of the recommendation model.

Description

Technical Field

[0001] The present disclosure relates to the technical field of recommendation system algorithms, and in particular to an object recommendation method, device, electronic device, and readable storage medium. Background Art

[0002] With the development of artificial intelligence technology, recommendation systems have been gradually applied to various platforms, including those used in information flow platforms, long and short video platforms, and e-commerce platforms. Recommendation systems are systems that recommend physical or non-physical items to users. Both physical and non-physical items are objects that the recommendation system needs to recommend. Based on user characteristics and object characteristics, the recommendation system selects target objects from hundreds of millions of objects and recommends them to users. The target objects are objects that the recommendation system predicts the user will be interested in. Current recommendation systems primarily learn to represent objects based on their identifier features. However, identifier features often vary between objects, and identifier features have poor transferability. Modeling identifier features is also insufficient to represent object characteristics, resulting in inaccurate predictions of user preferences. In recommendation systems, historical data often favors popular objects that have previously performed well. As the recommendation system continues to learn the characteristics of popular objects, it tends to continue recommending popular objects and ignore less popular objects, leading to a cold start problem. Summary of the Invention

[0003] In view of this, the embodiments of the present disclosure provide an object recommendation method, device, electronic device and readable storage medium to solve the problem in the prior art that recommendation results of the recommendation system are inaccurate due to recommendation based on identification features.

[0004] In a first aspect of an embodiment of the present disclosure, an object recommendation method is provided, comprising: obtaining a candidate object feature dataset, a user's interactive object feature dataset, and the user's feature data; performing multimodal vector representation on the feature data of each candidate object in the candidate object feature dataset through a pre-trained model to obtain a representation feature vector of each candidate object, and performing multimodal vector representation on the feature data of each interactive object in the interactive object feature dataset through a pre-trained model to obtain a representation feature vector of each interactive object; performing vector embedding on the user's feature data to obtain an initial feature vector of the user, and performing cross-processing on the initial feature vector of the user to obtain a feature vector of the user; performing nonlinear transformation on the representation feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object, and performing nonlinear transformation on the representation feature vector of each interactive object based on a recurrent neural network to obtain a temporal dependency relationship between the representation feature vectors of each interactive object to obtain an enhanced feature vector of each interactive object; performing probability prediction on each candidate object based on the user's feature vector, the enhanced feature vector of each candidate object, and the enhanced feature vector of each interactive object to obtain a predicted probability of each candidate object; and selecting a target object to recommend to the user based on the predicted probability of each candidate object.

[0005] According to a second aspect of the embodiments of the present disclosure, an object recommendation device is provided, comprising: an acquisition module for acquiring a candidate object feature data set, a user's interactive object feature data set, and user feature data; a multimodal characterization module for performing multimodal vector characterization on each candidate object feature data in the candidate object feature data set through a pre-trained model to obtain a characterization feature vector of each candidate object, and performing multimodal vector characterization on each interactive object feature data in the interactive object feature data set through a pre-trained model to obtain a characterization feature vector of each interactive object; a user characterization module for performing vector embedding on the user's feature data to obtain the user's initial feature vector, and performing cross-processing on the user's initial feature vector to obtain a to the user's feature vector; a vector enhancement module, which is used to perform a nonlinear transformation on the representation feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object, and to perform a nonlinear transformation on the representation feature vector of each interactive object based on a recurrent neural network to obtain the temporal dependency between the representation feature vectors of each interactive object to obtain an enhanced feature vector of each interactive object; a prediction module, which is used to perform a probability prediction on each candidate object based on the user's feature vector, the enhanced feature vectors of each candidate object and the enhanced feature vectors of each interactive object to obtain a predicted probability of each candidate object; a recommendation module, which is used to select a target object to recommend to the user based on the predicted probability of each candidate object.

[0006] According to a third aspect of an embodiment of the present disclosure, an electronic device is provided, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the computer program.

[0007] According to a fourth aspect of the embodiments of the present disclosure, a readable storage medium is provided, which stores a computer program. When the computer program is executed by a processor, the steps of the above method are implemented.

[0008] Compared with the prior art, the embodiments of the present disclosure have the following advantages: the present disclosure inputs the feature data of each candidate object in the candidate object feature dataset into the pre-trained model of the recommendation model for multimodal vector representation. The candidate object feature data includes the identification features and text features of the candidate object. Based on the pre-trained model, the candidate objects are multimodally represented to obtain high-quality candidate object representations, i.e., the representation feature vectors of each candidate object, which provide a basis for subsequent recommendation tasks. Furthermore, the feature data of each interactive object in the interactive object feature dataset is input into the pre-trained model to perform multimodal vector representation on the feature data of each interactive object in the interactive object feature dataset. The identification features and text features of the interactive object are used for multimodal modeling to obtain the representation feature vectors of each interactive object. The feature vectors contain rich information about the interactive object and can be used in subsequent recommendation tasks to make personalized recommendations to users. Vector embedding is performed on user feature data, converting discrete feature data into continuous, numerical vectors. This results in the user's initial feature vector, which helps computers identify the user's feature data. This initial feature vector is then fed into a deep neural network for transformation and combination, generating a richer and more meaningful feature representation. This, in turn, helps the recommendation system better understand user feature information, thereby providing more accurate personalized recommendations. The candidate object representation feature vectors are then fed into a recurrent neural network for nonlinear transformation, generating more complex, nonlinear feature relationships. This captures the intrinsic attributes and relationships of the candidate objects, resulting in more expressive and semantically informative enhanced feature vectors for the candidate objects, providing stronger support for subsequent recommendation decisions. The representation feature vectors of each interacting object are then fed into a recurrent neural network for nonlinear transformation, capturing the temporal dependencies between the representation feature vectors of each interacting object. This allows for a better understanding of the user's dynamic behavior and preferences, learning richer and more diverse feature representations, and generating enhanced feature vectors for each interacting object. This enhances the recommendation model's ability to predict user preferences for each candidate object, resulting in more accurate recommendations. The user's feature vector, the enhanced feature vectors of the candidate objects, and the enhanced feature vectors of each interactive object are concatenated and fed into a neural network for feature transformation. The input data undergoes a series of nonlinear transformations, gradually transforming from lower-level feature representations to higher-level feature representations. This approach can capture and abstract complex data patterns and relationships, perform probabilistic predictions on candidate objects, and output the predicted probability of the candidate object, which represents the user's preference or interest in the candidate object as predicted by the recommendation model. A probabilistic prediction is performed on each candidate object, resulting in a predicted probability for each candidate object. Based on the predicted probability of each candidate object, the candidate with the highest predicted probability is selected as the target object and recommended to the user.The present disclosure performs multimodal modeling on the feature data of candidate objects and the feature data of interactive objects through a pre-trained model to obtain an object representation including rich information, thereby solving the problem in the prior art of inaccurate recommendation results of the recommendation system caused by recommendations based on identification features, and improving the accuracy of the recommendation model of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] In order to more clearly illustrate the technical solutions in the embodiments of the present disclosure, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present disclosure. For ordinary technicians in this field, other drawings can be obtained based on these drawings without any creative work.

[0010] Figure 1 is a schematic diagram of an application scenario of an embodiment of the present disclosure;

[0011] Figure 2 is a flowchart of an object recommendation method provided by an embodiment of the present disclosure;

[0012] Figure 3 is a flowchart of another object recommendation method provided by an embodiment of the present disclosure;

[0013] Figure 4 is a flowchart of another object recommendation method provided by an embodiment of the present disclosure;

[0014] Figure 5 is a structural diagram of an object recommendation device provided by an embodiment of the present disclosure;

[0015] Figure 6 It is a structural diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION

[0016] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present disclosure. However, it will be apparent to those skilled in the art that the present disclosure may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present disclosure with unnecessary detail.

[0017] It should be noted that the user information (including but not limited to terminal device information, user personal information, etc.) and data (including but not limited to data used for display, data for analysis, etc.) involved in this disclosure are all information and data authorized by the user or fully authorized by all parties.

[0018] An object recommendation method and apparatus according to an embodiment of the present disclosure will be described in detail below with reference to the accompanying drawings.

[0019] Figure 1 FIG2 is a schematic diagram of an application scenario of an embodiment of the present disclosure. The application scenario may include terminal devices 1, 2, and 3, a server 4, and a network 5.

[0020] The terminal devices 1, 2 and 3 can be hardware or software. When the terminal devices 1, 2 and 3 are hardware, they can be various electronic devices with display screens and supporting communication with the server 4, including but not limited to smart phones, tablet computers, laptop computers and desktop computers; when the terminal devices 1, 2 and 3 are software, they can be installed in the above electronic devices. The terminal devices 1, 2 and 3 can be implemented as multiple software or software modules, or as a single software or software module, and the embodiments of the present disclosure are not limited to this. Furthermore, various applications can be installed on the terminal devices 1, 2 and 3, such as data processing applications, instant messaging tools, social platform software, search applications, shopping applications, etc.

[0021] Server 4 can be a server that provides various services, for example, a backend server that receives requests sent by terminal devices that establish communication connections with it. The backend server can receive and analyze the requests sent by the terminal devices, and generate processing results. Server 4 can be a single server, a server cluster consisting of multiple servers, or a cloud computing service center, all of which are not limited in the present embodiment.

[0022] It should be noted that the server 4 can be either hardware or software. When the server 4 is hardware, it can be various electronic devices that provide various services to the terminal devices 1, 2, and 3. When the server 4 is software, it can be multiple software programs or software modules that provide various services to the terminal devices 1, 2, and 3, or it can be a single software program or software module that provides various services to the terminal devices 1, 2, and 3. This is not limited in the present embodiment.

[0023] The network 5 can be a wired network connected by coaxial cable, twisted pair and optical fiber, or it can be a wireless network that can interconnect various communication devices without wiring, such as Bluetooth, Near Field Communication (NFC), infrared, etc. The embodiments of the present disclosure are not limited to this.

[0024] The user can establish a communication connection with the server 4 via the network 5 through the terminal devices 1, 2, and 3 to receive or send information. Specifically, the server 4 obtains a candidate object feature dataset, a user's interactive object feature dataset, and the user's feature data; performs multimodal vector representation on the feature data of each candidate object in the candidate object feature dataset using a pre-trained model to obtain a representation feature vector for each candidate object, and performs multimodal vector representation on the feature data of each interactive object in the interactive object feature dataset using a pre-trained model to obtain a representation feature vector for each interactive object; performs vector embedding on the user's feature data to obtain an initial feature vector for the user, and performs cross processing on the initial feature vector for the user to obtain a feature vector for the user; performs nonlinear transformation on the representation feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector for each candidate object; and performs nonlinear transformation on the representation feature vector of each interactive object based on a recurrent neural network to obtain a temporal dependency between the representation feature vectors of each interactive object to obtain an enhanced feature vector for each interactive object; performs probability prediction on each candidate object based on the user's feature vector, the enhanced feature vector of each candidate object, and the enhanced feature vector of each interactive object to obtain a predicted probability for each candidate object; and selects a target object for recommendation to the user based on the predicted probability of each candidate object.

[0025] It should be noted that the specific types, quantities and combinations of the terminal devices 1, 2 and 3, the server 4 and the network 5 can be adjusted according to the actual needs of the application scenario, and the embodiments of the present disclosure do not limit this.

[0026] Figure 2 It is a flowchart of an object recommendation method provided by an embodiment of the present disclosure. Figure 2 The object recommendation method can be Figure 1 The server executes. Figure 2 As shown, the object recommendation method includes:

[0027] Step 201: Acquire a candidate object feature dataset, a user's interactive object feature dataset, and user feature data.

[0028] In some embodiments, the object recommendation method of the present disclosure can be applied to recommendation systems of various platforms, including housing rental and purchase platforms, long video platforms, short video platforms, news information platforms, etc. The above-mentioned candidate object feature data set can be obtained from the platform database through a recall method. The objects in the platform database can reach hundreds of millions, and the recall method can obtain candidate object feature data of thousands of levels. The present disclosure does not limit the number of candidate objects. The candidate objects can be in the thousands or tens of thousands. The object recommendation method of the present disclosure can be applied to a housing rental and purchase platform. Specifically, the above-mentioned candidate objects and interactive objects can be houses. The candidate object feature data includes multiple identification features and multiple text features of the candidate object. The identification features of the candidate object can be an identifier that reflects the unique identity of the candidate object individual, which can be used to distinguish different objects or users and can effectively represent the uniqueness of the object or user. Specifically, when the object recommendation method provided by the present disclosure is applied to a housing rental platform, the identification features of the candidate object can be the unit ID of the house, the building ID of the house, the ID of the project to which the house belongs, and so on. The text features of a candidate object are semantic information used to describe and represent the content of the candidate object. For example, the text features of a candidate object can be the title text of a house on the platform, the dictionary information of the building to which the house belongs, the supporting facilities information of the house, the information of the neighborhood where the house is located, etc. In this case, the above-mentioned houses are all candidate objects. The interactive object feature data set of the above-mentioned user can be the feature data of multiple interactive objects that the user has clicked, collected, or browsed. The feature data of the interactive object includes multiple identification features and multiple text features of the interactive object. Specifically, the interactive object is a house. The text features of the interactive object can be the title text of the house on the platform, the dictionary information of the building to which the house belongs, the supporting facilities information of the house, the information of the neighborhood where the house is located, the scene of the user's interaction with the house, the type of user's interactive behavior, the time of the user's interaction with the house, etc. The identification features of the interactive object can be the unit ID of the house, the building ID of the house, the ID of the project to which the house belongs, etc. The feature data of the user in this disclosure can be the user's age, gender, identification features of the user, user activity, etc. For objects, by combining identification features with text features, the characteristics of the object can be more comprehensively represented, which can help the recommendation model better understand the information and attributes of the object and the characteristics of the user, make more accurate recommendations to the user, and improve the accuracy of the recommendation model.

[0029] In step 202, a multimodal vector representation is performed on each candidate object feature data in the candidate object feature data set through a pre-trained model to obtain a representation feature vector of each candidate object, and a multimodal vector representation is performed on each interactive object feature data in the interactive object feature data set through a pre-trained model to obtain a representation feature vector of each interactive object.

[0030] In some embodiments, the house rental platform can use a recommendation model to input the candidate object feature data set, the user's interactive object feature data set, and the user's feature data into the recommendation model. The recommendation model makes a probability prediction for each candidate object based on the above input data to obtain the predicted probability of each candidate object. The predicted probability of the candidate object can represent the user's preference for the candidate object predicted by the recommendation model.

[0031] In some embodiments, the recommendation model includes a pre-trained model. In the present disclosure, the multimodal vector representation can be for converting different feature data into a unified vector representation to facilitate subsequent deep learning processing. The candidate object feature data includes multiple text features and multiple identification features of the candidate object, and the interactive object feature data includes multiple text features and multiple identification features of the interactive object. The candidate object feature data is input into the trained pre-trained model to process the candidate object feature data. The processing process includes complex nonlinear transformation and feature extraction. After processing by the pre-trained model, the candidate object feature data is converted into a high-dimensional, vector representation, that is, the representation feature vector of the candidate object, which can better capture the intrinsic structure and semantic information of the input data, and obtain high-quality candidate object feature data representation. The feature data of each interactive object is input into the trained pre-trained model to process the feature data of each interactive object. The processing process includes complex nonlinear transformation and feature extraction. After processing by the pre-trained model, the feature data of each interactive object is converted into a high-dimensional, vector representation, that is, the representation feature vector of each interactive object, which can better capture the intrinsic structure and semantic information of each interactive object feature data, and obtain high-quality representation of each interactive object feature data. Through this multimodal vector representation process, we can obtain feature vectors representing each candidate object and each interactive object, which can be further used in other processing steps of the recommendation model, such as probability prediction. This multimodal vector representation method can effectively integrate different types of feature data for candidate objects and interactive objects, improving the accuracy and generalization capabilities of the recommendation system.

[0032] Step 203 , performing vector embedding on the user's feature data to obtain the user's initial feature vector, and performing cross processing on the user's initial feature vector to obtain the user's feature vector.

[0033] In some embodiments, a user's feature data includes multiple discrete features and multiple continuous features of the user. In this disclosure, the multiple discrete features of the user may include the user's gender, the user's identification feature, the user's city, etc. In this embodiment of the disclosure, the multiple discrete features of the user may include the user's activity level, the user's age, etc. By encoding each discrete feature of the user through hash coding, the user's discrete features can be uniformly represented as a binary string, i.e., the encoding result of each discrete feature of the user, thereby facilitating subsequent storage, calculation, and other steps. By feeding the encoding results of each user's discrete features into the first embedding layer for vector embedding, the encoding results are converted from hash-coded binary form to continuous vector representations. This helps capture semantic information and similarities between features, enabling the recommendation model to better understand and process the input feature data, resulting in each user's discrete feature vectors. Each user's continuous features are then bucketed, dividing the range of their continuous features into discrete buckets and mapping their values ​​to the corresponding buckets, resulting in bucketed results for each continuous feature. These bucketed results are then fed into the first embedding layer for vector embedding, resulting in each user's continuous feature vectors, converting the discretized features into vectors. The user's discrete feature vectors and continuous feature vectors are concatenated to obtain a higher-dimensional initial feature vector, integrating all information from the user's discrete and continuous features. The initial feature vectors are then cross-processed and fed into a deep neural network to learn complex patterns and relationships in the input data, resulting in the user's feature vector. This is then used for subsequent tasks such as probability prediction to achieve more precise object recommendations and improve the accuracy of the recommendation model.

[0034] In step 204, a nonlinear transformation is performed on the representation feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object, and a nonlinear transformation is performed on the representation feature vector of each interactive object based on a recurrent neural network to obtain a temporal dependency relationship between the representation feature vectors of each interactive object to obtain an enhanced feature vector of each interactive object.

[0035] In some embodiments, the recommendation model also includes a recurrent neural network. Specifically, the recurrent neural network can be a recurrent neural network (RNN), a long short-term memory network (LSTM), a gated recurrent unit (GRU), etc. When the above-mentioned recurrent neural network is an RNN, the RNN includes multiple memory units. The representation feature vector of the candidate object is input into the RNN. The memory unit of the RNN can be calculated based on the hidden state of the previous moment and the input at the current moment to obtain the hidden state at the current moment. Each element of the representation feature vector of the input candidate object is processed one by one to capture the temporal dependency. In each memory unit of the RNN, a nonlinear transformation is performed through an activation function so that the input feature vector can capture more complex patterns and relationships after being processed by the RNN, thereby obtaining an enhanced feature vector of the candidate object. The representation feature vectors of multiple candidate objects are input into the RNN for nonlinear transformation to obtain an enhanced feature vector of each candidate object. The representation feature vectors of each interactive object are input into the RNN. The RNN's memory unit can calculate the hidden state at the previous moment and the input at the current moment to obtain the hidden state at the current moment. The representation feature vectors of each interactive object are processed one by one to capture temporal dependencies. In each memory unit of the RNN, a nonlinear transformation is performed through the activation function so that the representation feature vectors of each interactive object can capture more complex patterns and relationships after being processed by the RNN, resulting in enhanced feature vectors for each interactive object. Nonlinear transformation of the representation feature vectors through recurrent neural networks can further improve the quality of the feature vectors and enhance the performance of the recommendation system.

[0036] Step 205 : Probability prediction is performed on each candidate object based on the user's feature vector, the enhanced feature vectors of each candidate object, and the enhanced feature vectors of each interactive object to obtain a predicted probability of each candidate object.

[0037] In some embodiments, a probability prediction is performed on each candidate object based on the user's feature vector, the enhanced feature vectors of each candidate object, and the enhanced feature vectors of each interactive object, to obtain a predicted probability of whether the user will click on each candidate object. The predicted probability of each candidate object can be a value between 0 and 1. The higher the predicted probability, the more likely the recommendation model predicts that the corresponding candidate object will be clicked by the user. The predicted probability of each candidate object is a prerequisite for subsequent ranked recommendation. Selecting multiple candidate objects with higher predicted probabilities for recommendation to the user improves the accuracy of the recommendation model and enhances the user experience on this platform.

[0038] Step 206 : Based on the predicted probability of each candidate object, select a target object to recommend to the user.

[0039] In some embodiments, the predicted probability of each candidate object reflects the user's preference for each candidate object predicted by the recommendation model. A preset threshold can be set to compare the predicted probability of each candidate object with the preset threshold. When the predicted probability of the candidate object is greater than or equal to the preset threshold, the corresponding candidate object is determined as the target object recommended to the user. The number of target objects is not limited in this disclosure and can be any reasonable number. The above-mentioned preset threshold is set according to the test results in the specific embodiment and is not limited in this disclosure. By calculating the predicted probability of each candidate object and sorting and selecting based on the above-mentioned predicted probability, the user can be provided with target objects that are highly matched with his interests and needs, thereby improving the recommendation effect of the recommendation model and solving the problem of inaccurate recommendation results of the recommendation system based on identification features in the prior art.

[0040] In some embodiments, after selecting a target object to recommend to a user, the method further includes updating the parameters of the recommendation model based on user feedback. User feedback may include information about the user's interest in the recommended target object and their interaction (e.g., clicks, purchases, comments, etc.). This feedback information can be used as new data input into the recommendation model to update the user's feature vector, thereby improving the accuracy and personalization of subsequent recommendations.

[0041] Based on the method provided in the aforementioned embodiment, the present disclosure inputs the feature data of each candidate object in the candidate object feature dataset into a pre-trained model of the recommendation model for multimodal vector representation. The candidate object feature data includes the identification features and text features of the candidate object. Based on the pre-trained model, the candidate objects are multimodally represented to obtain high-quality candidate object representations, namely, the representation feature vectors of each candidate object, which provide a basis for subsequent recommendation tasks. Furthermore, the feature data of each interactive object in the interactive object feature dataset is input into the pre-trained model to perform multimodal vector representation on the feature data of each interactive object in the interactive object feature dataset. Multimodal modeling is performed using the identification features and text features of the interactive object to obtain the representation feature vector of each interactive object. This feature vector contains rich information about the interactive object and can be used in subsequent recommendation tasks to make personalized recommendations to users. Vector embedding is performed on user feature data, converting discrete feature data into continuous, numerical vectors. This results in the user's initial feature vector, which helps computers identify the user's feature data. This initial feature vector is then fed into a deep neural network for transformation and combination, generating a richer and more meaningful feature representation. This, in turn, helps the recommendation system better understand user feature information, thereby providing more accurate personalized recommendations. The candidate object representation feature vectors are then fed into a recurrent neural network for nonlinear transformation, generating more complex, nonlinear feature relationships. This captures the intrinsic attributes and relationships of the candidate objects, resulting in more expressive and semantically informative enhanced feature vectors for the candidate objects, providing stronger support for subsequent recommendation decisions. The representation feature vectors of each interacting object are then fed into a recurrent neural network for nonlinear transformation, capturing the temporal dependencies between the representation feature vectors of each interacting object. This allows for a better understanding of the user's dynamic behavior and preferences, learning richer and more diverse feature representations, and generating enhanced feature vectors for each interacting object. This enhances the recommendation model's ability to predict user preferences for each candidate object, resulting in more accurate recommendations. The user's feature vector, the enhanced feature vectors of the candidate objects, and the enhanced feature vectors of each interactive object are concatenated and fed into a neural network for feature transformation. The input data undergoes a series of nonlinear transformations, gradually transforming from lower-level feature representations to higher-level feature representations. This approach can capture and abstract complex data patterns and relationships, perform probabilistic predictions on candidate objects, and output the predicted probability of the candidate object, which represents the user's preference or interest in the candidate object as predicted by the recommendation model. A probabilistic prediction is performed on each candidate object, resulting in a predicted probability for each candidate object. Based on the predicted probability of each candidate object, the candidate with the highest predicted probability is selected as the target object and recommended to the user.The present disclosure performs multimodal modeling on the feature data of candidate objects and the feature data of interactive objects through a pre-trained model to obtain an object representation including rich information, thereby solving the problem in the prior art of inaccurate recommendation results of the recommendation system caused by recommendations based on identification features, and improving the accuracy of the recommendation model of the recommendation system.

[0042] In some embodiments, the candidate object feature data includes multiple text features of the candidate object and multiple identification features of the candidate object; a multimodal vector representation is performed on each candidate object feature data in the candidate object feature data set through a pre-trained model to obtain a representation feature vector of each candidate object, including: inputting each text feature of each candidate object into a text encoder for encoding processing to obtain each text encoding result of each candidate object, and inputting each text encoding result of each candidate object into a whitening layer for fixed embedding to obtain a target feature vector of each text of each candidate object; vector embedding is performed on multiple identification features of each candidate object to obtain an initial feature vector of each identification feature of each candidate object; and the initial feature vector of each identification feature of each candidate object is input into a first multi-layer perceptron for feature transformation to obtain a feature vector of each identification feature of each candidate object; the feature vector of each identification feature of each candidate object is input into a whitening layer for fixed embedding to obtain a target feature vector of each identification feature of each candidate object; based on the target feature vector of each text of each candidate object and the target feature vector of each identification feature of each candidate object, the representation feature vector of each candidate object is determined.

[0043] In some embodiments, the identification features of a candidate object may be identifiers that reflect the unique identity of the candidate object, and can be used to distinguish different objects. Specifically, when the object recommendation method provided by the present disclosure is applied to a housing rental platform, the identification features of the candidate object may be the unit ID of the house, the building ID of the house, the ID of the project to which the house belongs, etc. The text features of the candidate object are semantic information used to describe and represent the content of the candidate object. For example, the text features of the candidate object may be the title text of the house on the platform, the dictionary information of the real estate project to which the house belongs, the supporting facilities information of the house, the information of the residential community in which the house is located, etc.

[0044] In some embodiments, the pre-training model includes a text encoder, a whitening layer, a second ebbedding layer, and a first multi-layer perceptron. The text encoder can be MacBert. The text features of each candidate object are input into the MacBert model for encoding processing to capture the semantic information in the text features of each candidate object, and the text features are converted into corresponding semantic vectors to obtain the text encoding results of each candidate object. The text encoding results of each candidate object are input into the whitening layer of the pre-training model, and the text encoding results of each candidate object are fixedly embedded. The text encoding results of each candidate object are mapped to a fixed embedding space so that the distribution of the text encoding results of each candidate object is more uniform. The feature vectors can be further adjusted and standardized to obtain the target feature vectors of each text of each candidate object. The target feature vectors of each text of each candidate object contain the semantic information of the text features and the standardized distribution features. The multiple identification features of each candidate object are input into the second ebbedding layer for vector embedding, and the multiple discrete identification features of each candidate object are converted into continuous vector representations to obtain the initial feature vectors of each identification feature of each candidate object. The initial feature vectors of each identification feature of each candidate object are input into the first multi-layer perceptron for feature transformation, further extracting and enhancing the representation capability of the features, and obtaining richer and more accurate feature vectors of each identification feature of each candidate object. The feature vectors of each identification feature of each candidate object are input into the whitening layer for fixed embedding, which can further adjust and standardize the feature vectors of each identification feature of each candidate object, and obtain more accurate and standardized target feature vectors of each identification feature of each candidate object. Based on the target feature vectors of each text of each candidate object and the target feature vectors of each identification feature of each candidate object, the representation feature vector of each candidate object is obtained. The representation feature vector of each candidate object integrates the semantic information of the text and identification features, providing a rich and accurate data foundation for subsequent recommendation algorithms, which can improve the accuracy and personalization of the recommendation system.

[0045] In some embodiments, the feature data of the interactive object includes multiple text features of the interactive object and multiple identification features of the interactive object; a multimodal vector representation is performed on each interactive object feature data in the interactive object feature data set through a pre-trained model to obtain a representation feature vector of each interactive object, including: inputting each text feature of each interactive object into a text encoder for encoding processing to obtain each text encoding result of each interactive object, and inputting each text encoding result of each interactive object into a whitening layer for fixed embedding to obtain a target feature vector of each text of each interactive object; vector embedding is performed on multiple identification features of each interactive object to obtain an initial feature vector of each identification feature of each interactive object; and the initial feature vector of each identification feature of each interactive object is input into a first multi-layer perceptron for feature transformation to obtain a feature vector of each identification feature of each interactive object; the feature vector of each identification feature of each interactive object is input into a whitening layer for fixed embedding to obtain a target feature vector of each identification feature of each interactive object; based on the target feature vector of each text of each interactive object and the target feature vector of each identification feature of each interactive object, the representation feature vector of each interactive object is determined.

[0046] In some embodiments, based on the target feature vectors of each text of each candidate object and the target feature vectors of each identification feature of each candidate object, the representation feature vector of each candidate object is determined, including: splicing the target feature vectors of each text of each candidate object and the target feature vectors of each identification feature of each candidate object to obtain the initial fused feature vector of each candidate object; inputting the initial fused feature vector of each candidate object into the second multi-layer perceptron for feature transformation to obtain the fused feature vector of each candidate object, and normalizing the fused feature vector of each candidate object to obtain the representation feature vector of each candidate object.

[0047] In some embodiments, the target feature vectors of each text of each candidate object and the target feature vectors of each identification feature of each candidate object can be spliced ​​and merged in order or by dimension to form a feature vector of higher dimension, i.e., the initial fused feature vector of each candidate object. The initial fused feature vector of each candidate object integrates multiple attributes of the candidate object. The second multi-layer perceptron can be a deep learning model that can perform complex nonlinear transformations on the input data. Inputting the initial fused feature vector of each candidate object into the second multi-layer perceptron for feature transformation can further extract and enhance the representation capability of the features, thereby obtaining a richer and more accurate fused feature vector of each candidate object. The fused feature vector of each candidate object is normalized to adjust the scale of the fused feature vector of each candidate object so that it is comparable. Through normalization, the scale differences between different features can be eliminated to obtain the representation feature vector of each candidate object. The representation feature vector of each candidate object can fully reflect the attributes of the candidate object, providing strong support for the accuracy and personalization of the recommendation system, improving the accuracy of the recommendation model, and obtaining more accurate recommendation results.

[0048] The first multilayer perceptron is primarily used to process and transform the initial feature vectors of the identification features. By receiving these initial feature vectors and performing complex nonlinear transformations, it further extracts and enhances the representational power of the identification features. The second multilayer perceptron is primarily used to process and transform higher-level initial fused feature vectors. By receiving these initial fused feature vectors and performing feature transformations, it further enhances the representational power of the features. Through the collaborative operation of the first and second multilayer perceptrons, various feature representations of candidate objects can be effectively extracted and enhanced, providing a foundation for subsequent computational processing.

[0049] In some embodiments, based on the target feature vectors of each text of each interactive object and the target feature vectors of each identification feature of each interactive object, the representation feature vector of each interactive object is determined, including: splicing the target feature vectors of each text of each interactive object and the target feature vectors of each identification feature of each interactive object to obtain the initial fused feature vector of each interactive object; inputting the initial fused feature vector of each interactive object into the second multi-layer perceptron for feature transformation to obtain the fused feature vector of each interactive object, and normalizing the fused feature vector of each interactive object to obtain the representation feature vector of each interactive object.

[0050] refer to Figure 3The pre-training model 300 may include a text encoder 301, a second embedding layer 302, a first multi-layer perceptron 303, a whitening layer 304, a first splicing processing module 305, a second multi-layer perceptron 306, and a normalization processing module 307. The text features of each candidate object are input into the text encoder 301 for encoding processing to obtain the text encoding results of each candidate object, and the text encoding results of each candidate object are input into the whitening layer 304 for fixed embedding to obtain the target feature vector of each text of each candidate object. The multiple identification features of each candidate object are input into the second embedding layer 302 for vector embedding to obtain the initial feature vector of each identification feature of each candidate object, and the initial feature vector of each identification feature of each candidate object is input into the first multi-layer perceptron 303 for feature transformation to obtain the feature vector of each identification feature of each candidate object. The feature vector of each identification feature of each candidate object is input into the whitening layer 304 for fixed embedding to obtain the target feature vector of each identification feature of each candidate object. The target feature vectors of each text of each candidate object and the target feature vectors of each identification feature of each candidate object are input into the first splicing processing module 305 for splicing to obtain the initial fused feature vector of each candidate object. The initial fused feature vector of each candidate object is input into the second multi-layer perceptron 306 for feature transformation to obtain the fused feature vector of each candidate object, and the fused feature vector of each candidate object is input into the normalization processing module 307 for normalization to obtain the representation feature vector of each candidate object. By performing multimodal vector representation on the feature data of each candidate object through the pre-training model 300, the representation feature vector of each candidate object can be obtained. The representation feature vector of each candidate object can fully reflect the attributes of the candidate object, provide strong support for the accuracy and personalization of the recommendation system, improve the accuracy of the recommendation model, and obtain more accurate recommendation results.

[0051] In some embodiments, a nonlinear transformation is performed on the characterization feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object, including: inputting the characterization feature vector of each candidate object into the recurrent neural network, calculating the current hidden layer state based on the characterization feature vector of each candidate object and the state of the previous hidden layer, and inputting the current hidden layer state into the activation function layer of the recurrent neural network to obtain an enhanced feature vector of each candidate object.

[0052] In some embodiments, in a recurrent neural network, the state of the current hidden layer is calculated based on the characterization feature vector of the current input and the state of the previous hidden layer. Based on the characterization feature vector of the candidate object currently input and the state of the previous hidden layer, the state of the current hidden layer is calculated through specific weights and activation functions, historical information is captured and context-related feature representations are generated, and the current hidden layer state is input into an activation function layer, such as a ReLU (Rectified Linear Unit) function or a sigmoid function, to learn more complex feature relationships and obtain enhanced feature vectors of the candidate objects. Through the conversion and enhancement of the recurrent neural network, more contextual information and temporal dependencies are captured to obtain enhanced feature vectors of the candidate objects. Inputting the characterization feature vectors of each candidate object into the recurrent neural network to obtain enhanced feature vectors of each candidate object can further enhance the richness and context relevance of the feature representation, and provide more accurate data for subsequent probability prediction tasks.

[0053] In some embodiments, a nonlinear transformation is performed on the representation feature vectors of each interactive object based on a recurrent neural network to obtain the temporal dependency between the representation feature vectors of each interactive object and obtain an enhanced feature vector of each interactive object, including: inputting the representation feature vectors of each interactive object into the recurrent neural network based on the user's interaction order, calculating the current hidden layer state based on the representation feature vectors of each interactive object and the state of the previous hidden layer, and inputting the current hidden layer state into the activation function layer of the recurrent neural network to obtain an enhanced feature vector of each interactive object.

[0054] In some embodiments, based on the user's interaction order, the characterization feature vectors of each interactive object are input into a recurrent neural network, so that the computer recognizes the historical interaction information between the user and each interactive object, and based on the characterization feature vectors of each interactive object currently input and the state of the previous hidden layer, the state of the current hidden layer is calculated by a specific weight and activation function, so that the recurrent neural network can use the historical interaction information to generate a richer and more context-related feature representation for each interactive object, and input the current hidden layer state into an activation function layer, such as a ReLU function or a sigmoid function, to learn a more complex feature relationship and obtain an enhanced feature vector of the interactive object. Through the conversion and enhancement of the recurrent neural network, more contextual information and temporal dependencies are captured to obtain an enhanced feature vector of the interactive object. The characterization feature vectors of each interactive object are input into the recurrent neural network to obtain the enhanced feature vectors of each interactive object, which can further enhance the richness and context relevance of the feature representation of the interactive object and provide more accurate data for subsequent probability prediction tasks.

[0055] In some embodiments, probability prediction is performed on each candidate object based on the user's feature vector, the enhanced feature vectors of each candidate object, and the enhanced feature vectors of each interactive object to obtain the predicted probability of each candidate object, including: splicing the user's feature vector, the enhanced feature vectors of each candidate object, and the enhanced feature vectors of each interactive object to obtain each fused feature vector; inputting each fused feature vector into a third multi-layer perceptron for feature transformation, so that probability prediction processing is performed on each candidate object to obtain the predicted probability of each candidate object.

[0056] In some embodiments, the user's feature vector is spliced ​​with the enhanced feature vectors of each candidate object and the enhanced feature vectors of each interactive object, and feature information from different sources is integrated to obtain each fused feature vector, and each fused feature vector is input into a third multi-layer perceptron for feature transformation. Each fused feature vector is processed and transformed at a deeper level to enhance the complexity and discrimination of the feature representation. The probability of each candidate object being clicked by the user or other interactive behaviors is predicted based on the results of each feature transformation. The predicted probability of each candidate object can be used for subsequent recommendation decisions, such as sorting the candidate objects according to the size of the probability value, thereby providing personalized recommendations to users.

[0057] refer to Figure 4The recommendation model may include a pre-trained model 300, a first embedding layer 401, a deep neural network 402, a recurrent neural network 403, a second splicing processing module 404, and a third multi-layer perceptron 405. The pre-trained model 300 performs multimodal vector representation on the feature data of each candidate object in the candidate object feature data set to obtain a representation feature vector of each candidate object, and the pre-trained model 300 performs multimodal vector representation on the feature data of each interactive object in the interactive object feature data set to obtain a representation feature vector of each interactive object. The user's feature data is input into the first embedding layer 401 for vector embedding to obtain the user's initial feature vector, and the user's initial feature vector is input into the deep neural network 402 for cross processing to obtain the user's feature vector. Based on the recurrent neural network 403, a nonlinear transformation is performed on the representation feature vector of each candidate object to obtain an enhanced feature vector of each candidate object, and based on the recurrent neural network 403, a nonlinear transformation is performed on the representation feature vector of each interactive object to obtain the temporal dependency between the representation feature vectors of each interactive object to obtain an enhanced feature vector of each interactive object. The user's feature vector, the enhanced feature vectors of each candidate object, and the enhanced feature vectors of each interactive object are input into the second splicing processing module 404 for splicing processing to obtain fused feature vectors. Each fused feature vector is input into the third multi-layer perceptron 405 for feature transformation, so that probabilistic prediction processing is performed on each candidate object to obtain a predicted probability for each candidate object. Based on the predicted probabilities of each candidate object, a target object is selected for recommendation to the user. This provides the user with target objects that are highly compatible with their interests and needs, thereby improving the recommendation model's effectiveness and resolving the issue of inaccurate recommendation results in recommendation systems based on identification features in the prior art.

[0058] In some embodiments, before performing multimodal vector representation on each candidate object feature data in the candidate object feature data set through a pre-training model, the method further includes: obtaining a training data set of the pre-training model, the training data set containing multiple training object feature data, and each training object feature data is arranged based on the order of user interaction; selecting any one of the training object feature data in the training data set as the first object feature data, and the training object feature data after the first object feature data as the second object feature data, and all the training object feature data except the first object feature data and the second object feature data in the training data set as the third object feature data; selecting the first object feature data and the second object feature data to construct a positive sample pair, and selecting the first object feature data and the third object feature data to construct a multi-object feature pair. negative sample pairs; input the positive sample pairs and each negative sample pair into the pre-training model, perform multimodal vector representation on the feature data of the first object through the pre-training model to obtain the representation feature vector of the first object, perform multimodal vector representation on the feature data of the second object through the pre-training model to obtain the representation feature vector of the second object, perform vector representation on the feature data of each third object through the pre-training model to obtain the representation feature vector of each third object; based on the representation feature vector of the first object, the representation feature vector of the second object and the representation feature vector of each third object, obtain the loss value corresponding to the information-noise contrast loss function; and update the parameters of the pre-training model based on the loss value, and obtain a pre-training model that has been trained when the loss value is less than or equal to the preset value.

[0059] In some embodiments, the training process of the recommendation model includes two stages: pre-training and fine-tuning. Pre-training is the process of training a pre-trained model using an unsupervised learning training method. The above is the training process of the pre-trained model. A training data set for the pre-trained model is obtained. The training data set contains multiple training object feature data, and each training object feature data is arranged based on the order of user interaction. Any training object feature data is selected from the training data set as the first object feature data, and then the training object feature data in the order after the first object feature data is selected as the second object feature data. The remaining training object feature data is used as the third object feature data. Based on the first object feature data and the second object feature data, a positive sample pair is constructed. Based on the first object feature data and the multiple third object feature data, multiple negative sample pairs can be constructed, and each negative sample pair contains a first object feature data and a third object feature data.

[0060] The positive sample pairs and each negative sample pair are input into a pre-training model, and the first object feature data includes multiple text features of the first object and multiple identification features of the first object.

[0061] The multiple identification features of the first object may include the unit ID of a house, the building ID of a house, the ID of the project to which the house belongs, etc. The multiple text features of the first object may include the title text of the house on the platform, dictionary information of the real estate project to which the house belongs, information about the house's supporting facilities, information about the residential complex where the house is located, scenarios in which users interact with the house, types of user interaction behaviors, duration of user interaction with the house, etc. Random masking is performed on the text features of the first object to obtain masked text features of the first object. A random masking mechanism may be used to perform random masking on the text features. Perform text encoding processing on the occluded text features of the first object to obtain the occluded text encoding result of the first object, and input the occluded text encoding result of the first object into the whitening layer for fixed embedding to obtain the target feature vector of the occluded text of the first object; perform vector embedding on the identification feature of the first object to obtain the initial feature vector of the identification feature of the first object, and input the initial feature vector of the identification feature of the first object into the first multi-layer perceptron for feature transformation to obtain the feature vector of the identification feature of the first object; input the feature vector of the identification feature of the first object into the whitening layer for fixed embedding to obtain the target feature vector of the identification feature of the first object; and obtain the representation feature vector of the first object based on the target feature vector of the occluded text of the first object and the target feature vector of the identification feature of the first object.

[0062] The multiple identification features of the second object can include the unit ID of the house, the building ID of the house, the ID of the project to which the house belongs, etc. The multiple text features of the second object can include the title text of the house on the platform, the dictionary information of the real estate project to which the house belongs, the supporting facilities information of the house, the information of the residential complex where the house is located, the scene of user interaction with the house, the type of user interaction behavior, the duration of user interaction with the house, etc. The text features of the second object are randomly masked to obtain the masked text features of the second object. The random mask mechanism can be used to perform random masking on the text features. Perform text encoding processing on the occluded text features of the second object to obtain the occluded text encoding result of the second object, and input the occluded text encoding result of the second object into the whitening layer for fixed embedding to obtain the target feature vector of the occluded text of the second object; perform vector embedding on the identification features of the second object to obtain the initial feature vector of the identification features of the second object, and input the initial feature vector of the identification features of the second object into the first multi-layer perceptron for feature transformation to obtain the feature vector of the identification features of the second object; input the feature vector of the identification features of the second object into the whitening layer for fixed embedding to obtain the target feature vector of the identification features of the second object; and obtain the representation feature vector of the second object based on the target feature vector of the occluded text of the second object and the target feature vector of the identification features of the second object.

[0063] The multiple identification features of the third object can include the unit ID of the house, the building ID of the house, the ID of the project to which the house belongs, etc. The multiple text features of the third object can include the title text of the house on the platform, the dictionary information of the real estate project to which the house belongs, the supporting facilities information of the house, the information of the residential community where the house is located, the scene of user interaction with the house, the type of user interaction behavior, the duration of user interaction with the house, etc. The text features of each third object are randomly masked to obtain the masked text features of each third object. The random mask mechanism can be used to perform random masking on the text features. Perform text encoding processing on the occluded text features of each third object to obtain the occluded text encoding results of each third object, and input the occluded text encoding results of each third object into the whitening layer for fixed embedding to obtain the target feature vector of the occluded text of each third object; perform vector embedding on the identification features of each third object to obtain the initial feature vector of the identification features of each third object, and input the initial feature vector of the identification features of each third object into the first multi-layer perceptron for feature transformation to obtain the feature vector of the identification features of each third object; input the feature vector of the identification features of each third object into the whitening layer for fixed embedding to obtain the target feature vector of the identification features of each third object; and obtain the representation feature vector of each third object based on the target feature vector of the occluded text of each third object and the target feature vector of the identification features of each third object.

[0064] In some embodiments, based on the characterization feature vector of the first object, the characterization feature vector of the second object, and the characterization feature vectors of each third object, a loss value corresponding to the information noise contrast loss (Info Noise Contrastive Estimation loss, InfoNCE) function can be obtained, and based on the calculated loss value, the parameters of the pre-trained model are updated through the back propagation algorithm until the loss value is less than or equal to the preset value, and the training is stopped to obtain a pre-trained model. At this point, the pre-trained model has learned how to extract characterization features from multimodal data to facilitate prediction and recommendation based on these features.

[0065] In some embodiments, based on the characterization feature vector of the first object, the characterization feature vector of the second object, and the characterization feature vectors of each third object, a loss value corresponding to the information-noise contrast loss function is obtained, including: multiplying the characterization feature vector of the first object and the characterization feature vector of the second object to obtain the product of the positive samples; multiplying the characterization feature vector of the first object with the characterization feature vectors of each third object respectively to obtain the product of each negative sample; adding the products of each negative sample to obtain the sum of the negative sample products; and obtaining the loss value corresponding to the information-noise contrast loss function based on the product of the positive samples and the sum of the negative sample products.

[0066] In some embodiments, the InfoNCE loss function needs to meet the following requirements: that is, when the first object feature data and the second object feature data in the positive sample pair are similar, and the first object feature data and all third object feature data in each negative sample pair are dissimilar, the loss value corresponding to the information-noise contrast loss function should be relatively low. Conversely, if the first object feature data and the second object feature data are dissimilar, or the first object feature data and other third object feature data are similar, the loss value corresponding to the information-noise contrast loss function should be large, thereby penalizing the model and prompting the pre-trained model to update its parameters. By multiplying the characterization feature vector of the first object with the characterization feature vector of the second object to obtain the product of the positive sample, the similarity or correlation between the positive sample pairs can be measured. The characterization feature vector of the first object is multiplied by the characterization feature vector of each third object respectively to obtain the product of each negative sample. The products of each negative sample are added together to obtain the sum of the negative sample products, which can measure the similarity or correlation between the negative sample pairs. Based on the product of positive samples and the sum of the products of negative samples, the loss value corresponding to the information-noise contrast loss function can be obtained. This loss value can be used to guide the training process of the pre-training model, optimize the parameters of the pre-training model, and improve the prediction accuracy and recommendation effect of the recommendation model.

[0067] All the above optional technical solutions can be arbitrarily combined to form optional embodiments of the present disclosure, and will not be described in detail here.

[0068] The following are embodiments of the apparatus disclosed herein, which can be used to implement the method embodiments disclosed herein. For details not disclosed in the apparatus embodiments disclosed herein, please refer to the method embodiments disclosed herein.

[0069] Figure 5 Schematic diagram of an object recommendation device provided by an embodiment of the present disclosure. Figure 5 As shown, the object recommendation device includes:

[0070] An acquisition module 501 is configured to acquire a candidate object feature dataset, a user's interactive object feature dataset, and user feature data;

[0071] A multimodal characterization module 502 is configured to perform multimodal vector characterization on each candidate object feature data in the candidate object feature dataset using a pre-trained model to obtain a characterization feature vector for each candidate object, and to perform multimodal vector characterization on each interactive object feature data in the interactive object feature dataset using a pre-trained model to obtain a characterization feature vector for each interactive object;

[0072] The user characterization module 503 is used to perform vector embedding on the user's feature data to obtain the user's initial feature vector, and perform cross processing on the user's initial feature vector to obtain the user's feature vector;

[0073] A vector enhancement module 504 is configured to perform a nonlinear transformation on the representation feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object, and to perform a nonlinear transformation on the representation feature vector of each interactive object based on a recurrent neural network to obtain a temporal dependency relationship between the representation feature vectors of each interactive object to obtain an enhanced feature vector of each interactive object;

[0074] A prediction module 505 is configured to perform probability prediction on each candidate object based on the user's feature vector, the enhanced feature vectors of each candidate object, and the enhanced feature vectors of each interactive object, to obtain a predicted probability of each candidate object;

[0075] The recommendation module 506 is configured to select a target object to recommend to the user based on the predicted probability of each candidate object.

[0076] According to the technical solution provided by the embodiments of the present disclosure, the multimodal characterization module 502 inputs the feature data of each candidate object in the candidate object feature dataset into the pre-trained model of the recommendation model for multimodal vector characterization. The candidate object feature data includes the identification features and text features of the candidate object. Based on the pre-trained model, the candidate objects are multimodally characterized to obtain high-quality candidate object representations, namely, representation feature vectors of each candidate object, which provide a basis for subsequent recommendation tasks. Furthermore, the feature data of each interactive object in the interactive object feature dataset is input into the pre-trained model for multimodal vector characterization. Multimodal modeling is performed using the identification features and text features of the interactive objects to obtain representation feature vectors of each interactive object. This feature vector contains rich information about the interactive object and can be used in subsequent recommendation tasks to make personalized recommendations to users. The user characterization module 503 embeds the user's feature data into vectors, converting the user's discrete feature data into continuous, numerical vectors, i.e., obtaining the user's initial feature vector, which helps the computer identify the user's feature data. The user's initial feature vector is then input into a deep neural network for transformation and combination to obtain the user's feature vector, generating a richer and more meaningful feature representation, i.e., the user's feature vector, which can help the recommendation system better understand the user's feature information, thereby subsequently providing more accurate personalized recommendation results. The vector enhancement module 504 inputs the candidate object's representation feature vector into a recurrent neural network for nonlinear transformation, learning more complex, nonlinear feature relationships, capturing the intrinsic attributes and relationships of the candidate object, and obtaining a more expressive and semantically informative enhanced feature vector for the candidate object, providing stronger support for subsequent recommendation decisions. The representation feature vectors of each interactive object are input into a recurrent neural network for nonlinear transformation to obtain the temporal dependency between the representation feature vectors of each interactive object, better understanding the user's dynamic behavior and preferences, learning richer and more diverse feature representations, and obtaining an enhanced feature vector for each interactive object, which can better support the subsequent recommendation model in predicting the user's preferences for each candidate object, thereby obtaining more accurate recommendation results. The prediction module 505 concatenates the user's feature vector, the enhanced feature vectors of the candidate objects, and the enhanced feature vectors of each interactive object. The concatenation result is then fed into a neural network for feature transformation. The input data undergoes a series of nonlinear transformations, gradually transforming from lower-level feature representations to higher-level feature representations. This allows for the capture and abstraction of complex data patterns and relationships, and a probabilistic prediction is made for the candidate objects. The output is the predicted probability of the candidate object, which represents the user's preference or interest in the candidate object as predicted by the recommendation model. A probabilistic prediction is performed for each candidate object, resulting in a predicted probability for each candidate object. Based on the predicted probabilities of each candidate object, the recommendation module 506 selects the candidate object with the highest predicted probability as the target object and recommends it to the user.The present disclosure performs multimodal modeling on the feature data of candidate objects and the feature data of interactive objects through a pre-trained model to obtain an object representation including rich information, thereby solving the problem in the prior art of inaccurate recommendation results of the recommendation system caused by recommendations based on identification features, and improving the accuracy of the recommendation model of the recommendation system.

[0077] In some embodiments, the candidate object feature data includes multiple text features of the candidate object and multiple identification features of the candidate object; the multimodal representation module 502 is configured to perform multimodal vector representation on each candidate object feature data in the candidate object feature data set through a pre-trained model to obtain a representation feature vector of each candidate object, including inputting each text feature of each candidate object into a text encoder for encoding processing to obtain each text encoding result of each candidate object, and inputting each text encoding result of each candidate object into a whitening layer for fixed embedding to obtain a target feature vector of each text of each candidate object; performing vector embedding on multiple identification features of each candidate object to obtain an initial feature vector of each identification feature of each candidate object; and inputting the initial feature vector of each identification feature of each candidate object into a first multi-layer perceptron for feature transformation to obtain a feature vector of each identification feature of each candidate object; inputting the feature vector of each identification feature of each candidate object into a whitening layer for fixed embedding to obtain a target feature vector of each identification feature of each candidate object; based on the target feature vector of each text of each candidate object and the target feature vector of each identification feature of each candidate object, determine the representation feature vector of each candidate object.

[0078] In some embodiments, the multimodal representation module 502 is configured to determine the representation feature vector of each candidate object based on the target feature vector of each text of each candidate object and the target feature vector of each identification feature of each candidate object, including splicing the target feature vector of each text of each candidate object and the target feature vector of each identification feature of each candidate object to obtain the initial fused feature vector of each candidate object; inputting the initial fused feature vector of each candidate object into the second multi-layer perceptron for feature transformation to obtain the fused feature vector of each candidate object, and normalizing the fused feature vector of each candidate object to obtain the representation feature vector of each candidate object.

[0079] In some embodiments, the vector enhancement module 504 is configured to perform a nonlinear transformation on the characterization feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object, including inputting the characterization feature vector of each candidate object into the recurrent neural network, calculating the current hidden layer state based on the characterization feature vector of each candidate object and the state of the previous hidden layer, and inputting the current hidden layer state into the activation function layer of the recurrent neural network to obtain an enhanced feature vector of each candidate object.

[0080] In some embodiments, the vector enhancement module 504 is configured to perform a nonlinear transformation on the representation feature vectors of each interactive object based on a recurrent neural network, obtain the temporal dependency between the representation feature vectors of each interactive object, and obtain the enhanced feature vectors of each interactive object, including: inputting the representation feature vectors of each interactive object into the recurrent neural network based on the user's interaction order, calculating the current hidden layer state based on the representation feature vectors of each interactive object and the state of the previous hidden layer, and inputting the current hidden layer state into the activation function layer of the recurrent neural network to obtain the enhanced feature vectors of each interactive object.

[0081] In some embodiments, the prediction module 505 is configured to perform splicing processing on the user's feature vector, the enhanced feature vectors of each candidate object, and the enhanced feature vectors of each interactive object to obtain each fused feature vector; and input each fused feature vector into the third multi-layer perceptron for feature transformation, so as to perform probability prediction processing on each candidate object and obtain the predicted probability of each candidate object.

[0082] In some embodiments, the object recommendation device is configured to obtain a training data set of a pre-trained model before performing vector representation on each candidate object feature data in the candidate object feature data set through the pre-trained model, wherein the training data set contains multiple training object feature data, and each training object feature data is arranged based on the order of user interaction; any one training object feature data in the training data set is selected as the first object feature data, and the training object feature data after the first object feature data is the second object feature data, and all the first object feature data and the second object feature data in the training data set are used as the third object feature data; the first object feature data and the second object feature data are selected to construct a positive sample pair, and the first object feature data and the third object feature data are selected to construct a positive sample pair. Create multiple negative sample pairs; input the positive sample pairs and each negative sample pair into the pre-training model, perform multimodal vector representation on the feature data of the first object through the pre-training model to obtain the representation feature vector of the first object, perform multimodal vector representation on the feature data of the second object through the pre-training model to obtain the representation feature vector of the second object, perform vector representation on the feature data of each third object through the pre-training model to obtain the representation feature vector of each third object; based on the representation feature vector of the first object, the representation feature vector of the second object and the representation feature vector of each third object, obtain the loss value corresponding to the information-noise contrast loss function; and update the parameters of the pre-training model based on the loss value, and obtain a trained pre-training model when the loss value is less than or equal to the preset value.

[0083] In some embodiments, the object recommendation device is configured to obtain a loss value corresponding to the information-noise contrast loss function based on the characterization feature vector of the first object, the characterization feature vector of the second object, and the characterization feature vectors of each third object, including multiplying the characterization feature vector of the first object and the characterization feature vector of the second object to obtain the product of the positive samples; multiplying the characterization feature vector of the first object with the characterization feature vectors of each third object respectively to obtain the product of each negative sample; adding the products of each negative sample to obtain the sum of the negative sample products; and obtaining the loss value corresponding to the information-noise contrast loss function based on the product of the positive samples and the sum of the negative sample products.

[0084] It should be understood that the size of the serial numbers of the steps in the above embodiments does not mean the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of the present disclosure.

[0085] Figure 6 FIG. 6 is a schematic diagram of an electronic device 6 provided in an embodiment of the present disclosure. Figure 6As shown, the electronic device 6 of this embodiment includes: a processor 601, a memory 602, and a computer program 603 stored in the memory 602 and executable by the processor 601. When the processor 601 executes the computer program 603, the steps of the above-described method embodiments are implemented. Alternatively, when the processor 601 executes the computer program 603, the functions of the modules / units in the above-described device embodiments are implemented.

[0086] The electronic device 6 may be a desktop computer, a notebook, a PDA, a cloud server or other electronic device. The electronic device 6 may include but is not limited to a processor 601 and a memory 602. It will be understood by those skilled in the art that Figure 6 This is merely an example of the electronic device 6 and does not limit the electronic device 6 . The electronic device 6 may include more or fewer components than shown in the figure, or different components.

[0087] The processor 601 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.

[0088] The memory 602 can be an internal storage unit of the electronic device 6, such as a hard disk or memory of the electronic device 6. The memory 602 can also be an external storage device of the electronic device 6, such as a plug-in hard disk, a smart media card (SMC), a secure digital (SD) card, a flash memory card, etc. The memory 602 can also include both an internal storage unit of the electronic device 6 and an external storage device. The memory 602 is used to store computer programs and other programs and data required by the electronic device.

[0089] Those skilled in the art will clearly understand that for the sake of convenience and brevity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In actual applications, the above-mentioned functions can be distributed and completed by different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The above-mentioned integrated units can be implemented in the form of hardware or in the form of software functional units.

[0090] If the integrated module / unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a readable storage medium (such as a computer-readable storage medium). Based on this understanding, the present disclosure implements all or part of the processes in the above-mentioned embodiment method, and can also be completed by instructing the relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium, and the computer program can implement the steps of the above-mentioned various method embodiments when executed by the processor. The computer program may include computer program code, and the computer program code may be in source code form, object code form, executable file or some intermediate form, etc. The computer-readable storage medium may include: any entity or device that can carry computer program code, recording medium, U disk, mobile hard disk, magnetic disk, optical disk, computer memory, read-only memory (ROM), random access memory (RAM), electric carrier signal, telecommunication signal and software distribution medium, etc.

[0091] The above embodiments are only used to illustrate the technical solutions of the present disclosure, rather than to limit them. Although the present disclosure has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present disclosure, and should all be included in the scope of protection of the present disclosure.

Claims

1. A method for object recommendation, characterized in that: include: Acquire a candidate object feature dataset, a user's interactive object feature dataset, and feature data of the user; Performing multimodal vector representation on each candidate object feature data in the candidate object feature dataset using a pre-trained model to obtain a representation feature vector for each candidate object, and performing multimodal vector representation on each interactive object feature data in the interactive object feature dataset using the pre-trained model to obtain a representation feature vector for each interactive object; Performing vector embedding on the feature data of the user to obtain an initial feature vector of the user, and performing cross processing on the initial feature vector of the user to obtain a feature vector of the user; performing a nonlinear transformation on the representation feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object, and performing a nonlinear transformation on the representation feature vector of each interactive object based on the recurrent neural network to obtain a temporal dependency relationship between the representation feature vectors of each interactive object to obtain an enhanced feature vector of each interactive object; Performing probability prediction on each candidate object based on the user's feature vector, the enhanced feature vector of each candidate object, and the enhanced feature vector of each interactive object to obtain a predicted probability of each candidate object; Selecting a target object to recommend to the user based on the predicted probability of each candidate object; The candidate object feature data includes a plurality of text features of the candidate object and a plurality of identification features of the candidate object; The method of performing multimodal vector representation on each candidate object feature data in the candidate object feature dataset using a pre-trained model to obtain a representation feature vector of each candidate object includes: Inputting each text feature of each candidate object into a text encoder for encoding processing to obtain each text encoding result of each candidate object, and inputting each text encoding result of each candidate object into a whitening layer for fixed embedding to obtain a target feature vector of each text of each candidate object; Performing vector embedding on multiple identification features of each candidate object to obtain an initial feature vector of each identification feature of each candidate object; and inputting the initial feature vector of each identification feature of each candidate object into a first multi-layer perceptron for feature transformation to obtain a feature vector of each identification feature of each candidate object; Inputting the feature vectors of each identification feature of each candidate object into the whitening layer for fixed embedding to obtain the target feature vectors of each identification feature of each candidate object; Based on the target feature vectors of the respective texts of the respective candidate objects and the target feature vectors of the respective identification features of the respective candidate objects, a representation feature vector of each candidate object is determined.

2. The object recommendation method according to claim 1, characterized in that: The determining of the representation feature vector of each candidate object based on the target feature vector of each text of each candidate object and the target feature vector of each identification feature of each candidate object includes: splicing the target feature vectors of each text of each candidate object and the target feature vectors of each identification feature of each candidate object to obtain an initial fused feature vector of each candidate object; The initial fused feature vectors of each candidate object are input into a second multi-layer perceptron for feature transformation to obtain a fused feature vector of each candidate object, and the fused feature vectors of each candidate object are normalized to obtain a representation feature vector of each candidate object.

3. The object recommendation method according to claim 1, characterized in that: The performing of nonlinear transformation on the representation feature vector of each candidate object based on the recurrent neural network to obtain an enhanced feature vector of each candidate object includes: Inputting the characterization feature vector of each candidate object into a recurrent neural network, calculating a current hidden layer state based on the characterization feature vector of each candidate object and the state of the previous hidden layer, and inputting the current hidden layer state into the activation function layer of the recurrent neural network to obtain an enhanced feature vector of each candidate object; The performing of nonlinear transformation on the representation feature vectors of each of the interactive objects based on the recurrent neural network, obtaining the temporal dependency relationship between the representation feature vectors of each of the interactive objects, and obtaining the enhanced feature vectors of each of the interactive objects includes: The characterization feature vectors of each interactive object are input into the recurrent neural network based on the user's interaction order, the current hidden layer state is calculated based on the characterization feature vectors of each interactive object and the state of the previous hidden layer, and the current hidden layer state is input into the activation function layer of the recurrent neural network to obtain the enhanced feature vectors of each interactive object.

4. The object recommendation method according to claim 1, characterized in that: The performing probability prediction on each candidate object based on the feature vector of the user, the enhanced feature vector of each candidate object, and the enhanced feature vector of each interactive object to obtain the predicted probability of each candidate object includes: performing splicing processing on the user's feature vector, the enhanced feature vectors of each candidate object, and the enhanced feature vectors of each interactive object to obtain fused feature vectors; Each fused feature vector is input into a third multi-layer perceptron for feature transformation, so as to perform probability prediction processing on each candidate object and obtain a predicted probability of each candidate object.

5. The object recommendation method according to claim 1, characterized in that: Before performing multimodal vector representation on each candidate object feature data in the candidate object feature dataset using the pre-trained model, the method further includes: Obtaining a training data set of a pre-trained model, wherein the training data set includes a plurality of training object feature data, and each of the training object feature data is arranged based on a user interaction order; Selecting any one training object feature data in the training data set as first object feature data, the training object feature data following the first object feature data as second object feature data, and all the training object feature data except the first object feature data and the second object feature data as third object feature data; selecting the first object feature data and the second object feature data to construct a positive sample pair, and selecting the first object feature data and the third object feature data to construct multiple negative sample pairs; Inputting the positive sample pair and each of the negative sample pairs into the pre-trained model, performing multimodal vector representation on the feature data of the first object using the pre-trained model to obtain a representation feature vector of the first object, performing multimodal vector representation on the feature data of the second object using the pre-trained model to obtain a representation feature vector of the second object, and performing vector representation on each of the feature data of the third object using the pre-trained model to obtain a representation feature vector of each of the third objects; Based on the characterization feature vector of the first object, the characterization feature vector of the second object, and the characterization feature vectors of each of the third objects, a loss value corresponding to the information-noise contrast loss function is obtained; and based on the loss value, the parameters of the pre-trained model are updated, and when the loss value is less than or equal to a preset value, a pre-trained model that has been trained is obtained.

6. The object recommendation method according to claim 5, characterized in that: The obtaining, based on the characterization feature vector of the first object, the characterization feature vector of the second object, and the characterization feature vectors of each of the third objects, a loss value corresponding to the information-noise contrast loss function includes: multiplying the characterizing feature vector of the first object and the characterizing feature vector of the second object to obtain the product of the positive sample; Multiplying the characterizing feature vector of the first object by the characterizing feature vectors of each of the third objects respectively to obtain products of each of the negative samples; Adding the products of the negative samples to obtain the negative sample product sum; Based on the product of the positive samples and the sum of the products of the negative samples, a loss value corresponding to the information-noise contrast loss function is obtained.

7. An object recommendation device, characterized in that: include: An acquisition module, configured to acquire a candidate object feature dataset, a user's interactive object feature dataset, and the user's feature data; a multimodal characterization module, configured to perform multimodal vector characterization on each candidate object feature data in the candidate object feature dataset using a pre-trained model to obtain a characterization feature vector for each candidate object, and to perform multimodal vector characterization on each interactive object feature data in the interactive object feature dataset using the pre-trained model to obtain a characterization feature vector for each interactive object; A user characterization module is configured to perform vector embedding on the user's feature data to obtain the user's initial feature vector, and perform cross processing on the user's initial feature vector to obtain the user's feature vector; a vector enhancement module, configured to perform a nonlinear transformation on the representation feature vector of each candidate object based on a recurrent neural network to obtain an enhanced feature vector of each candidate object, and to perform a nonlinear transformation on the representation feature vector of each interactive object based on the recurrent neural network to obtain a temporal dependency relationship between the representation feature vectors of each interactive object to obtain an enhanced feature vector of each interactive object; a prediction module, configured to perform probability prediction on each candidate object based on the user's feature vector, the enhanced feature vector of each candidate object, and the enhanced feature vector of each interactive object, to obtain a predicted probability of each candidate object; A recommendation module, configured to select a target object to recommend to the user based on the predicted probability of each candidate object; The candidate object feature data includes multiple text features of the candidate object and multiple identification features of the candidate object; the multimodal representation module is specifically used to: input each text feature of each candidate object into a text encoder for encoding processing to obtain each text encoding result of each candidate object, and input each text encoding result of each candidate object into a whitening layer for fixed embedding to obtain a target feature vector of each text of each candidate object; perform vector embedding on multiple identification features of each candidate object to obtain an initial feature vector of each identification feature of each candidate object; and input the initial feature vector of each identification feature of each candidate object into a first multi-layer perceptron for feature transformation to obtain a feature vector of each identification feature of each candidate object; input the feature vector of each identification feature of each candidate object into the whitening layer for fixed embedding to obtain a target feature vector of each identification feature of each candidate object; based on the target feature vector of each text of each candidate object and the target feature vector of each identification feature of each candidate object, determine the representation feature vector of each candidate object.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 6 are implemented.

9. A readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 6 are implemented.

Images