Cold start object recommendation method and device, computer device and storage medium
By utilizing a meta-mapping network for multi-layer perception processing and mapping transformation during cold start, the problem of inaccurate recommendations during cold start is solved, achieving accurate recommendations even in the absence of interaction data.
Patent Information
- Application Number
- CN202210326097.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-03-30
- Publication Date
- 2026-05-19
- Estimated Expiration
- 2042-03-30
AI Technical Summary
Traditional object recommendation methods cannot achieve accurate recommendations in the cold start state because of the lack of interaction data or the limited amount of interaction data, making it impossible to determine user preferences.
By acquiring the content features and initial preference features of the target object in the cold start state, multi-layer perceptual processing is performed using a meta-mapping network to determine the feature mapping relationship and perform mapping transformation to obtain updated preference features that approximate the spatial distribution of the target preference features, and finally recommendation analysis is performed.
It improves recommendation accuracy in cold start conditions, enabling it to learn accurate user preference features even in the absence of interaction data, thus achieving more accurate recommendation results.
Smart Images

Figure CN116932862B_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of computer technology, and in particular to a method, apparatus, computer device, storage medium, and computer program product for recommending cold start objects. Background Technology
[0002] With the development of computer technology, more and more human-computer interaction scenarios have emerged. For example, users' daily interactions with objects such as short videos, images, and articles generally require users to actively search for their preferred objects. Given the massive amount of interactive objects provided by the system, the process of actively searching is complex and consumes significant data processing resources. To solve this problem, proactively recommending objects that users prefer can greatly improve the effectiveness of interaction and significantly facilitate users' daily lives and entertainment.
[0003] However, traditional object recommendation methods generally recommend objects to users based on user-object interaction data to determine the users who prefer the object. But this object recommendation method relies on user-object interaction data. In the cold start state, due to the lack of interaction data or the limited interaction data for objects, accurate object recommendations cannot be achieved. Summary of the Invention
[0004] Therefore, it is necessary to provide a cold-start object recommendation method, apparatus, computer device, computer-readable storage medium, and computer program product that can improve the accuracy of recommendations, addressing the aforementioned technical problems.
[0005] Firstly, this application provides a cold start object recommendation method. The method includes:
[0006] Obtain the content features and initial preference features of the target object in the cold start state;
[0007] The content features are processed by a meta-mapping network to determine the feature mapping relationship that matches the target object; the meta-mapping network is trained based on the interaction data of the recommended object in a simulated cold start state.
[0008] The initial preference features are mapped and transformed according to the feature mapping relationship to obtain updated preference features that approximate the target preference feature spatial distribution; the target preference feature spatial distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended object;
[0009] Based on the updated preference features, a recommendation analysis is performed on the target object to obtain recommendation results for the target object.
[0010] Secondly, this application also provides a cold start object recommendation device. The device includes:
[0011] The feature acquisition module is used to acquire the content features and initial preference features of the target object in the cold start state;
[0012] The feature processing module is used to perform multi-layer perceptual processing on the content features based on the meta-mapping network to determine the feature mapping relationship that matches the target object; the meta-mapping network is trained based on the interaction data of the recommended object in a simulated cold start state;
[0013] The mapping transformation module is used to perform mapping transformation on the initial preference features according to the feature mapping relationship to obtain updated preference features that approximate the target preference feature space distribution; the target preference feature space distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended object;
[0014] The recommendation analysis module is used to perform recommendation analysis on the target object based on the updated preference features, and obtain recommendation results for the target object.
[0015] In one embodiment, the feature acquisition module includes a content feature acquisition module and an initial preference feature acquisition module;
[0016] The content feature acquisition module is used to acquire content description data of the target object, perform vector transformation on the content description data, and obtain the content features of the target object.
[0017] The initial preference feature acquisition module is used to acquire the preference features corresponding to each of the recommended objects that satisfy the target preference feature spatial distribution; perform global average pooling on each preference feature, and use the obtained global average preference feature as the initial preference feature of the target object of the cold start.
[0018] In one embodiment, the target object includes a target video; the content feature acquisition module is further configured to acquire supplementary descriptive information for the target video and extract key video frames from the target video; perform image recognition processing on the key video frames to obtain video content information in the key video frames; and determine the supplementary descriptive information and the video content information as the content description data of the target object.
[0019] In one embodiment, the feature mapping relationship includes a stretching function; the mapping transformation module includes a stretching module, used to stretch the feature distribution of the initial preference feature according to the stretching function, integrate the content feature into the stretched space of the initial preference feature, and obtain an updated preference feature that approximates the target preference feature space distribution.
[0020] In one embodiment, the feature processing module includes a stretching function determination module, which performs multilayer perceptual processing on the content features based on the meta-stretching network parameters in the meta-mapping network to obtain a stretching function that stretches the initial preference features to approximate the spatial distribution of the target preference features.
[0021] In one embodiment, the feature mapping relationship further includes an offset function; the stretching module is used to stretch the feature distribution of the initial preference feature based on the stretching function, and to integrate the content feature into the stretching space of the initial preference feature to obtain the intermediate preference feature;
[0022] The mapping transformation module further includes an offset module, which is used to offset and adjust the feature distribution of the intermediate preference features based on the offset function, so as to remove noise data in the intermediate preference features and obtain updated preference features.
[0023] In one embodiment, the feature processing module includes an offset function determination module, used to acquire interaction data for the target object; and to perform multi-layer perceptual processing on the interaction data based on the meta-offset network parameters in the meta-mapping network to obtain an offset function that makes the intermediate preference features approach the spatial distribution of the target preference features.
[0024] In one embodiment, the cold start object recommendation device further includes a feature aggregation module, which is used to aggregate the interaction preference features in the interaction data to obtain aggregated interaction features;
[0025] The offset function determination module is further configured to perform multi-layer perceptual processing on the aggregated interaction features based on the meta-offset network parameters in the meta-mapping network, to obtain an offset function that makes the intermediate preference features approach the spatial distribution of the target preference features.
[0026] In one embodiment, the cold-start object recommendation device further includes a meta-mapping network training module, used to obtain recommendation feedback labels from the cold-start simulation data of the recommended objects and determine the recommendation results for the recommended objects; calculate the loss data between the recommendation results and the recommendation feedback labels; and iteratively train the parameters of the initial meta-mapping network based on the loss data using a gradient descent optimization algorithm until the training stopping condition is met, thereby obtaining the meta-mapping network.
[0027] In one embodiment, the meta-mapping network training module is further configured to, based on the initial meta-mapping network, convert the initial preference features of the recommended object into updated preference features that conform to the target preference feature spatial distribution; calculate, respectively, a first loss data between the recommendation result corresponding to the initial preference features and the recommendation feedback label, and a second loss data between the recommendation result corresponding to the updated preference features and the recommendation feedback label; the first loss data is used to update the parameters of the initial meta-mapping network; the second loss data is used to update the initial preference features.
[0028] In one embodiment, the cold start object recommendation device further includes a training data acquisition module, used to select sample objects from the sample objects whose interaction data meets the interaction conditions and whose recommendation effect reaches the expected effect; construct triples based on the recommended object data, the receiving object data and the recommendation feedback tags in the interaction data of the sample objects; and use the triples as interaction data for parameter training of the initial meta-mapping network.
[0029] Thirdly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the computer program to perform the following steps:
[0030] Obtain the content features and initial preference features of the target object in the cold start state;
[0031] The content features are processed by a meta-mapping network to determine the feature mapping relationship that matches the target object; the meta-mapping network is trained based on the interaction data of the recommended object in a simulated cold start state.
[0032] The initial preference features are mapped and transformed according to the feature mapping relationship to obtain updated preference features that approximate the target preference feature spatial distribution; the target preference feature spatial distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended object;
[0033] Based on the updated preference features, a recommendation analysis is performed on the target object to obtain recommendation results for the target object.
[0034] Fourthly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, performs the following steps:
[0035] Obtain the content features and initial preference features of the target object in the cold start state;
[0036] The content features are processed by a meta-mapping network to determine the feature mapping relationship that matches the target object; the meta-mapping network is trained based on the interaction data of the recommended object in a simulated cold start state.
[0037] The initial preference features are mapped and transformed according to the feature mapping relationship to obtain updated preference features that approximate the target preference feature spatial distribution; the target preference feature spatial distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended object;
[0038] Based on the updated preference features, a recommendation analysis is performed on the target object to obtain recommendation results for the target object.
[0039] Fifthly, this application also provides a computer program product. The computer program product includes a computer program that, when executed by a processor, performs the following steps:
[0040] Obtain the content features and initial preference features of the target object in the cold start state;
[0041] The content features are processed by a meta-mapping network to determine the feature mapping relationship that matches the target object; the meta-mapping network is trained based on the interaction data of the recommended object in a simulated cold start state.
[0042] The initial preference features are mapped and transformed according to the feature mapping relationship to obtain updated preference features that approximate the target preference feature spatial distribution; the target preference feature spatial distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended object;
[0043] Based on the updated preference features, a recommendation analysis is performed on the target object to obtain recommendation results for the target object.
[0044] The aforementioned cold-start object recommendation method, apparatus, computer device, storage medium, and computer program product acquire the content features and initial preference features of the target object in a cold-start state. This is achieved by using a meta-mapping network trained on interaction data of the recommended object in a simulated cold-start state. The content features undergo multi-layer perceptual processing to determine the feature mapping relationship matching the target object. The initial preference features are then mapped and transformed according to this relationship to obtain updated preference features that approximate the spatial distribution of the target preference features. Since the spatial distribution of the updated preference features approximates the spatial distribution of the preference features corresponding to the recommended object, it can be determined that the target object in a cold-start state can learn accurate preference features. Furthermore, recommendation analysis is performed on the target object based on the updated preference features to obtain more accurate recommendation results. Attached Figure Description
[0045] Figure 1 This is an application environment diagram of the cold start object recommendation method in one embodiment;
[0046] Figure 2 This is a flowchart illustrating a cold start object recommendation method in one embodiment;
[0047] Figure 3 This is a schematic diagram of a target object in a cold start state in one embodiment;
[0048] Figure 4 This is a schematic diagram of the structure of a multilayer sensing processor in one embodiment;
[0049] Figure 5 This is a schematic diagram of the interface recommended for displaying the cold start object in one embodiment;
[0050] Figure 6 This is a flowchart illustrating the cold start object recommendation method in another embodiment;
[0051] Figure 7 This is a flowchart illustrating the process of determining the content features of a short video in one embodiment;
[0052] Figure 8 This is a structural block diagram of a cold start object recommendation device in one embodiment;
[0053] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation
[0054] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0055] The solutions provided in this application can involve technologies such as Artificial Intelligence (AI) and Machine Learning (ML). Artificial intelligence is the theory, technology, and application system that uses digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, perceive the environment, acquire knowledge, and use that knowledge to obtain optimal results. In other words, artificial intelligence is a comprehensive technology in computer science that attempts to understand the essence of intelligence and produce a new kind of intelligent machine that can react in a way similar to human intelligence. Artificial intelligence studies the design principles and implementation methods of various intelligent machines, enabling them to have perception, reasoning, and decision-making functions. Machine learning is a multidisciplinary field involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory, and many other disciplines. It specifically studies how computers can simulate or implement human learning behavior to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is the core of artificial intelligence and the fundamental way to give computers intelligence; its applications span all areas of artificial intelligence. Based on technologies such as artificial intelligence and machine learning, it is possible to acquire the content features and initial preference features of the target object in a cold start state. Based on the meta-mapping network trained by the interaction data of the recommended object in a simulated cold start state, multi-layer perceptual processing is performed on the content features to determine the feature mapping relationship that matches the target object. The initial preference features are then mapped and transformed according to the feature mapping relationship to obtain updated preference features that approximate the spatial distribution of the target preference features. This makes the spatial distribution of the updated preference features approximate the spatial distribution of the preference features corresponding to the recommended object. Based on the updated preference features, recommendation analysis is performed on the target object to obtain accurate recommendation results.
[0056] The cold start object recommendation method provided in this application embodiment can be applied to, for example, Figure 1In the application environment shown, terminal 102 communicates with server 104 via a network. A data storage system can store the data that server 104 needs to process. The data storage system can be integrated onto server 104 or placed on the cloud or other servers. Server 104 acquires the content features and initial preference features of the target object in a cold start state. Based on a meta-mapping network trained on the interaction data of the recommended object in a simulated cold start state, it performs multi-layer perceptual processing on the content features to determine the feature mapping relationship matching the target object. According to the feature mapping relationship, it performs mapping transformation on the initial preference features to obtain updated preference features that approximate the spatial distribution of the target preference features. Server 104 performs recommendation analysis on the target object based on the updated preference features to obtain recommendation results for the target object. When the recommendation results meet the recommendation conditions, the target object is recommended to the terminal 102 where the user indicated by the recommendation results is located for display. Terminal 102 can be, but is not limited to, various desktop computers, laptops, smartphones, tablets, IoT devices, and portable wearable devices. IoT devices can be smart speakers, smart TVs, smart air conditioners, smart in-vehicle devices, etc. Portable wearable devices can include smartwatches, smart bracelets, and head-mounted devices. Server 104 can be implemented using a standalone server or a server cluster consisting of multiple servers.
[0057] In one embodiment, such as Figure 2 As shown, a cold start object recommendation method is provided. This application embodiment can be applied to information recommendation and display applications, accurately recommending information that users may be interested in. This recommendation scenario can be text recommendation, video recommendation, or music recommendation, etc. This application embodiment will use a short video recommendation scenario as an example to describe in detail the cold start object recommendation method provided by this application. This scenario is only an example and does not limit the scope of protection of the technical solution of this application. The method is applied to... Figure 1 Taking a computer device as an example, the explanation includes the following steps:
[0058] Step 202: Obtain the content features and initial preference features of the target object in the cold start state.
[0059] The target object refers to the recommended object presented on the terminal device that can interact with the receiving object, including but not limited to displayed product objects, audio and video objects, e-book objects, etc. The receiving object is an actual object, including but not limited to the user and / or the device used by the user, such as a client. The specific form of interaction between the receiving object and the recommended object can vary depending on the scenario, including but not limited to: liking, forwarding, browsing, ordering, collecting, sharing, commenting, etc. Specifically, the target object for the receiving object to interact with may also be different in different scenarios. For example, in a multimedia scenario, the target object can be a multimedia file that has been liked, collected, or forwarded, such as music, short videos, articles, etc.; in an e-commerce scenario, the target object can be items added to the shopping cart or items that have already been purchased; in an advertising scenario, the target object can be an advertisement that has been viewed, clicked, or collected, etc.
[0060] In a recommendation system, the number of recommended objects is constantly increasing. For newly added recommended objects, there is a lack of interaction data with the target audience. The process of entering the recommendation system without referable interaction data is called the cold start process, and the state of the newly added recommended object during the cold start process is called the cold start state. Specifically, taking a short video recommendation system as an example, a newly uploaded short video by a creator is a target object in the cold start state. Since there are no other users' interaction records such as watching, liking, or sharing the short video, the recommendation system has difficulty identifying the user preferences to which the short video belongs, which can easily lead to inaccurate target audience recommendations and affect the recommendation conversion rate. This is the cold start problem of short videos.
[0061] Content features are features extracted from content data that matches the target object. This content data includes a detailed description of the target object. For example... Figure 3 As shown, the content data may include available information related to the content, such as the title, category, and description, that the user fills in when uploading the target object, as well as information obtained by extracting content from the target object.
[0062] Preference features are used to characterize the preferences of a target object for the receiving objects it interacts with. Initial preference features are the preference features assigned to the target object in the cold start state. Specifically, initial preference features can be preference features randomly generated by the recommendation system, or they can be generated based on the feature distribution characteristics of all existing recommendation objects that have sufficient interaction data and have been fully learned.
[0063] Specifically, in response to a newly added target object event, the computer device searches for the corresponding interaction data, determines the current state of the target object based on the amount of interaction data, and determines the state of the newly added target object as a cold start state when the amount of interaction data is zero or less than a set data threshold. Based on the content data contained in the target object, the computer device extracts the content features of the target object and determines the initial preference features of the target object based on the object identification information.
[0064] Step 204: Perform multi-layer perceptual processing on content features based on the meta-mapping network to determine the feature mapping relationship that matches the target object; the meta-mapping network is trained based on the interaction data of the recommended object in a simulated cold start state.
[0065] Among them, the meta-mapping network is a network model trained on the interaction data of the recommended object in the simulated cold start state. The recommended object can be a recommended object in the recommendation system whose existing interaction data reaches the quantity threshold and whose recommendation success rate reaches the probability threshold. The quantity threshold indicates that the recommended object has sufficient interaction data, and the probability threshold indicates that the recommended object's preference data has been fully learned in the recommendation system, enabling accurate recommendations.
[0066] Simulated cold start refers to the cold start state simulated by filtering recommended objects through interactive data in a non-cold start state. By simulating the cold start state of recommended objects, we can fully learn the spatial distribution characteristics of preference features in the cold start state and the preference features of recommended objects in the non-cold start state. This allows us to train the network parameters in the initial meta-mapping network so that the trained meta-mapping network can generate a one-to-one feature mapping relationship for the target object. Then, through the feature mapping relationship, the initial preference features of the target object are transformed into updated preference features that approximate the spatial distribution of the preference features corresponding to the recommended object.
[0067] Multilayer perceptron processing refers to the process of mapping data from multiple dimensions onto a single data point. Specifically, multilayer perceptron mapping can be implemented using a multilayer perceptron (also known as an artificial neural network model). Figure 4As shown, a multilayer perceptron includes an input and output layer, and can have multiple hidden layers. The layers in a multilayer perceptron are fully connected. Specifically, the bottom layer of a multilayer perceptron is the input layer, followed by hidden layers, and finally the output layer. The neurons in the hidden layers are fully connected to the input layer. Assuming the input layer is represented by a vector X, the output of the hidden layer is f(W1X+b1), where W1 is the connection coefficient between the input and hidden layers, and b1 is the bias from the input layer to the hidden layer. The function f can be a commonly used sigmoid function or a tanh function. The regression from the hidden layer to the output layer is a multi-class logistic regression, also known as softmax regression. The output of the output layer is softmax(W2X1+b2), where X1 represents the output of the hidden layer f(W1X+b1), W2 is the connection coefficient between the hidden and output layers, and b2 is the bias from the hidden layer to the output layer.
[0068] Multilayer perceptron processing essentially uses content features as input data to a multilayer perceptron (MLP). Logical analysis is performed based on the trained parameters within each layer of the MLP, and the output is a feature mapping relationship matching the target object. The parameters of each layer in the MLP include the connection weights and biases between layers, such as W1, b1, W2, and b2. The parameters of each layer in the MLP are determined through model training. The specific training process includes: first, randomly initializing all parameters of the MLP; then iteratively training, continuously calculating gradients and updating parameters until a training termination condition is met. This termination condition can be at least one of the following: accuracy meets accuracy requirements, error meets error requirements, or the number of iterations meets a requirement.
[0069] Feature mapping relationships are data processing rules that must be followed when mapping and transforming initial preference features. Specifically, feature mapping relationships can include at least one of the following processing logics: feature space stretching, feature space shifting, etc. Specifically, feature mapping relationships can involve stretching the initial preference features in their feature space to incorporate content features, shifting the initial preference features in their feature space to filter out noise features, or first stretching the initial preference features in their feature space to incorporate content features, and then shifting the feature space to filter out noise features. The specific approach can be set according to the actual needs of the scenario.
[0070] Specifically, after identifying the content features of the target object, the computer device performs multi-layer perceptual processing on the content features of the target object based on a meta-mapping network trained using interaction data of the recommended object in a simulated cold start state, thereby obtaining a feature mapping relationship that matches the target object. When there are multiple target objects, the computer device can simultaneously perform multi-layer perceptual processing on each target object using multiple processes based on the trained meta-mapping network, obtaining a feature mapping relationship that matches each target object.
[0071] Step 206: Perform a mapping transformation on the initial preference features according to the feature mapping relationship to obtain updated preference features that approximate the target preference feature spatial distribution; the target preference feature spatial distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended object.
[0072] Among them, mapping transformation is the process of transforming the feature space distribution of the initial preference features according to the feature mapping relationship, such as stretching the feature space or shifting the feature space of the initial preference features according to the feature mapping relationship.
[0073] The target preference feature space distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended objects. Specifically, a recommendation system trained on a large amount of interaction data of the recommended objects allocates a specific feature space to the preference features of the recommended objects. This means that the preference features of the recommended objects satisfy a certain distribution, i.e., the target preference feature space distribution. However, in the cold start state, before a target object records enough interaction data, the spatial distribution of its preference features differs from the target preference feature space distribution. After the target object in the cold start state learns preference information from sufficient interaction data, its preference feature distribution will gradually converge to the target preference feature space distribution.
[0074] Updated preference features are the result of mapping and transforming the initial preference features. The feature space of updated preference features is closer to the distribution of the target preference feature space than the initial preference features. Recommendation analysis based on updated preference features can avoid the drawback of a large difference between the initial preference features and the target preference feature space distribution due to insufficient interaction data in the cold start state. Recommendation analysis based on updated preference features that are close to the distribution of the target preference feature space yields more accurate recommendation results.
[0075] Specifically, after obtaining the feature mapping relationship between the two target objects, the computer device performs a mapping transformation on the initial preference features corresponding to the target objects according to the feature mapping relationship, and obtains updated preference features that approximate the spatial distribution of the target preference features, and then performs recommendation analysis based on the updated preference features.
[0076] Step 208: Perform recommendation analysis on the target object based on the updated preference features to obtain recommendation results for the target object.
[0077] The recommendation analysis can be an analysis of the target audience's preferred group of recipients, or an analysis to determine whether a recommendation is needed for a specific recipient. Accordingly, when the recommendation analysis is for the target audience's preferred group of recipients, the recommendation result is to recommend the target audience to recipients who meet the recommendation criteria based on updated preference features. When the recommendation analysis is to determine whether a recommendation is needed for a specific recipient, the recommendation result includes recommending the target audience to the specified recipient or not recommending the target audience to the specified recipient. For example... Figure 5 As shown, when user A uploads a new short video 1 on the first terminal, short video 1 is in a cold start state in the recommendation system. Through the above-mentioned cold start object recommendation method, the second terminal is determined to be the preferred receiving object of short video 1, and thus short video 1 is recommended and displayed on the second terminal.
[0078] The aforementioned cold-start object recommendation method acquires the content features and initial preference features of the target object in a cold-start state. It is based on a meta-mapping network trained on the interaction data of the recommended object in a simulated cold-start state. The content features are processed by multi-layer perception to determine the feature mapping relationship that matches the target object. The initial preference features are then mapped and transformed according to the feature mapping relationship to obtain updated preference features that approximate the spatial distribution of the target preference features. Since the spatial distribution of the updated preference features approximates the spatial distribution of the preference features corresponding to the recommended object, it can be determined that the target object in the cold-start state can learn accurate preference features. Then, recommendation analysis is performed on the target object based on the updated preference features to obtain more accurate recommendation results.
[0079] In one embodiment, such as Figure 6 As shown, the content features and initial preference features of the target object for cold start are obtained, including:
[0080] Step 602: Obtain the content description data of the target object, perform vector transformation on the content description data, and obtain the content features of the target object.
[0081] Step 604: Obtain the preference features corresponding to each of the recommended objects that satisfy the target preference feature spatial distribution.
[0082] Step 606: Perform global average pooling on each preference feature, and use the resulting global average preference feature as the initial preference feature of the target object for cold start.
[0083] Content description data refers to the raw data used to represent detailed information about the target object, which may include title, category, brief description, and detailed content. Vector transformation refers to the process of converting content description data into continuous vectors. Vector transformation can be achieved through embedding techniques, which are techniques that transform raw data into continuous vectors. The resulting continuous vectors can be called embedded features. Computer recognition processes the raw content description data by inputting it into the embedding layer, obtaining the embedded features corresponding to the content description data. Recommended objects that satisfy the target preference feature space distribution are those in the recommendation system that have sufficient interaction data and have been fully learned. Global average pooling refers to the process of averaging global features and outputting the average feature value.
[0084] In a specific application, taking short videos as the target object as an example, when a short video enters the recommendation system during a cold start, the recommendation model randomly initializes an initial preference feature for the target object. However, this randomly initialized initial preference feature contains almost no information related to the cold start short video, failing to achieve accurate recommendation results. By performing a global average pooling operation on the existing short video preference features, which have sufficient interaction data and have been fully learned, a global average preference feature is obtained and used as a generalized preference feature. When a cold start short video enters the recommendation system, i.e., for the cold start video in the zero-sample stage, the generalized preference feature is used as the initial preference feature for the cold start short video. This preference feature initialization method provides the cold start short video with an initial preference feature that is as close as possible to the spatial distribution of the target preference feature.
[0085] In this embodiment, by performing a global average pooling operation on the preference features of the recommended object, the global average preference features are obtained as the initial preference features of the target object. This makes the initial preference features as close as possible to the spatial distribution of the target preference features in the initial stage, thereby accelerating the progress of the target object adapting to the recommendation system for accurate recommendation.
[0086] In one embodiment, the target object includes a target video; obtaining the content description data of the target object includes: obtaining supplementary description information for the target video and extracting key video frames from the target video; performing image recognition processing on the key video frames to obtain video content information in the key video frames; and determining the supplementary description information and video content information as the content description data of the target object.
[0087] Among them, such as Figure 7As shown, supplementary descriptive information can be titles, categories, brief descriptions, and other usable information directly edited by the user for the target video, related to the video content. Key video frames in the target video can be video frames extracted from the target video according to specified nodes, such as extraction at specified time intervals or between specified frames. Image recognition processing specifically refers to the process of obtaining content information from key video frames, such as recognizing text in a video or object categories in an image.
[0088] Specifically, the computer device acquires associated information for the target video, obtains supplementary descriptive information for the target video, and extracts key video frames from the target video according to the video frame extraction parameters. By performing image recognition processing on the key video frames, the video content information in the key video frames is obtained. The computer device determines the supplementary descriptive information and video content information as the content description data of the target object and performs feature embedding processing to obtain the content features of the target object.
[0089] In this embodiment, for the target video, by extracting key video frames from the target video and performing image recognition processing, the supplementary descriptive content such as the video title, category, and description, as well as the content in the video frames, can be fully utilized to obtain content features. This allows for the extraction of the target video's preference information as much as possible, thus providing a foundation for subsequent accurate recommendations.
[0090] In one embodiment, the feature mapping relationship includes a stretching function; the mapping transformation of the initial preference features according to the feature mapping relationship to obtain updated preference features that approximate the target preference feature space distribution includes: stretching the feature distribution of the initial preference features according to the stretching function, incorporating the content features into the stretched space of the initial preference features, and obtaining updated preference features that approximate the target preference feature space distribution.
[0091] The stretching function is used to implement the initial preference feature D. c To the target preference feature space distribution D t The function of the stretching operation of the mapping can obtain the stretching space through the stretching operation, and then integrate the content features of the target object into the stretching space to obtain the updated preference features that approximate the distribution of the target preference feature space.
[0092] Specifically, stretching the feature distribution space can be achieved by specifying the stretching direction and stretching distance, such as stretching along the row direction of the feature vector or stretching along the column direction of the feature vector, or stretching along both the row and column directions at the same time. The stretching direction and stretching distance can be determined based on the stretching function.
[0093] In this embodiment, by performing a stretching transformation on the target object's preference features in a direction that adapts to the target preference feature space, the progress of the target object adapting to the recommendation system is accelerated, and recommendation results for the target object are obtained quickly and accurately.
[0094] In one embodiment, multi-layer perceptual processing of content features based on a meta-mapping network is performed to determine the feature mapping relationship that matches the target object. This includes: performing multi-layer perceptual processing of content features based on the meta-stretching network parameters in the meta-mapping network to obtain a stretching function that stretches the initial preference features to approximate the spatial distribution of the target preference features.
[0095] Among them, the meta-stretched network is used to realize the spatial distribution D of the initial preference features. c To the target preference feature space distribution D t The stretching operation of the mapping is performed by a stretching function generated by the meta-stretching network for each target object. The meta-stretching network parameters are the offset parameters obtained through training in the meta-mapping network and are required to generate the stretching operation on the preference features.
[0096] Specifically, the content features of a target object often contain some preference information; objects with similar preference features tend to have relatively similar content features. By using content features as input to a meta-stretching network, the network can extract preference-related information from the content features. The resulting stretching function stretches the target object's preference features towards D... t The meta-stretching network stretches the content features in a specific direction, incorporating preference-related information into the preference features. The expression for the meta-stretching network is as follows:
[0097]
[0098] Where w scale These are the parameters in the meta-stretched network, which can be iteratively updated through network training. h represents the multilayer perceptron model, specifically a two-layer multilayer perceptron model. Let D be the stretching function. c To D t The stretching operation of the mapping can be expressed as ⊙ indicates element-wise multiplication.
[0099] In this embodiment, the content features are processed by multi-layer perceptual processing through the meta-stretching network parameters in the meta-mapping network to obtain the stretching function. This makes the spatial distribution of the updated preference features obtained based on the stretching function approximate the spatial distribution of the preference features corresponding to the recommended object. This ensures that the target object in the cold start state can learn accurate preference features, thereby obtaining more accurate recommendation results for the target object.
[0100] In one embodiment, the feature mapping relationship further includes an offset function; stretching the feature distribution of the initial preference features according to the stretching function, and incorporating the content features into the stretched space of the initial preference features to obtain updated preference features that approximate the target preference feature space distribution, including: stretching the feature distribution of the initial preference features based on the stretching function, incorporating the content features into the stretched space of the initial preference features to obtain intermediate preference features; and adjusting the feature distribution of the intermediate preference features based on the offset function to remove noise data in the intermediate preference features to obtain updated preference features.
[0101] The offset function is used to implement the initial preference feature D. c To the target preference feature space distribution D t The function performs an offset operation on the mapping, filtering out noisy data from the stretched intermediate preference features to obtain updated preference features that approximate the target preference feature space distribution. The offset processing of the feature distribution space can be implemented based on a specified offset direction and offset distance, such as offsetting along the row direction or column direction of the feature vector, or sequentially offsetting along the row or column directions, etc. The offset direction and offset distance can be determined based on the offset function.
[0102] Specifically, the computer device first stretches the feature distribution of the initial preference features corresponding to the target object based on the stretching function to determine the stretching space of the initial preference features. Then, the content features corresponding to the target object are integrated into the stretching space of the initial preference features to obtain the stretched intermediate preference features. Then, based on the offset function, the feature distribution of the intermediate preference features is offset and adjusted to remove noise data in the intermediate preference features, so as to obtain the updated preference features corresponding to the target object.
[0103] In this embodiment, by sequentially performing stretching transformation and offset adjustment on the target object's preference features in the direction of adapting to the target preference feature space distribution, the updated preference features can be made closer to the target preference feature space distribution, thereby accelerating the progress of the target object adapting to the recommendation system and quickly and accurately obtaining recommendation results for the target object.
[0104] In one embodiment, multi-layer perceptual processing of content features based on a meta-mapping network is performed to determine the feature mapping relationship that matches the target object. This includes: acquiring interaction data for the target object; and performing multi-layer perceptual processing of the interaction data based on the meta-offset network parameters in the meta-mapping network to obtain an offset function that shifts the intermediate preference features toward the spatial distribution of the target preference features.
[0105] The interaction data for the target object can be a small amount of interaction data acquired during the cold start phase, or randomly generated interaction data following a random distribution. The meta-offset network parameters are the parameters obtained through training in the meta-mapping network, used to generate the offset parameters for offsetting preference features.
[0106] In a specific application, the meta-offset network is mainly responsible for offsetting and adjusting the stretched intermediate preference features. This involves filtering out noisy data from the intermediate preference features and adjusting the updated preference features D. c Closer to the target preference feature space distribution D t Taking short videos as the target audience as an example, learning preference information from interaction data is a key aspect of cold-start short video preference features. c Close to D t The most direct method. However, some noise is inevitable in the interaction data, such as accidental double-tap of short videos. Noise can have a significant negative impact on the learning of preference features. The Meta-Offset Network identifies noise in the interaction data of cold-start short videos and generates offset parameters with filtering effect for each cold-start short video.
[0107] In this embodiment, by acquiring interaction data for the target object, and based on the meta-offset network parameters in the meta-mapping network, multi-layer perceptual processing is performed on the interaction data to obtain an offset function that makes the intermediate preference features approach the spatial distribution of the target preference features. This allows for offset adjustment of the intermediate preference features based on the interaction data, retaining the useful information learned by the intermediate preference features from the interaction data, filtering out erroneous information generated by noise, and improving the fit between the updated preference features and the spatial distribution of the target preference features.
[0108] In one embodiment, the method further includes: aggregating interaction preference features in the interaction data to obtain aggregated interaction features; and performing multi-layer perceptual processing on the interaction data based on the meta-offset network parameters in the meta-mapping network to obtain an offset function that shifts the intermediate preference features toward the spatial distribution of the target preference features, including: performing multi-layer perceptual processing on the aggregated interaction features based on the meta-offset network parameters in the meta-mapping network to obtain an offset function that shifts the intermediate preference features toward the spatial distribution of the target preference features.
[0109] Interaction preference features refer to the preference characteristics of each receiving object that interacts with the target object. Aggregating interaction preference features involves combining these features, for example, by averaging the interaction preference features using a mean function. In a specific application, taking a short video as the target object, the aggregated interaction features could be the average preference characteristics of all users who liked short video i.
[0110] In a specific application, taking short video i as the target object, the global positive feedback set of short video i, that is, the set of preference features of each user who has liked short video i, is represented as U(i) = {u j |y(i, u) = 1}. In the few-sample phase of the cold start problem, |U(i)| > 0. The meta-offset network plays a role in the few-sample phase of the cold start problem. The representation of the meta-offset network is as follows:
[0111]
[0112] Where w shift These are the learnable parameters in the meta-offset network, and g is typically a two-layer multilayer perceptron model. It is an aggregate function. For offset function. Based on stretch function. and offset function From the initial preference feature D c To the target preference feature space distribution D t The mapping can be represented as:
[0113]
[0114] in This indicates element-wise addition.
[0115] In this embodiment, by acquiring interaction data for the target object and aggregating the interaction preference features in the interaction data, the aggregated interaction features used for multi-layer perception processing can characterize the average level of the interaction data. This allows the obtained offset parameters to generate an offset function that can more effectively filter out erroneous information caused by noise, thereby improving the fit between the updated preference features and the spatial distribution of the target preference features.
[0116] In one embodiment, the parameter training process of the meta-mapping network includes: obtaining recommendation feedback labels from the interaction data of the recommended object and determining the recommendation result for the recommended object; calculating the first loss data of the recommendation result and the recommendation feedback labels; and iteratively training the parameters of the initial meta-mapping network based on the first loss data using a gradient descent optimization algorithm until the training stopping condition is met, thereby obtaining the meta-mapping network.
[0117] The recommendation feedback label characterizes whether the recipient interacts with the recommended object. This label can be binary, including positive and negative feedback labels. In a specific application, the label uses 0 and 1, where 0 represents negative feedback (e.g., a user swiping past a recommended short video) and 1 represents positive feedback (e.g., a user liking a recommended short video). The recommendation result based on interaction data refers to the result obtained through recommendation analysis processing based on the initial preference features of the recommended object and the interaction data. The initial meta-mapping network refers to the meta-mapping network in its initial state. Iterative training updates the parameters of the initial meta-mapping network to obtain the final meta-mapping network.
[0118] Specifically, the prediction of the recommendation result for the recommended object can be expressed as:
[0119]
[0120] Where θ represents the parameters of the recommendation model.
[0121] Log loss data is commonly used as the loss function for binary classification tasks, i.e., the optimization objective, and is calculated as follows: The loss data between y and y is as follows:
[0122]
[0123] Where φ represents the embedding layer parameters, i.e., the embedding of the receiving object and the target object. The recommendation result is then calculated. After obtaining the loss data between the recommendation feedback label y, the learnable parameters θ and φ in the model can be optimized using gradient descent-based optimization methods, allowing the recommendation model to learn the relationship between the receiving object and the target object.
[0124] In this embodiment, recommendation feedback labels are obtained from the cold start simulation data of the recommended object, and the recommendation result for the recommended object is determined. The loss data between the recommendation result and the recommendation feedback labels is calculated. Then, the parameters of the initial meta-mapping network are iteratively trained based on the loss data through the gradient descent optimization algorithm until the training stopping condition is met. This enables the obtained meta-mapping network to generate more accurate feature mapping relationships for the target object during application, thereby improving the updated preference features of the mapping transformation to be closer to the target preference feature space distribution and obtaining accurate recommendation results.
[0125] In one embodiment, the method further includes: selecting sample objects from the sample objects whose interaction data meets the interaction conditions and whose recommendation effect reaches the expected effect; constructing triples based on the recommended object data, the receiving object data, and the recommendation feedback tags in the interaction data of the sample objects; and using the triples as interaction data for parameter training of the initial meta-mapping network.
[0126] Specifically, the interactive data used for parameter training of the initial meta-mapping network includes three entities (i, j, y): short videos, users, and tags, where i and j are the indices of the short video and user, respectively, and y is the tag. Both users and short videos have associated features. User features include user ID, gender, age, etc.; short video features include short video ID, title, category, description, etc. The embedding of the ID feature is called the preference feature, and the embedding of other features is called the content feature. Taking short videos as an example, the embedding of the short video ID obtained through the embedding layer is called the preference feature, while the embeddings of the short video title, category, and description obtained through the embedding layer are collectively called the content features; the same applies to users.
[0127] In one embodiment, the method further includes: converting the initial preference features of the recommended object into updated preference features that conform to the spatial distribution of the target preference features, based on the initial meta-mapping network. Further, calculating the loss data for the recommendation results and recommendation feedback labels includes: calculating first loss data for the recommendation results and recommendation feedback labels corresponding to the initial preference features, and second loss data for the recommendation results and recommendation feedback labels corresponding to the updated preference features; wherein the first loss data is used to update the parameters of the initial meta-mapping network; and the second loss data is used to update the initial preference features.
[0128] Specifically, the initial meta-mapping network includes an initial meta-stretching network and a meta-offset network. The training process for the initial meta-stretching network and meta-offset network is as follows:
[0129] Obtain the pre-trained recommendation model f θ,φ (·), randomly initialize the learnable parameters w of the meta-stretched network h. scale The learnable parameters w of the meta-offset network g shift ;
[0130] Repeat the following process until the recommendation model converges:
[0131] (1) From the interaction data about short video i Extract a small amount of interactive data.
[0132] (2) Initial preference features v based on short video i i Determine the first recommendation result, and calculate the first loss data based on the first recommendation result and the recommendation feedback tags in the interaction data.
[0133] (3) Through calculation formula Obtain the updated preference features of short video i in:
[0134]
[0135] w scale These are the parameters in the meta-stretched network, w shift These are the parameters in the meta-offset network, where h and g represent the multilayer perceptron model. For stretching functions, The offset function is represented by ⊙, which indicates element-wise multiplication. This indicates element-wise addition.
[0136] (4) Update preference features based on short video i Determine the second recommendation result, and calculate the second loss data based on the second recommendation result and the recommendation feedback tags in the interaction data.
[0137] (5) Optimize using gradient descent algorithm, based on the first loss data Update parameter w scale w shift Based on the second loss data For preference feature v i Update.
[0138] (6) The loop ends when the recommended model converges.
[0139] Specifically, the preference features of short video i, v i and content feature set Where n represents the total number of content feature categories for short videos, and u represents the user's preference features. j and content feature set Where m represents the total number of categories of user content features, u, v, That is, the vector dimension of each feature is k-dimensional. After specifying short videos and users as combinations of features, the triplet sample can be represented as ([v i X i u j S j ],y). For the prediction of interaction behavior, the recommendation result can be expressed as:
[0140]
[0141] Where θ represents the parameters of the recommendation model. Log loss data is often used as the loss data function for binary classification tasks, i.e., the optimization objective, and is calculated... The loss data between y and y is as follows:
[0142]
[0143] Where φ represents the embedding layer parameters, i.e., the embeddings of all users and short videos. The prediction is calculated. After obtaining the loss data between the short video and the label y, the learnable parameters θ and φ in the model can be optimized using gradient descent-based optimization methods, allowing the recommendation model to learn the relationship between the short video and the user.
[0144] Based on v i calculate The process is as follows:
[0145]
[0146] Similarly, it can be based on calculate
[0147] because Only with w scale w shift Related, Only with v i Related, therefore w scale w shift and v i The learning processes do not affect each other, so the introduction of the meta-mapping network will not affect the original recommendation model's recommendation performance for existing short videos.
[0148] In this embodiment, by calculating the first loss data of the recommendation result and recommendation feedback label corresponding to the initial preference feature and the second loss data of the recommendation result and recommendation feedback label corresponding to the updated preference feature respectively, the parameters of the initial meta-mapping network and the initial preference feature are updated. This ensures that the learning of the parameters of the initial meta-mapping network and the learning of the initial preference feature do not affect each other, so that the introduction of the meta-mapping network will not affect the recommendation effect of the original recommendation model on the existing recommendation objects, while achieving accurate recommendation of cold-start target objects.
[0149] This application also provides an application scenario for short video recommendation, in which the cold start object recommendation method described above is applied. Specifically, the cold start object recommendation method is applied in this application scenario as follows:
[0150] Recently, short videos have gradually become the type of internet content with the longest user dwell time and the highest data consumption. Major short video platforms rely on massive amounts of user and short video data, leveraging powerful recommendation system technologies to recommend popular, diverse, and personalized content to users, increasing user dwell time and strengthening user stickiness. However, whether it's traditional recommendation technologies, such as collaborative filtering (including SVD++ and matrix factorization), or deep learning-based neural recommendation systems, such as click-through rate prediction models like deep interest networks and deep interest evolution networks, all require a large amount of user-short video interaction data to allow the model to learn user preferences for short videos from this data.
[0151] For short videos lacking interactive data—that is, newly uploaded short videos by creators—the recommendation system struggles to identify the user preferences associated with the video due to the absence of other users' interactions such as watching, liking, and sharing. This leads to inaccurate target audience recommendations and impacts conversion rates; this is the cold start problem for short videos. Based on current daily data volumes of billions, records related to newly released short videos typically account for only 1-5%, reaching tens of millions. To address this issue, a cold start short video recommendation method based on meta-network mapping is proposed for short video platforms with cold start recommendation requirements. It is integrated into various existing deep learning-based recommendation models, overcoming the shortcomings of existing models in the zero-sample and few-sample stages of the cold start problem.
[0152] The cold-start short video recommendation method specifically includes initialization processing of preference features for short videos and two types of meta-networks: meta-stretching networks and meta-offset networks, wherein:
[0153] The initialization process for the preference features of short videos includes: for cold-start short videos in the zero-sample stage, a global average pooling operation is performed on the preference features of all existing short videos with sufficient interaction data and sufficient learning (hereinafter referred to as general short videos) to obtain generalized preference features as the initial preference features of cold-start short videos.
[0154] The meta-stretching network takes the content features of the cold-start short video as input and generates a stretching function specific to the current cold-start short video. It performs a stretching transformation on the preference features of the cold-start short video in the direction that adapts to the current recommendation system, i.e., the preference feature space of the ideal general short video, thus accelerating the process of adapting the cold-start short video to the recommendation system. The meta-offset network analyzes the preference features of all interactive users of the cold-start short video and generates an offset function that can guide the filtering of noise in the interactive data.
[0155] Specifically, recommendation systems trained on large amounts of general short video interaction data allocate a specific feature space to the preference features of general short videos, which means that the preference features of general short videos satisfy the target preference feature space distribution D. t Before sufficient interaction data is recorded, the preference feature distribution D of cold-start short videos... c With D t There is a gap; after cold-start short videos learn preference information from sufficient interaction data, their preference feature distribution will gradually shift towards D. t Convergence, the preference feature initialization process provides distance D for cold-start short videos t Minimize the initial preference features as much as possible to reduce their proximity to D t The difficulty. Furthermore, the two meta-networks provide D for cold-starting short videos. c To Dt The mapping has a significant effect on both the zero-sample and few-sample phases of the cold start problem.
[0156] The above technical solution is implemented based on the following parts: cold start recommendation scenario definition, preference feature initialization, training of meta-stretch network and meta-offset network, and model application.
[0157] Part 1: Cold Start Recommendation Scenarios Definition
[0158] Specifically, the recommendation model focuses on binary classification (0-1) recommendation tasks, predicting the probability (0-1) of a user liking a short video; 0 represents negative feedback, meaning the user directly scrolls past the short video; 1 represents positive feedback, meaning the user likes the short video. Each sample is a triple, including three entities: short video, user, and tag, i.e., (i, j, y), where i and j are the indices of the short video and user, respectively, and y is the tag. Both users and short videos have associated features. Features corresponding to user information include user ID, gender, age, etc.; features corresponding to short video information include short video ID, title, category, description, etc. The recommendation system uses embedding technology, a technique that transforms the original features of users and short videos into continuous vectors; these continuous vectors are called embeddings. After the original features pass through the embedding layer in the recommendation system, the embeddings of various features of users and short videos are obtained. Typically, the embedding of the ID feature is called the preference feature, and the embedding of other features is called the content feature. Taking short videos as an example, the embedding obtained from the short video's ID through the embedding layer is called preference feature, while the embeddings obtained from the short video's title, category, and description through the embedding layer are collectively called content features; the same applies to users. Since the ID is unique for each short video, learning ID features relies solely on the interaction data of the corresponding short video—this is the root cause of the cold start problem in short videos. Content features, on the other hand, typically appear repeatedly in multiple short videos, with ample interaction data, allowing almost all content features to be fully learned.
[0159] After the user and short video features are processed through the embedding layer, the preference features v of short video i can be obtained. i and content feature set Where n represents the total number of content feature categories for short videos. Similarly, the preference features u of user u can be obtained. j and content feature set Where m represents the total number of categories of user content features, u, v, That is, the vector dimension of each feature is k-dimensional, and the size of k is set according to the actual business effect. After specifying short videos and users as features, the original triplet sample can be represented as ([v i X i u j S jThe recommendation results for interactive behaviors can be represented as ], y).
[0160]
[0161] Where θ represents the parameters of a typical recommendation model. Log loss data is often used as the loss function for binary classification tasks, i.e., the optimization objective, and is calculated... The loss data between y and y is as follows:
[0162]
[0163] Where φ represents the embedding layer parameters, i.e., the embeddings of all users and short videos. The prediction is calculated. After obtaining the loss data between the short video and the label y, the learnable parameters θ and φ in the model can be optimized using gradient descent-based optimization methods, allowing the recommendation model to learn the relationship between the short video and the user.
[0164] Part Two: Preference Feature Initialization
[0165] Typically, when a short video is introduced into a recommendation system during a cold start, the system randomly initializes a preference feature for it. However, this randomly initialized preference feature contains almost no information relevant to the cold-start video, making it difficult to achieve satisfactory recommendation results. By performing global average pooling on the preference features of all existing general short videos, a globally average preference feature is obtained. As a generalized preference feature, when a short video enters the recommendation system during the zero-sample stage, the generalized preference feature is used as the initial preference feature for the short video. This preference feature initialization method provides the cold-start short video with a distribution D of distance from the general preference feature. t Minimizing the initial preference features significantly enhances the performance of the two meta-networks, making it far superior to random initialization methods.
[0166] Part 3: Training of Meta-Stretch Network and Meta-Offset Network.
[0167] After obtaining the generalized initial preference features, the meta-stretching network and the meta-offset network will generate stretching and offset functions specific to each cold start video. These two functions together constitute the distribution of the cold start preference features D. c Distribution D of general preference characteristics t The mapping improves the recommendation effect of short videos during cold starts.
[0168] Meta-stretch network responsible for D c To D tThe stretching operation is performed by the meta-stretching network, which generates a stretching function for each cold-start video. Typically, the content features of short videos often contain some preference information; short videos with similar preference features tend to have relatively similar content features. Therefore, by using content features as input to the meta-stretching network, the network can extract preference-related information from the content features. The resulting stretching function stretches the preference features of the cold-start short video towards D... t The meta-stretching network stretches the content features in a specific direction, incorporating preference-related information into the preference features. The expression for the meta-stretching network is as follows:
[0169]
[0170] Where w scale These are the learnable parameters in the meta-stretched network, where h is typically a two-layer multilayer perceptron model. This is the stretching function obtained from D. c To D t The stretching operation of the mapping can be expressed as ⊙ indicates element-wise multiplication.
[0171] The meta-offset network is mainly responsible for translating and adjusting the stretched preference features. This serves two purposes: filtering noise from the preference features and adjusting the D... c Closer to D t Typically, learning preference information from interaction data is a key approach for cold-start short videos, focusing on preference features derived from D. c Close to D t The most direct method. However, some noise is inevitable in the interaction data, such as accidental double-tap of short videos. Especially in the few-sample stage, noise can have a significant negative impact on the learning of preference features. The meta-offset network identifies noise in the interaction data of cold-start short videos and generates an offset function with filtering effect for each cold-start short video. The offset function shifts and adjusts the preference features of the cold-start short video, retaining the useful information learned from the interaction data and filtering out erroneous information generated by noise.
[0172] To represent the meta-offset network, a global positive feedback set for short video i is introduced, which is the set of preference features of all users who have liked short video i, denoted as U(i) = {u j |y(i, u) = 1}. In the zero-sample stage of the cold start problem, |U(i)| = 0; in the few-sample stage, |U(i)| > 0. Clearly, the meta-offset network plays a role in the few-sample stage of the cold start problem. The meta-offset network is represented as follows:
[0173]
[0174] Where wshift These are the learnable parameters in the meta-offset network, and g is typically a two-layer multilayer perceptron model. This is an aggregation function (e.g., a simple averaging function, i.e., calculating the average preference features of all users who have liked short video i). This is the resulting offset function. Finally, we have... and By D c To D t The mapping can be represented as:
[0175]
[0176] in This indicates element-wise addition.
[0177] Given a trained recommendation model f θ,φ (·), This recommendation model can make good recommendations for general short videos, but for cold-start short videos, the recommendation model performs poorly because the preference features have not been sufficiently trained. The training process of the models in modules three and four is explained below, as shown in the following algorithm:
[0178] Obtain the pre-trained recommendation model f θ,φ (·), randomly initialize the learnable parameters w of the meta-stretched network h. scale The learnable parameters w of the meta-offset network g shift ;
[0179] Repeat the following process until the recommendation model converges:
[0180] (1) From the interaction data about short video i Extract a small amount of interactive data.
[0181] (2) Initial preference features v based on short video i i Determine the first recommendation result, and calculate the first loss data based on the first recommendation result and the recommendation feedback tags in the interaction data.
[0182] (3) Through calculation formula Obtain updated preference features
[0183] (4) Update preference features based on short video i Determine the second recommendation result, and calculate the second loss data based on the second recommendation result and the recommendation feedback tags in the interaction data.
[0184] (5) Optimize using gradient descent algorithm, based on the first loss data Update parameter w scalew shift Based on the second loss data For preference feature v i Update.
[0185] (6) The loop ends when the recommendation model converges.
[0186] Among them, based on v i calculate The process is as follows:
[0187]
[0188] Similarly, it can be based on calculate The above training process uses a single short video i as an example; in actual training, it will be performed in batches. Because... Only with w scale w shift Related, Only with v i Related, therefore w scale w shift and v i The learning processes do not interfere with each other, so the introduction of the meta-network will not affect the original recommendation model's recommendation performance for general short videos.
[0189] Part Four: Model Application.
[0190] After the meta-network training is completed, when cold-start short videos enter the recommendation system, their preference features will be transformed into v. warm Make predictions until v learns preference information from enough interaction data, then use v directly for prediction.
[0191] It should be understood that although the steps in the flowcharts of the above embodiments are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the above embodiments may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.
[0192] Based on the same inventive concept, this application also provides a cold start object recommendation apparatus for implementing the cold start object recommendation method described above. The solution provided by this apparatus is similar to the implementation described in the above method; therefore, the specific limitations in one or more cold start object recommendation apparatus embodiments provided below can be found in the limitations of the cold start object recommendation method described above, and will not be repeated here.
[0193] In one embodiment, such as Figure 8 As shown, a cold start object recommendation device 800 is provided, including a feature acquisition module 802, a feature processing module 804, a mapping transformation module 806, and a recommendation analysis module 808, wherein:
[0194] Feature acquisition module 802 is used to acquire the content features and initial preference features of the target object in the cold start state;
[0195] The feature processing module 804 is used to perform multi-layer perceptual processing on the content features based on the meta-mapping network to determine the feature mapping relationship that matches the target object; the meta-mapping network is trained based on the interaction data of the recommended object in a simulated cold start state;
[0196] The mapping transformation module 806 is used to perform mapping transformation on the initial preference features according to the feature mapping relationship to obtain updated preference features that approximate the target preference feature space distribution; the target preference feature space distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended object;
[0197] The recommendation analysis module 808 is used to perform recommendation analysis on the target object based on the updated preference features, and obtain recommendation results for the target object.
[0198] In one embodiment, the feature acquisition module includes a content feature acquisition module and an initial preference feature acquisition module;
[0199] The content feature acquisition module is used to acquire content description data of the target object, perform vector transformation on the content description data, and obtain the content features of the target object.
[0200] The initial preference feature acquisition module is used to acquire the preference features corresponding to each of the recommended objects that satisfy the target preference feature spatial distribution; perform global average pooling on each preference feature, and use the obtained global average preference feature as the initial preference feature of the target object of the cold start.
[0201] In one embodiment, the target object includes a target video; the content feature acquisition module is further configured to acquire supplementary descriptive information for the target video and extract key video frames from the target video; perform image recognition processing on the key video frames to obtain video content information in the key video frames; and determine the supplementary descriptive information and the video content information as the content description data of the target object.
[0202] In one embodiment, the feature mapping relationship includes a stretching function; the mapping transformation module includes a stretching module, used to stretch the feature distribution of the initial preference feature according to the stretching function, integrate the content feature into the stretched space of the initial preference feature, and obtain an updated preference feature that approximates the target preference feature space distribution.
[0203] In one embodiment, the feature processing module includes a stretching function determination module, which performs multilayer perceptual processing on the content features based on the meta-stretching network parameters in the meta-mapping network to obtain a stretching function that stretches the initial preference features to approximate the spatial distribution of the target preference features.
[0204] In one embodiment, the feature mapping relationship further includes an offset function; the stretching module is used to stretch the feature distribution of the initial preference feature based on the stretching function, and to integrate the content feature into the stretching space of the initial preference feature to obtain the intermediate preference feature;
[0205] The mapping transformation module further includes an offset module, which is used to offset and adjust the feature distribution of the intermediate preference features based on the offset function, so as to remove noise data in the intermediate preference features and obtain updated preference features.
[0206] In one embodiment, the feature processing module includes an offset function determination module, used to acquire interaction data for the target object; and to perform multi-layer perceptual processing on the interaction data based on the meta-offset network parameters in the meta-mapping network to obtain an offset function that makes the intermediate preference features approach the spatial distribution of the target preference features.
[0207] In one embodiment, the cold start object recommendation device further includes a feature aggregation module, which is used to aggregate the interaction preference features in the interaction data to obtain aggregated interaction features;
[0208] The offset function determination module is further configured to perform multi-layer perceptual processing on the aggregated interaction features based on the meta-offset network parameters in the meta-mapping network, to obtain an offset function that makes the intermediate preference features approach the spatial distribution of the target preference features.
[0209] In one embodiment, the cold-start object recommendation device further includes a meta-mapping network training module, used to obtain recommendation feedback labels from the cold-start simulation data of the recommended objects and determine the recommendation results for the recommended objects; calculate the loss data between the recommendation results and the recommendation feedback labels; and iteratively train the parameters of the initial meta-mapping network based on the loss data using a gradient descent optimization algorithm until the training stopping condition is met, thereby obtaining the meta-mapping network.
[0210] In one embodiment, the meta-mapping network training module is further configured to, based on the initial meta-mapping network, convert the initial preference features of the recommended object into updated preference features that conform to the target preference feature spatial distribution; calculate, respectively, a first loss data between the recommendation result corresponding to the initial preference features and the recommendation feedback label, and a second loss data between the recommendation result corresponding to the updated preference features and the recommendation feedback label; the first loss data is used to update the parameters of the initial meta-mapping network; the second loss data is used to update the initial preference features.
[0211] In one embodiment, the cold start object recommendation device further includes a training data acquisition module, used to select sample objects from the sample objects whose interaction data meets the interaction conditions and whose recommendation effect reaches the expected effect; construct triples based on the recommended object data, the receiving object data and the recommendation feedback tags in the interaction data of the sample objects; and use the triples as interaction data for parameter training of the initial meta-mapping network.
[0212] Each module in the aforementioned cold start object recommendation device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in the processor of a computer device in hardware form or independent of it, or stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each module.
[0213] The aforementioned cold-start object recommendation device acquires the content features and initial preference features of the target object in a cold-start state. It is based on a meta-mapping network trained on the interaction data of the recommended object in a simulated cold-start state. The device performs multi-layer perceptual processing on the content features to determine the feature mapping relationship that matches the target object. The initial preference features are then mapped and transformed according to the feature mapping relationship to obtain updated preference features that approximate the spatial distribution of the target preference features. Since the spatial distribution of the updated preference features approximates the spatial distribution of the preference features corresponding to the recommended object, it can be determined that the target object in the cold-start state can learn accurate preference features. Then, based on the updated preference features, the device performs recommendation analysis on the target object to obtain more accurate recommendation results.
[0214] In one embodiment, a computer device is provided, which may be a server, and its internal structure diagram may be as follows: Figure 9 As shown, this computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system, computer programs, and a database. The internal memory provides the environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The database stores interactive data. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When executed by the processor, the computer program implements a cold start object recommendation method.
[0215] Those skilled in the art will understand that Figure 9 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.
[0216] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.
[0217] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.
[0218] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.
[0219] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data shall comply with the relevant laws, regulations and standards of the relevant countries and regions.
[0220] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.
[0221] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.
[0222] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.
Claims
1. A cold start object recommendation method, characterized in that, The method includes: Obtain the content description data of the target object, perform vector transformation on the content description data, and obtain the content features of the target object; Obtain the preference features corresponding to each of the recommended objects that satisfy the target preference feature spatial distribution; the target preference feature spatial distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended objects; Global average pooling is performed on each of the aforementioned preference features, and the resulting global average preference features are used as the initial preference features of the target object in the cold start state. Based on the meta-stretching network parameters in the meta-mapping network, the content features are subjected to multi-layer perceptual processing to obtain a stretching function that stretches the initial preference features to approximate the spatial distribution of the target preference features. Obtain interaction data for the target object; Based on the meta-offset network parameters in the meta-mapping network, multi-layer perceptual processing is performed on the interaction data to obtain an offset function that shifts the intermediate preference features to the spatial distribution of the target preference features; the meta-mapping network is trained based on the interaction data of the recommendation object in a simulated cold start state. Based on the stretching function, the feature distribution of the initial preference feature is stretched, and the content feature is incorporated into the stretched space of the initial preference feature to obtain the intermediate preference feature; Based on the offset function, the feature distribution of the intermediate preference features is offset and adjusted to remove noisy data in the intermediate preference features, thereby obtaining updated preference features; Based on the updated preference features, a recommendation analysis is performed on the target object to obtain recommendation results for the target object.
2. The method according to claim 1, characterized in that, The target object includes the target video; The acquisition of the content description data of the target object includes: Obtain supplementary descriptive information for the target video and extract key video frames from the target video; Image recognition processing is performed on the key video frames to obtain the video content information in the key video frames; The supplementary description information and the video content information are determined as the content description data of the target object.
3. The method according to claim 1, characterized in that, The method further includes: The interaction preference features in the interaction data are aggregated to obtain aggregated interaction features; The step of performing multi-layer perceptual processing on the interactive data based on the meta-offset network parameters in the meta-mapping network to obtain an offset function that shifts the intermediate preference features toward the spatial distribution of the target preference features includes: Based on the meta-offset network parameters in the meta-mapping network, the aggregated interaction features are subjected to multi-layer perceptual processing to obtain an offset function that shifts the intermediate preference features toward the spatial distribution of the target preference features.
4. The method according to any one of claims 1 to 3, characterized in that, The parameter training process of the meta-mapping network includes: Obtain recommendation feedback tags from the cold start simulation data of the recommended object, and determine the recommendation result for the recommended object; Calculate the loss data between the recommendation result and the recommendation feedback label; The parameters of the initial meta-mapping network are iteratively trained using the gradient descent optimization algorithm based on the loss data until the training stopping condition is met, thus obtaining the meta-mapping network.
5. The method according to claim 4, characterized in that, The method further includes: Based on the initial meta-mapping network, the initial preference features of the recommended object are transformed into updated preference features that conform to the spatial distribution of the target preference features; The calculation of the loss data between the recommendation result and the recommendation feedback label includes: Calculate the first loss data of the recommendation result corresponding to the initial preference feature and the recommendation feedback label, and the second loss data of the recommendation result corresponding to the updated preference feature and the recommendation feedback label, respectively; The first loss data is used to update the parameters of the initial meta-mapping network; the second loss data is used to update the initial preference features.
6. The method according to claim 4, characterized in that, The method further includes: Select sample objects whose interaction data meets the interaction conditions and whose recommendation effect achieves the expected effect; Based on the recommended object data, the received object data, and the recommendation feedback tags in the interaction data of the sample objects, triplet sets are constructed. The triples are used as interactive data for parameter training of the initial meta-mapping network.
7. A cold start object recommendation device, characterized in that, The device includes: The feature acquisition module is used to acquire content description data of the target object, perform vector transformation on the content description data to obtain the content features of the target object; acquire the preference features corresponding to each of the recommended objects that satisfy the target preference feature spatial distribution; perform global average pooling on each of the preference features, and use the obtained global average preference features as the initial preference features of the target object in the cold start state; the target preference feature spatial distribution is used to characterize the spatial distribution of the preference features corresponding to the recommended objects; The feature processing module is used to perform multi-layer perceptual processing on the content features based on the meta-stretching network parameters in the meta-mapping network to obtain a stretching function that stretches the initial preference features to approximate the spatial distribution of the target preference features; acquire interaction data for the target object; and perform multi-layer perceptual processing on the interaction data based on the meta-offset network parameters in the meta-mapping network to obtain an offset function that offsets the intermediate preference features to approximate the spatial distribution of the target preference features; the meta-mapping network is trained based on the interaction data of the recommended object in a simulated cold start state; The mapping transformation module is used to stretch the feature distribution of the initial preference features based on the stretching function, integrate the content features into the stretched space of the initial preference features to obtain intermediate preference features; and to offset the feature distribution of the intermediate preference features based on the offset function to remove noise data in the intermediate preference features to obtain updated preference features. The recommendation analysis module is used to perform recommendation analysis on the target object based on the updated preference features, and obtain recommendation results for the target object.
8. The cold start object recommendation device according to claim 7, characterized in that, The target object includes a target video; the feature acquisition module is further configured to acquire supplementary descriptive information for the target video and extract key video frames from the target video; perform image recognition processing on the key video frames to obtain video content information in the key video frames; and determine the supplementary descriptive information and the video content information as the content description data of the target object.
9. The cold start object recommendation device according to claim 7, characterized in that, The device further includes a feature aggregation module, used to aggregate interaction preference features in the interaction data to obtain aggregated interaction features; The feature processing module is further configured to perform multi-layer perceptual processing on the aggregated interaction features based on the meta-offset network parameters in the meta-mapping network, to obtain an offset function that makes the intermediate preference features approach the spatial distribution of the target preference features.
10. The cold start object recommendation device according to any one of claims 7 to 9, characterized in that, The device further includes a meta-mapping network training module, used to obtain recommendation feedback tags from cold start simulation data of the recommended object and determine the recommendation result for the recommended object; Calculate the loss data between the recommendation result and the recommendation feedback label; The parameters of the initial meta-mapping network are iteratively trained using the gradient descent optimization algorithm based on the loss data until the training stopping condition is met, thus obtaining the meta-mapping network.
11. The cold start object recommendation device according to claim 10, characterized in that, The meta-mapping network training module is also used to convert the initial preference features of the recommended object into updated preference features that conform to the target preference feature space distribution based on the initial meta-mapping network. Calculate the first loss data of the recommendation result corresponding to the initial preference feature and the recommendation feedback label, and the second loss data of the recommendation result corresponding to the updated preference feature and the recommendation feedback label, respectively; the first loss data is used to update the parameters of the initial meta-mapping network; the second loss data is used to update the initial preference feature.
12. The cold start object recommendation device according to claim 10, characterized in that, The device also includes a training data acquisition module, used to select sample objects from the sample objects whose interaction data meets the interaction conditions and whose recommendation effect reaches the expected effect; Based on the recommended object data, received object data, and recommendation feedback tags in the interaction data of the sample objects, triples are constructed; the triples are used as interaction data for parameter training of the initial meta-mapping network.
13. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.
14. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.
15. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.
