Content recall method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention provides a content recall method and device, electronic equipment and a storage medium. The method comprises the steps of obtaining candidate content features corresponding to candidate content and a target feature sequence corresponding to a target content sequence; fusing the candidate content features and the target feature sequence to obtain a fused feature sequence; according to the fusion feature sequence, determining a correlation weight between each target content feature in the target feature sequence and the candidate content feature; determining a target browsing content feature according to a correlation weight between each target content feature in the target feature sequence and the candidate content feature and the target feature sequence; and performing recall processing on the candidate content according to the target browsing content feature, the object feature corresponding to the target object and the candidate content feature corresponding to the candidate content. Through the method, the accuracy of the recalled candidate content can be effectively improved.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to a content recall method, device, electronic device and storage medium. Background Art

[0002] At present, with the emergence of various applications, in order to meet the needs of different users, applications generally provide a large amount of information (such as videos, images, articles, etc.) for users to choose from. Since the application provides a lot of information, users need to spend a lot of time searching and browsing the information they are interested in. In order to save user time, some applications use recommendation systems to recommend content to users in a personalized way. Specifically, the recommendation system determines the information that the user may like based on the user's historical browsing content, and then recalls information similar to the historical browsing content from the information library. However, in the related art, when content is recalled based on historical browsing content, the accuracy of content recall is low. Summary of the invention

[0003] In view of this, the embodiments of the present application propose a content recall method, device, electronic device and storage medium, which can more accurately recall content.

[0004] In a first aspect, an embodiment of the present application provides a content recall method, the method comprising: obtaining candidate content features corresponding to candidate content and a target feature sequence corresponding to a target content sequence, the target content sequence including content attribute information corresponding to multiple historical browsing contents of a target object triggering a specified behavior; fusing the candidate content features and the target feature sequence to obtain a fused feature sequence; determining, based on the fused feature sequence, a correlation weight between each target content feature in the target feature sequence and the candidate content feature; determining a target browsing content feature based on the correlation weight between each target content feature in the target feature sequence and the candidate content feature and the target feature sequence; and recalling the candidate content based on the target browsing content feature, the object feature corresponding to the target object, and the candidate content feature corresponding to the candidate content.

[0005] In the second aspect, the embodiment of the present application provides a content recall device, which includes: a data acquisition module, a feature fusion module, a weight determination module, a feature determination module, and a recall processing module. The data acquisition module is used to obtain candidate content features corresponding to the candidate content and a target feature sequence corresponding to the target content sequence, wherein the target content sequence includes content attribute information corresponding to multiple historical browsing contents that trigger a specified behavior of the target object; the feature fusion module is used to fuse the candidate content features and the target feature sequence to obtain a fused feature sequence; the weight determination module is used to determine the correlation weight between each target content feature in the target feature sequence and the candidate content feature according to the fused feature sequence; the feature determination module is used to determine the target browsing content feature according to the correlation weight between each target content feature in the target feature sequence and the candidate content feature and the target feature sequence; the recall processing module is used to recall the candidate content according to the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content.

[0006] In one possible implementation, the recall processing module includes a fusion submodule, a probability prediction submodule and a recall processing submodule; the fusion submodule is used to fuse the target browsing content features, the object features corresponding to the target object and the candidate content features corresponding to the candidate content to obtain a target fusion feature; the probability prediction submodule is used to predict the probability of the target object performing the specified behavior on the candidate content based on the fusion feature; the recall submodule is used to recall the candidate content when the probability is greater than a probability threshold.

[0007] In one possible implementation, the fused feature sequence includes multiple fused features, each fused feature is obtained by fusing the candidate content with a historical browsing content feature; the weight determination module includes a compression processing submodule and a stretching processing submodule, the compression processing submodule is used to compress and reduce the dimension of each fused feature in the fused feature sequence to obtain a fused feature sequence after dimension reduction; the stretching processing submodule is used to stretch the fused feature sequence after dimension reduction to obtain the correlation weight between each target content feature in the target feature sequence and the candidate content feature.

[0008] In one possible implementation, the compression processing submodule includes a compression processing unit and a first activation unit; the compression processing unit is used to compress each fused feature in the fused feature sequence through a compression parameter matrix to obtain a first intermediate feature sequence; the first activation unit is used to activate the first intermediate feature sequence to obtain the reduced-dimensional fused feature sequence.

[0009] In one possible implementation, the stretching processing submodule includes a stretching processing unit and a second activation unit; the stretching processing unit is used to perform dimensionality increase processing on the reduced fusion feature sequence through a stretching parameter matrix to obtain a second intermediate feature sequence; the second activation unit is used to perform activation processing on the second intermediate feature sequence to obtain the correlation weight between each target content feature and the candidate content feature.

[0010] In one possible implementation, the feature determination module is further used to perform weighted summation of the target content features in the target feature sequence according to the correlation weights between each target content feature in the target feature sequence and the candidate content features to obtain a target browsing content feature vector; and perform pooling processing on the target browsing content feature vector to obtain a target browsing content feature.

[0011] In one possible implementation, the feature fusion module is further configured to perform a dot product process on the candidate content features and the target feature sequence to obtain the fused feature sequence.

[0012] In one possible implementation, the data acquisition module includes an information extraction submodule and a feature extraction submodule. The information extraction submodule is used to extract a target content sequence based on the information browsing record of the target object; the feature extraction submodule is used to extract features from the target content sequence to obtain a target feature sequence.

[0013] In one possible implementation, the data acquisition module is also used to obtain target portrait information of the target object and context information for the target object to acquire information, wherein the context information includes at least one of information acquisition time, information acquisition location, and mood of the target object; and feature extraction is performed on the target portrait information and context information respectively to obtain object features including portrait features corresponding to the target portrait information and context features corresponding to the context information.

[0014] In a third aspect, an embodiment of the present application provides an electronic device, comprising a processor and a memory; one or more programs are stored in the memory and configured to be executed by the processor to implement the above method.

[0015] In a fourth aspect, an embodiment of the present application provides a computer-readable storage medium, in which a program code is stored, wherein the above method is executed when the program code is executed by a processor.

[0016] In a fifth aspect, an embodiment of the present application provides a computer program product or a computer program, the computer program product or the computer program including computer instructions, the computer instructions being stored in a computer-readable storage medium. A processor of a computer device obtains the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the computer device performs the above method.

[0017] The embodiment of the present application provides a content recall method, device, electronic device and storage medium. A fused feature sequence is obtained by fusing candidate content features and a target feature sequence; and when determining the correlation weight between each target content feature and the candidate content feature in the target feature sequence based on the fused feature sequence, it is fully considered that the probability of the candidate content feature executing a specified behavior is related to the historical browsing content that executes the specified behavior, that is, the calculation of the correlation weight takes into account the correlation between the candidate content feature and the global context of the target feature sequence, so that it can be ensured that according to the correlation weight between each target content feature and the candidate content feature in the target feature sequence and the target feature sequence, the determined target browsing content feature can accurately express the features of the content that the target object is more interested in, or the features of the content that the target object pays attention to, and then according to the target browsing content feature, the object feature corresponding to the target object and the candidate content feature corresponding to the candidate content, the recalled content is the content that the target object is more interested in, or the content that the target object pays attention to, that is, the accuracy of the recalled content is effectively guaranteed. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0019] Figure 1 An application scenario diagram of a content recall method provided by an embodiment of the present application is shown;

[0020] Figure 2 A schematic diagram showing a flow chart of a content recall method proposed in an embodiment of the present application is shown;

[0021] Figure 3 A model structure view of a content recall model provided by an embodiment of the present application is shown;

[0022] Figure 4 Another schematic diagram of a process of a content recall method proposed in an embodiment of the present application is shown;

[0023] Figure 5A schematic diagram showing a display interface of an application proposed in an embodiment of the present application is shown;

[0024] Figure 6 A schematic diagram showing another display interface of an application proposed in an embodiment of the present application is shown;

[0025] Figure 7 A connection block diagram of a content recall device proposed in an embodiment of the present application is shown;

[0026] Figure 8 A structural block diagram of an electronic device for executing the method of an embodiment of the present application is shown. DETAILED DESCRIPTION

[0027] Example embodiments will now be described more fully with reference to the accompanying drawings. However, example embodiments can be implemented in a variety of forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided so that this application will be more comprehensive and complete and fully convey the concept of the example embodiments to those skilled in the art.

[0028] In addition, described feature, structure or characteristic can be combined in one or more embodiments in any suitable manner. In the following description, many specific details are provided to provide a full understanding of the embodiments of the present application. However, those skilled in the art will appreciate that the technical scheme of the present application can be put into practice without one or more of the specific details, or other methods, components, devices, steps, etc. can be adopted. In other cases, known methods, devices, realizations or operations are not shown or described in detail to avoid blurring the various aspects of the application.

[0029] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities may be implemented in software form, or in one or more hardware modules or integrated circuits, or in different networks and / or processor devices and / or microcontroller devices.

[0030] The flowcharts shown in the accompanying drawings are only exemplary and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps can be decomposed, and some operations / steps can be combined or partially combined, so the actual execution order may change according to actual conditions.

[0031] It should be noted that the "multiple" mentioned in this article refers to two or more. "And / or" describes the association relationship of the associated objects, indicating that there can be three relationships. For example, A and / or B can mean: A exists alone, A and B exist at the same time, and B exists alone. The character " / " generally indicates that the associated objects before and after are in an "or" relationship.

[0032] It should also be noted that in the embodiments of the present application, the acquisition of user data (such as one or more of job type, education level, age, gender, life stage (for example, whether married), work location, and preferred products, etc.) requires the user's permission or consent, and the collection, use, processing and storage of the above information must comply with the regulations of the region.

[0033] With the research and advancement of artificial intelligence technology, artificial intelligence technology has been studied and applied in many fields and played an increasingly important role.

[0034] Artificial Intelligence (AI) is the theory, method, 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 knowledge to obtain the best 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 intelligent machine that can respond in a similar way to human intelligence. Take the application of artificial intelligence in machine learning as an example:

[0035] Among them, machine learning (ML) is a multi-disciplinary interdisciplinary subject involving probability theory, statistics, approximation theory, convex analysis, algorithm complexity theory and other disciplines. It specializes in studying how computers 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 make computers intelligent. Its applications are spread across all areas of artificial intelligence. Take the application of machine learning in content recall as an example:

[0036] At present, when recalling content based on the historical browsing content in the application used by the user, the information sequence composed of multiple historical operation data of the user in the application is mapped to a vector representation through the embedding layer, and the vector representations corresponding to each historical operation data in the information sequence are directly summed up to obtain a final vector representation (Pooling), which is used as the user's interest representation. After that, the candidate content is recalled based on the user's interest representation, user-side features and the features of the candidate content.

[0037] The inventor has found through research that when the methods in the related art perform content recall, the extraction of user interests lacks consideration of the correlation between candidate content and historical operation data, which makes the results of recall processing based on user interests inaccurate.

[0038] Based on this, the present application proposes a content recall method, which includes: obtaining candidate content features corresponding to candidate content and target feature sequences corresponding to target content sequences; fusing candidate content features and target feature sequences to obtain fused feature sequences; determining the correlation weights between each target content feature and candidate content features in the target feature sequence according to the fused feature sequences; determining target browsing content features according to the correlation weights between each target content feature and candidate content features in the target feature sequence and the target feature sequence; recalling candidate content according to target browsing content features, object features corresponding to target objects, and candidate content features corresponding to candidate content. By adopting the above method, in the process of content recall processing, it is fully considered that the probability of candidate content features executing a specified behavior is related to the historical browsing content that executes the specified behavior, that is, the candidate content features have a correlation with the global context of the target feature sequence, so that when the target browsing content features are obtained according to the correlation weights between each target content feature and the candidate content features, the obtained target browsing content features can accurately express the interests and hobbies of the target object, thereby making the recall processing result obtained by recalling the candidate content according to the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content more accurate.

[0039] The following describes an exemplary application of performing content recall provided by an embodiment of the present invention. The content recall method provided by an embodiment of the present invention can be applied to: Figure 1 In the server in the application environment shown.

[0040] Figure 1 is a schematic diagram of an application scenario according to an embodiment of the present application, such as Figure 1 As shown, the application scenario includes a terminal device 10 and a server 20 that is communicatively connected to the terminal device 10 via a network.

[0041] The terminal device 10 can run a client for displaying data (eg, displaying articles, videos, etc.), and the terminal device 10 can also publish the article to be processed through the server 20.

[0042] The network may be a wide area network or a local area network, or a combination of the two. The terminal device 10 may be a smart phone, a smart TV, a tablet computer, a laptop computer, or a desktop computer.

[0043] Server 20 can be an independent physical server, or a server cluster or distributed system composed of multiple physical servers. It can also be a cloud server that provides basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communications, middleware services, domain name services, security services, CDN, as well as big data and artificial intelligence platforms.

[0044] If using Figure 1 When the terminal device 10 and the server 20 in the method jointly realize content recall, the terminal device can send an information acquisition request to the server in response to the user's information acquisition instruction. When the server receives the information acquisition request, it can obtain the candidate content features corresponding to the candidate content and the target feature sequence corresponding to the target content sequence, wherein the target content sequence includes the content attribute information corresponding to multiple historical browsing contents of the target object triggering the specified behavior; the candidate content features and the target feature sequence are fused to obtain a fused feature sequence; according to the fused feature sequence, the correlation weight between each target content feature in the target feature sequence and the candidate content feature is determined; according to the correlation weight between each target content feature in the target feature sequence and the candidate content feature and the target feature sequence, the target browsing content feature is determined; according to the target browsing content feature, the object feature corresponding to the target object and the candidate content feature corresponding to the candidate content, the candidate content is recalled.

[0045] After recalling the candidate content, the server 20 may also send the recalled candidate content to the terminal device 10 so that the terminal device 10 displays the recalled candidate content. The server 20 may also obtain and store the operation behavior triggered by the user on the recalled candidate content in the terminal device.

[0046] It should be understood that the above-mentioned content recall method can also be executed by a terminal device or a server alone, and the specific process of the terminal device or the server alone executing the content recall method will not be described in detail here.

[0047] The embodiments of the present application will be described in detail below with reference to the accompanying drawings.

[0048] Please read Figure 2 , Figure 2 The present application also provides a content recall method, which can be applied to an electronic device, and the electronic device can be the terminal device 10 or the server 20 mentioned above, and the method includes:

[0049] Step S110: Acquire candidate content features corresponding to the candidate content and a target feature sequence corresponding to the target content sequence.

[0050] The target content sequence includes the content attribute information corresponding to each of the multiple historical browsing contents that trigger the specified behavior of the target object. The content attribute information can be content ID, content publisher ID, content category, etc. The target object refers to the user of the client that triggers the recall behavior. The above-mentioned candidate content can be articles, images, videos, etc. The above-mentioned candidate content, that is, the content to be recalled, wherein the candidate content includes candidate videos, candidate images, and candidate articles, etc. Each candidate content is pre-saved in the content library, and the content library is updated regularly or in real time. Accordingly, the historical browsing content includes at least one of the following: article viewing records, image viewing records, video viewing records, etc. Users can generate historical browsing content through operations such as clicking, double-clicking, and long pressing.

[0051] The target content sequence includes content attribute information corresponding to multiple historical browsing contents that trigger the specified behavior of the target object within the target time period, wherein the target time period can be a time period of a preset time from the current moment, or other time periods. Content attribute information historical browsing contents, for example, include information category, information ID, information publisher ID, etc. The electronic device can store the attribute information of each content, and the electronic device can also update the attribute information of each content regularly or in real time.

[0052] The above-mentioned specified behavior can be at least one of a like behavior, a forwarding behavior, a collection behavior, an in-depth browsing behavior, etc. The above-mentioned in-depth browsing behavior can specifically be that the content browsing time reaches a first preset time, or that the browsing time ratio reaches a preset ratio. Among them, the browsing time ratio can be based on the ratio between the user's actual browsing time and the second preset time. Different content has different corresponding second preset times. Exemplarily, the second preset time can be the time to normally complete the browsing of the content, such as the time it takes for a video to finish playing.

[0053] Taking the content as video content as an example, the above-mentioned specified behavior may be a deep play behavior, and the deep play behavior may refer to a behavior in which the play time of the video played by the user reaches a first preset time length.

[0054] The method of obtaining the candidate content features corresponding to the candidate content may be to obtain the pre-stored candidate content features corresponding to the candidate content, or to obtain the candidate content and extract the features of the candidate content to obtain the candidate content features. The above-mentioned method of extracting features may be to perform feature mapping (embedding processing) on ​​the candidate content to obtain the candidate content features, or may be to use a feature extraction network in an RNN model, LSTM model, GRU model, Transformer model, etc. to extract features.

[0055] The method of obtaining the target feature sequence corresponding to the target content sequence is similar to the method of obtaining the candidate content features corresponding to the candidate content, and will not be described in detail here.

[0056] It should be noted that the above-mentioned target content sequence can be obtained from the information browsing record of the target object.

[0057] That is, in the above step S110, obtaining the target feature sequence corresponding to the target content sequence includes: extracting the target content sequence based on the information browsing record of the target object; and performing feature extraction on the target content sequence to obtain the target feature sequence.

[0058] The method of extracting the content sequence from the historical browsing content corresponding to the target object may specifically be to extract multiple historical browsing contents corresponding to the content that triggered the specified behavior within a specified time period, that is, the generation time corresponding to each historical browsing content should be within a specified time period, such as two weeks, one week, three days or one day from the current time. The content attribute information corresponding to the multiple historical browsing contents is combined or the corresponding content attribute information is arranged according to the generation time of the historical browsing contents to obtain the target content sequence.

[0059] The above method of extracting features from the target content sequence to obtain the target feature sequence can refer to the specific method of obtaining the candidate content features corresponding to the candidate content mentioned above, which will not be described in detail here.

[0060] In one possible implementation mode of the present application, the candidate content features are obtained by performing feature mapping on the candidate content, and the target feature sequence is obtained by performing feature mapping on the target content sequence.

[0061] Step S120: Fusing the candidate content features and the target feature sequence to obtain a fused feature sequence.

[0062] The candidate content features and the target feature sequence may be fused in one or more fusion methods including dot product, bitwise addition, or bitwise addition followed by multiplication.

[0063] In one possible implementation of the present application, the above-mentioned method of fusing the candidate content features and the target feature sequence may specifically be:

[0064] Perform dot product processing on the candidate content features and the target feature sequence to obtain a fused feature sequence.

[0065] Step S130: Determine the correlation weight between each target content feature and the candidate content feature in the target feature sequence according to the fused feature sequence.

[0066] The target content feature refers to a feature obtained by feature mapping the content attribute information corresponding to a historical browsing content included in the target content sequence. The correlation weight between each target content feature and the candidate content feature reflects the similarity or correlation between the target content feature and the candidate content feature.

[0067] The above-mentioned step S130 can be specifically: using a similarity calculation method to calculate the similarity between the fused features corresponding to the target content features in each fused feature sequence and the candidate content features, and based on the similarity between the fused features corresponding to each target content feature and the candidate content features, determine the correlation weight between each target content feature and the candidate content features.

[0068] The above step S130 can also be to compress the fused feature sequence using a pooling operation to obtain a statistical vector for characterizing the fused feature sequence; based on the statistical vector, a weight vector for the fused feature sequence is obtained through machine learning, and the weight vector includes the correlation weights between each target content feature and the candidate content features in the target feature sequence.

[0069] The above step S130 may also use a correlation prediction algorithm or model to predict the fused feature sequence to obtain the correlation weights between each target content feature and the candidate content features in the target feature sequence.

[0070] It should be understood that the above-mentioned method of obtaining the correlation weights between each target content feature and the candidate content feature is only illustrative, and there may be more acquisition methods, which are not described one by one in the embodiments of the present application.

[0071] Step S140: determining the target browsing content feature according to the correlation weight between each target content feature and the candidate content feature in the target feature sequence and the target feature sequence.

[0072] In one possible implementation, the above step S140 may specifically be: according to the correlation weights between each target content feature in the target feature sequence and the candidate content features, weighted summing the target content features in the target feature sequence to obtain a target browsing content feature vector; and pooling the target browsing content feature vector to obtain the target browsing content feature.

[0073] In another possible implementation, the above step S140 can also be: obtaining a designated content feature from a plurality of target content features included in the target feature sequence, wherein the weight corresponding to the designated content feature satisfies a preset condition; the preset condition means that after the correlation weights corresponding to the target content features are sorted in descending order, the sorting order of the correlation weights corresponding to the designated content features is within a preset range, and the preset condition can also be that the correlation weight corresponding to the designated content feature is greater than a preset weight threshold. The correlation weights corresponding to the designated content features are normalized to obtain the normalized weights corresponding to the designated content features, and the designated content features are weighted and summed according to the normalized weights corresponding to the designated content features to obtain a target browsing content feature vector; the target browsing content feature vector is pooled to obtain the target browsing content feature.

[0074] Step S150: recalling the candidate contents according to the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate contents.

[0075] In one possible implementation, the above step S150 may be: fusing the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content to obtain the target fusion features; based on the fusion features, predicting the probability that the target object performs a specified behavior on the candidate content; if the probability is greater than a probability threshold, recalling the candidate content.

[0076] The above-mentioned fusion method can be at least one of splicing, bitwise multiplication or addition, etc. In one possible implementation of the present application, the above-mentioned fusion of the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content is specifically to splice the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content.

[0077] It should be understood that if there are multiple candidate contents, the above steps S110-S150 can be performed for each candidate content to predict the probability of the target object performing the specified behavior on each candidate content. Then, the target candidate information can be determined from the multiple candidate information based on the probability of performing the specified behavior on each candidate information, and the target candidate information can be recalled.

[0078] It should be understood that, as an implementable method, the above steps S110-S150 can be specifically performed by a pre-trained machine learning model (such as a content recall model), that is, the above-mentioned method of determining the target browsing content feature sequence and predicting the probability of the target object performing a specified behavior on the candidate content based on the fused features can be to obtain the target browsing content features based on the candidate content features and the target feature sequence using a feature extraction network in a content recall model trained with a machine learning model such as a deep neural network, a convolutional neural network, a Transformer network or a multi-layer perceptron, and perform deep feature extraction on the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content, and perform classification prediction on the extracted features based on the classification layer in the content recall model to obtain the probability of the target object performing the specified behavior on the candidate content, and recall the candidate content when the probability is greater than the probability threshold.

[0079] In this implementation mode, before executing the above step S110, a machine learning model (content recall model) should also be trained. The specific training process may be: obtaining sample data, the sample data including information corresponding to the sample object, a sample content sequence, a target sample content and a label corresponding to each target sample content, wherein the sample content sequence includes sample content attribute information corresponding to each of multiple historical sample browsing contents that trigger a specified behavior by the sample object; the label corresponding to the target sample content is used to indicate whether the sample object triggers a specified behavior for the target sample content; performing feature extraction on the target sample content to obtain a target sample content feature corresponding to the target sample content; and performing feature extraction on the sample content sequence to obtain a sample content feature sequence corresponding to the sample content sequence. The target sample content features and the sample content feature sequence are fused to obtain a fused sample feature sequence; based on the fused sample feature sequence, the correlation weight between each sample content feature in the sample content feature sequence and the target sample content feature is determined; based on the correlation weight between each sample content feature in the sample content feature sequence and the target sample content feature and the sample content feature sequence, the sample browsing content features are determined; based on the sample browsing content features, the sample object features corresponding to the sample object, and the target sample information feature vector, classification prediction is performed to obtain the prediction result of the sample object triggering a specified behavior on the target sample content; based on the prediction result and the label, the model loss is calculated, and the model parameters of the machine learning model (content recall model) are adjusted according to the model loss until the training end condition is met.

[0080] Among them, the label corresponding to the target sample content is y (the value is 0 or 1, representing negative samples and positive samples. The positive sample represents that the sample object triggers the specified behavior for the target sample content, and the negative sample represents that the sample object does not trigger the specified behavior for the target sample content). The loss function of its training can be p represents the probability that the sample object triggers a specified behavior for the target sample content (ie, the predicted result).

[0081] It should be understood that when the number of training times reaches a preset number of times or the model loss is less than a preset loss threshold, it is considered that the training of the content recall model has reached the training end conditions, and the content recall model that has reached the training end conditions can be used to execute the steps in the content recall method in the embodiments of the present application.

[0082] It should be noted that the data processing method at each stage of the above training process should be similar to the data processing method at each stage of the prediction process. Therefore, for the specific description of each step in the model training process, please refer to the specific description of each step in the content recall method above, and will not be repeated here one by one.

[0083] For example, Figure 3 As shown, a schematic diagram of the model structure of a machine learning model (content recall model) used in the content recall method is given. The machine learning model includes an embedding layer, a relevance prediction layer embedded with a relevance weight prediction model, a fully connected layer, a feature extraction network, and a classification layer.

[0084] The embedding layer is used to perform feature mapping on the information corresponding to the target object, the candidate content, and the target content sequence, respectively, to obtain the object features corresponding to the target object, the candidate content features corresponding to the candidate content, and the target feature sequence corresponding to the target content sequence.

[0085] The relevance prediction layer is used to perform dot product between the candidate content features and the target feature sequence to obtain a fused feature sequence, and use its embedded relevance weight prediction model to perform relevance prediction on the fused feature sequence to obtain the relevance weights between each target content feature in the target feature sequence and the candidate content features, and based on the relevance weights between each target content feature in the target feature sequence and the candidate content features, perform weighted summation of the target content features in the target feature sequence and perform pooling processing to obtain the target browsing content features.

[0086] The fully connected layer fully connects the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content to obtain the target fusion features.

[0087] The feature extraction network performs deep feature extraction on the target fusion features and inputs the extracted features into the classification layer.

[0088] The classification layer predicts the extracted deep features to obtain the probability that the target object performs a specified behavior on the candidate content, and recalls the candidate content when the predicted probability is greater than the probability threshold.

[0089] It should be understood that the model structure of the machine learning model used in the above-mentioned content recall method is only illustrative, and there may be more or fewer hierarchical structures as long as the content recall method of the present application can be implemented.

[0090] A content recall method provided in an embodiment of the present application obtains a fused feature sequence by fusing candidate content features and a target feature sequence; and when determining the correlation weights between each target content feature in the target feature sequence and the candidate content features based on the fused feature sequence, it is fully considered that the probability of the candidate content feature executing a specified behavior is related to the historical browsing content that executes the specified behavior, that is, the calculation of the correlation weight takes into account the correlation between the candidate content feature and the global context of the target feature sequence. Therefore, it can be ensured that the target browsing content features determined based on the correlation weights between each target content feature and the candidate content features in the target feature sequence and the target feature sequence can accurately express the features of the content in which the target object is more interested, or the features of the content that the target object pays attention to, and then the content recalled based on the target browsing content features, the object features corresponding to the target object and the candidate content features corresponding to the candidate content is the content in which the target object is more interested, or the content that the target object pays attention to, that is, the accuracy of the recalled content is effectively guaranteed.

[0091] See also Figure 4 Another embodiment of the present application provides a content recall method, the method comprising:

[0092] Step S210: Acquire candidate content features corresponding to the candidate content and a target feature sequence corresponding to the target content sequence.

[0093] The target content sequence includes content attribute information corresponding to each of the multiple historical browsing contents that trigger the specified behavior of the target object.

[0094] Step S220: Fusing the candidate content features and the target feature sequence to obtain a fused feature sequence.

[0095] The fused feature sequence includes multiple fused features, each of which is obtained by fusion of features mapped from candidate content and a content attribute corresponding to a historical browsing content. It should be understood that the target feature sequence includes features mapped from content attribute information corresponding to multiple historical browsing contents.

[0096] Step S230: compressing and reducing the dimension of each fused feature in the fused feature sequence to obtain a fused feature sequence after dimension reduction.

[0097] The above-mentioned method of reducing the dimensionality of each fused feature in the fused feature sequence can be a linear dimensionality reduction method (such as principal component analysis dimensionality reduction, linear discriminant analysis dimensionality reduction or univariate feature selection dimensionality reduction and other linear dimensionality reduction methods), or a nonlinear dimensionality reduction method (such as dimensionality reduction based on kernel functions or eigenvalues).

[0098] In an implementation manner of the present application, a nonlinear dimensionality reduction method may be adopted. Specifically, the above step S230 may include:

[0099] Step S232: compressing each fused feature in the fused feature sequence through a compression parameter matrix to obtain a first intermediate feature sequence.

[0100] The compression parameter matrix may be a parameter matrix obtained through pre-training, that is, it may be a parameter matrix involved in compressing the fused feature matrix in the aforementioned post-training content recall model stage.

[0101] Step S234: Activate the first intermediate feature sequence to obtain a fusion feature sequence after dimension reduction.

[0102] When the activation process is performed on the first intermediate feature sequence, the activation function may be used for processing. That is, the activation function may be used to perform the activation process on the first intermediate feature sequence to obtain a fusion feature sequence after dimension reduction.

[0103] Step S240: stretching the fusion feature sequence after dimension reduction to obtain the correlation weight between each target content feature and the candidate content feature in the target feature sequence.

[0104] The stretching method of the fusion feature sequence after dimension reduction can be: a linear stretching method (such as a linear interpolation method) or a nonlinear stretching method (such as a dimensionality increase method based on a kernel function or eigenvalue).

[0105] In an implementation manner of the present application, a nonlinear dimensionality increase method may be adopted. Specifically, the above step S240 may include:

[0106] Step S242: Performing dimension-up processing on the fused feature sequence after dimension reduction by stretching the parameter matrix to obtain a second intermediate feature sequence.

[0107] The up-scaled parameter matrix may be a parameter matrix obtained by pre-training, that is, it may be a parameter matrix involved in up-scaled the fusion feature matrix after dimension reduction in the aforementioned post-training content recall model stage.

[0108] Step S244: Activate the second intermediate feature sequence to obtain the correlation weight between each target content feature and the candidate content feature.

[0109] When the second intermediate feature sequence is activated, the activation function may be used for processing. That is, the second intermediate feature sequence may be activated by using the activation function to obtain the correlation weight between each target content feature and the candidate content feature.

[0110] It should be noted that the process of performing the lifting and compression operations in the above steps S230-S240 can be specifically performed using a weight prediction algorithm (SENet algorithm). The above weight prediction algorithm can be specifically embedded in a specific level of the content recall model to constitute a weight prediction layer of the content recall model. After compressing and lifting the fused feature sequence using the SENet algorithm (i.e., the above steps S230-S240), it is possible to learn the importance of different target content features, weight important target content features, and / or weaken target content features that do not contain much information.

[0111] Step S250: determining the target browsing content feature according to the correlation weight between each target content feature and the candidate content feature in the target feature sequence and the target feature sequence.

[0112] Step S260: recalling the candidate contents according to the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate contents.

[0113] By adopting the above-mentioned content recall method of the present application, when obtaining the correlation weight between the target content feature and the candidate content feature, the compression mechanism of SENet can be used to make the correlation evaluation of different target content features and the candidate content feature produce non-linear association, so that the correlation weight calculation is no longer performed independently between single target content features. In addition, since the correlation weight calculation of a single target content feature and a candidate content feature incorporates the information of the complete sequence correlation, and based on the sigmoid output of SENet, the correlation weight calculation no longer adopts the softmax normalization method, but adopts the activation function to avoid the normalization constraint on the sum of sequence correlations, and also avoids the polarization problem that the softmax approximate maximum value makes the Attention weight probability often concentrated on a small number of one or two behavior items, so that the feature vector (target browsing content feature) of interest to the user can be more accurately extracted from the target content sequence. When the obtained target browsing content is more accurate, the result obtained by recalling the candidate content based on the target browsing content feature, the object feature corresponding to the target object, and the candidate content feature corresponding to the candidate content is also more accurate.

[0114] Please refer to Figure 5 and Figure 6 As shown, the embodiment of the present application also provides a content recall method, which is specifically applied in a content interaction platform. The object corresponding to the content interaction platform can be retrieved through the content interaction platform. Figure 5 The interface shown in the figure can be used to interact with other users, watch short videos, and view dynamic information of other objects or users. Figure 6 For example, a user uses a content interaction platform to watch a short video, so that the content interaction platform recalls the video played by the user for more than 30 seconds. The method includes:

[0115] Obtain the original input features (including: information corresponding to the target object, such as gender, age, and whether it is the weekend, information of the video to be recalled, such as video ID, video publisher ID, video type, etc., and the video content sequence corresponding to the target object, which includes the video IDs, video publisher IDs, and video types corresponding to multiple historically browsed videos, etc.), map the original input features to vector representations through the parameter table corresponding to the Embedding layer, and obtain feature vectors corresponding to each original input feature, that is, obtain object features corresponding to the target object, candidate content features corresponding to the candidate video content, and target feature sequences corresponding to the video content sequence.

[0116] For the target feature sequence corresponding to the video content sequence, the target feature sequence includes the video content feature vectors corresponding to each historically browsed video; if the length of a video content feature vector exceeds the set length, the length of the video content feature vector can be truncated to L; for a video content feature vector whose length does not reach L, the placeholder can be padded on the far left <pad>, so that the length of the padded video content feature vector is L, so the target feature sequence corresponding to the video content sequence can be obtained as: Represents the video content feature vector corresponding to the i-th video in the video content sequence. Assume that the dimension of the video content feature vector is d V , then the dimension of the target feature sequence is d V ×L.

[0117] When the candidate content features and the target feature sequence are fused to obtain the fused feature sequence, since the candidate content features are composed of multiple parts (such as video ID, author ID, category, etc.), if the candidate content features are recorded as Q, its dimension is d Q . Furthermore, if d Q and the dimension d of the video content feature vector V Different from the above, the candidate content feature Q can be linearly transformed so that the dimension of the candidate content feature after linear transformation is d V ; That is, align the dimension of the candidate content features to d V Then we can get: Q′=K·Q T Q′ is the candidate content feature after linear transformation.

[0118] Among them, K represents the parameter matrix, whose dimension is d V ×d Q Then, dot product operations are performed on the linearly transformed candidate content features and the target feature sequence to obtain a vector dot product score sequence Z (ie, fusion feature sequence) of length L, where Z = Q′·V.

[0119] When determining the correlation weights between each target content feature and the candidate content feature in the target feature sequence according to the fused feature sequence, the above-mentioned vector dot product score sequence Z can be used as the input summary vector of SENet, and the following compression and pulling operations are performed.

[0120] Among them, compression (Squeeze), that is, compression and dimensionality reduction operation, can be specifically performed by compressing each fusion feature in the fusion feature sequence through a compression parameter matrix to obtain a first intermediate feature sequence; and the first intermediate feature sequence is activated to obtain a fusion feature sequence after dimensionality reduction. Exemplarily, the process of compression and dimensionality reduction can be described by the following formula: R sq =f sq (Z) = a1(W sq `Z T ), where R sq is the fusion feature sequence after dimensionality reduction, whose dimension is L / r, r represents the compression multiple; Wsq is the compression parameter matrix, whose dimension is a1 is a nonlinear activation function, which can be a sigmoid function.

[0121] Excitation, or the stretching operation, is a dimensionality-increasing operation on the intermediate result of compression and dimensionality reduction. Specifically, the fusion feature sequence after dimensionality reduction can be processed by the stretching parameter matrix to obtain a second intermediate feature sequence; the second intermediate feature sequence is activated to obtain the correlation weight between each target content feature and the candidate content feature. Exemplarily, the following excitation calculation formula can be used: R ex =f ex (R sq )=a2(W ex ·R sq ), where R ex It includes the correlation weights between each target content feature and the candidate content feature, whose dimension is L, and Wex is the stretching parameter matrix, whose dimension is a2 is the activation function, such as the sigmoid function Based on the sigmoid function, the obtained correlation weight can be a value between 0 and 1.

[0122] After obtaining the SENet weight (the correlation weight between each target content feature and the candidate content feature), a weighted summing Pooling operation can be performed on each target content feature in the target feature sequence to obtain the user's interest representation (target browsing content feature), that is,

[0123]

[0124] After processing the features corresponding to the target content sequence of the target object into interest representation, and then splicing other feature vectors, the input of the DNN network can be obtained. Among them, Embd (Profile & Context) represents object characteristics.

[0125] The estimated probability value of the final recall model for the candidate content playback time greater than 30 seconds is p=h(X), where h(·) represents the processing process of DNN. If the above estimated probability value is greater than the preset probability threshold, the candidate content can be recalled.

[0126] Taking the candidate video as a travel-related video, for user A, there are a total of 100 videos in his historical deep playback video sequence, of which only one is a travel-related video, and the remaining 99 are videos with otaku attributes such as the second dimension / anime and have nothing to do with travel. For user B, there are a total of 100 videos in his historical deep playback behavior, of which only one is a travel-related video, and the remaining 99 are videos that are related to travel to a certain extent, such as food.

[0127] Obviously, from the complete video sequence, it can be inferred that user B actually likes traveling more than user A. However, if the relevant technology is used, for user A's video sequence and the video to be estimated, the possible result of the attention weight calculation based on softmax normalization is: the weight of the video related to travel is 1, and the weight of the other 99 sequence videos is 0. For user B's video sequence and the video to be estimated, the possible result of the attention weight calculation based on softmax normalization is: the weight of the video related to travel is 0.505, and the weight of the other 99 food sequence videos is 0.005. That is, using the solution in the relevant technology, the weight of the relevance of user A's single travel behavior is more prominent than that of B, and the relevance of user B's travel behavior will be lowered by other related behaviors. Therefore, using the algorithm in the relevant technology, it can be obtained that 99 of user A's 100 behaviors are concentrated home-attribute videos that are not related to travel. Then his interest is mainly concentrated in the part that is not related to travel. His only travel-related video seriously deviates from the main interest distribution, which may have a considerable degree of random attributes and is not suitable for extraction as a representation of user A's interest. For user B, food videos and travel are somewhat related. At this time, for user B, a single travel behavior in his behavior sequence reflects the interest of user B in the association with food. Therefore, using the content recall method in the related technology will cause the recall result to be inaccurate.

[0128] By adopting the scheme of the present application, it is possible to obtain a fused feature sequence for each user by establishing a dot product between the candidate content features and the target feature sequence of the user; and when determining the correlation weights between each target content feature and the candidate content features in the target feature sequence based on the fused feature sequence, it is fully considered that the probability of the candidate content feature executing a specified behavior is related to the historical browsing content that executes the specified behavior, that is, the candidate content feature is correlated with the global context of the target feature sequence. Therefore, when determining the target browsing content features based on the correlation weights between each target content feature and the candidate content features in the target feature sequence and the target feature sequence, the obtained target browsing content features can accurately express the features of the target object's concerned content, and then the candidate content is recalled based on the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content, to ensure that the candidate content recalled for the target object is the content that the target object is concerned about, or the content that the target object has a high probability of being interested in, thereby ensuring that the accuracy of content recall is higher.

[0129] It should be understood that, although the steps in the flowcharts involved in the above embodiments are displayed in sequence according to the indication of the arrows, these steps are not necessarily executed in sequence according to the order indicated by the arrows. Unless there is a clear explanation in this article, the execution of these steps is not strictly limited in order, and these steps can be executed in other orders. Moreover, at least a part of the steps in the flowcharts involved in the above embodiments may include multiple steps or multiple stages, and these steps or stages are not necessarily executed at the same time, but can be executed at different times, and the execution order of these steps or stages is not necessarily carried out in sequence, but can be executed in turn or alternately with other steps or at least a part of the steps or stages in other steps.

[0130] See also Figure 7 Another embodiment of the present application provides a content recall device 300, which includes: a data acquisition module 310, a feature fusion module 320, a weight determination module 330, a feature determination module 340, and a recall processing module 350. The data acquisition module 310 is used to obtain candidate content features corresponding to the candidate content and a target feature sequence corresponding to the target content sequence, wherein the target content sequence includes content attribute information corresponding to multiple historical browsing contents that trigger a specified behavior of the target object; the feature fusion module 320 is used to fuse the candidate content features and the target feature sequence to obtain a fused feature sequence; the weight determination module 330 is used to determine the correlation weight between each target content feature in the target feature sequence and the candidate content feature according to the fused feature sequence; the feature determination module 340 is used to determine the target browsing content feature according to the correlation weight between each target content feature in the target feature sequence and the candidate content feature and the target feature sequence; the recall processing module 350 is used to recall the candidate content according to the target browsing content features, the object features corresponding to the target object, and the candidate content features corresponding to the candidate content.

[0131] In one possible implementation, the data acquisition module 310 includes an information extraction submodule and a feature extraction submodule. The information extraction submodule is used to extract a target content sequence based on the information browsing history of the target object; the feature extraction submodule is used to extract features from the target content sequence to obtain a target feature sequence.

[0132] In one possible implementation, the data acquisition module 310 is also used to obtain target portrait information of the target object and context information for the target object to acquire information, the context information including at least one of the information acquisition time, the information acquisition location, and the mood of the target object; and perform feature extraction on the target portrait information and the context information respectively to obtain object features including portrait features corresponding to the target portrait information and context features corresponding to the context information.

[0133] In one possible implementation, the feature fusion module 320 is further configured to perform a dot product process on the candidate content features and the target feature sequence to obtain a fused feature sequence.

[0134] In one possible implementation, the fused feature sequence includes multiple fused features, each fused feature is obtained by fusing candidate content with a historical browsing content feature; the weight determination module 330 includes a compression processing submodule and a stretching processing submodule, the compression processing submodule is used to compress and reduce the dimension of each fused feature in the fused feature sequence to obtain a fused feature sequence after dimension reduction; the stretching processing submodule is used to stretch the fused feature sequence after dimension reduction to obtain the correlation weight between each target content feature and the candidate content feature in the target feature sequence.

[0135] In one possible implementation, the compression processing submodule includes a compression processing unit and a first activation unit; the compression processing unit is used to compress each fused feature in the fused feature sequence through a compression parameter matrix to obtain a first intermediate feature sequence; the first activation unit is used to activate the first intermediate feature sequence to obtain a fused feature sequence after dimensionality reduction.

[0136] In one possible implementation, the stretching processing submodule includes a stretching processing unit and a second activation unit; the stretching processing unit is used to perform dimensionality increase processing on the fused feature sequence after dimension reduction through a stretching parameter matrix to obtain a second intermediate feature sequence; the second activation unit is used to activate the second intermediate feature sequence to obtain the correlation weight between each target content feature and the candidate content feature.

[0137] In one possible implementation, the feature determination module 340 is further used to perform weighted summation of the target content features in the target feature sequence according to the correlation weights between each target content feature in the target feature sequence and the candidate content features to obtain a target browsing content feature vector; and perform pooling processing on the target browsing content feature vector to obtain the target browsing content feature.

[0138] In one possible implementation, the recall processing module 350 includes a fusion submodule, a probability prediction submodule and a recall processing submodule; the fusion submodule is used to fuse the target browsing content features, the object features corresponding to the target object and the candidate content features corresponding to the candidate content to obtain the target fusion features; the probability prediction submodule is used to predict the probability of the target object performing a specified behavior on the candidate content based on the fusion features; the recall submodule is used to recall the candidate content when the probability is greater than the probability threshold.

[0139] Each module in the above-mentioned device can be implemented in whole or in part by software, hardware and a combination thereof. Each of the above-mentioned modules can be embedded in or independent of a processor in a computer device in the form of hardware, or can be stored in a memory in a computer device in the form of software, so that the processor can call and execute operations corresponding to each of the above modules. It should be noted that the device embodiment in this application corresponds to the aforementioned method embodiment. The specific principles in the device embodiment can be found in the contents of the aforementioned method embodiment, which will not be repeated here.

[0140] The following will be combined Figure 8 An electronic device provided by the present application is described.

[0141] See also Figure 8 Based on the content recall method provided in the above embodiment, the embodiment of the present application also provides another electronic device 100 including a processor 102 that can execute the above method. The electronic device 100 can be a server or a terminal device, and the terminal device can be a smart phone, tablet computer, computer, portable computer or other device.

[0142] The electronic device 100 further includes a memory 104 . The memory 104 stores a program that can execute the contents of the aforementioned embodiments, and the processor 102 can execute the program stored in the memory 104 .

[0143] Among them, the processor 102 may include one or more cores for processing data and a message matrix unit. The processor 102 uses various interfaces and lines to connect various parts of the entire electronic device 100, and executes various functions and processes data of the electronic device 100 by running or executing instructions, programs, code sets or instruction sets stored in the memory 104, and calling data stored in the memory 104. Optionally, the processor 102 can be implemented in at least one hardware form of digital signal processing (Digital Signal Processing, DSP), field programmable gate array (Field-Programmable Gate Array, FPGA), and programmable logic array (Programmable Logic Array, PLA). The processor 102 can integrate one or a combination of a central processing unit (Central Processing Unit, CPU), a graphics processing unit (Graphics Processing Unit, GPU) and a modem. Among them, the CPU mainly processes the operating system, user interface and application programs; the GPU is responsible for rendering and drawing display content; and the modem is used to process wireless communications. It can be understood that the above-mentioned modem may not be integrated into the processor 102, but may be implemented separately through a communication chip.

[0144] The memory 104 may include a random access memory (RAM) or a read-only memory (ROM). The memory 104 may be used to store instructions, programs, codes, code sets or instruction sets. The memory 104 may include a program storage area and a data storage area, wherein the program storage area may store instructions for implementing an operating system, instructions for implementing at least one function, instructions for implementing the following various method embodiments, etc. The data storage area may also store data acquired by the electronic device 100 during use (e.g., candidate content, content attribute information corresponding to each of a plurality of historical browsing contents, and candidate content features corresponding to the candidate content), etc.

[0145] The electronic device 100 may also include a network module and a screen. The network module is used to receive and send electromagnetic waves, realize the mutual conversion between electromagnetic waves and electrical signals, and thus communicate with a communication network or other devices, such as communicating with an audio playback device. The network module may include various existing circuit components for performing these functions, such as an antenna, a radio frequency transceiver, a digital signal processor, an encryption / decryption chip, a user identity module (SIM) card, a memory, etc. The network module may communicate with various networks such as the Internet, an intranet, a wireless network, or communicate with other devices via a wireless network. The above-mentioned wireless network may include a cellular telephone network, a wireless local area network, or a metropolitan area network.

[0146] In some embodiments, the electronic device 100 may further include: a peripheral interface 106 and at least one peripheral device. The processor 102, the memory 104 and the peripheral interface 106 may be connected via a bus or a signal line. Each peripheral device may be connected to the peripheral interface via a bus, a signal line or a circuit board. Specifically, the peripheral device includes: at least one of a radio frequency component 108, a positioning component 112, a camera 114, an audio component 116, a display screen 118 and a power supply 122.

[0147] The peripheral interface 106 may be used to connect at least one peripheral device related to I / O (Input / Output) to the processor 102 and the memory 104. In some embodiments, the processor 102, the memory 104, and the peripheral interface 106 are integrated on the same chip or circuit board; in some other embodiments, any one or two of the processor 102, the memory 104, and the peripheral interface 106 may be implemented on a separate chip or circuit board, which is not limited in the embodiments of the present application.

[0148] The radio frequency component 108 is used to receive and transmit RF (Radio Frequency) signals, also known as electromagnetic signals. The radio frequency component 108 communicates with the communication network and other communication devices through electromagnetic signals. The radio frequency component 108 converts electrical signals into electromagnetic signals for transmission, or converts received electromagnetic signals into electrical signals. Optionally, the radio frequency component 108 includes: an antenna system, an RF transceiver, one or more amplifiers, a tuner, an oscillator, a digital signal processor, a codec chipset, a user identity module card, and the like. The radio frequency component 108 can communicate with other terminals through at least one wireless communication protocol. The wireless communication protocol includes, but is not limited to: the World Wide Web, a metropolitan area network, an intranet, various generations of mobile communication networks (2G, 3G, 4G and 5G), a wireless local area network and / or a WiFi (Wireless Fidelity) network. In some embodiments, the radio frequency component 108 may also include circuits related to NFC (Near Field Communication), which is not limited in this application.

[0149] The positioning component 112 is used to locate the current geographic location of the electronic device to implement navigation or LBS (Location Based Service). The positioning component 112 can be a positioning component based on the US GPS (Global Positioning System), Beidou system or Galileo system.

[0150] The camera 114 is used to capture images or videos. Optionally, the camera 114 includes a front camera and a rear camera. Usually, the front camera is arranged on the front panel of the electronic device 100, and the rear camera is arranged on the back of the electronic device 100. In some embodiments, there are at least two rear cameras, which are any one of a main camera, a depth of field camera, a wide-angle camera, and a telephoto camera, so as to realize the fusion of the main camera and the depth of field camera to realize the background blur function, the fusion of the main camera and the wide-angle camera to realize the panoramic shooting and VR (Virtual Reality) shooting function or other fusion shooting functions. In some embodiments, the camera 114 may also include a flash. The flash can be a monochrome temperature flash or a dual-color temperature flash. A dual-color temperature flash refers to a combination of a warm light flash and a cold light flash, which can be used for light compensation at different color temperatures.

[0151] The audio component 116 may include a microphone and a speaker. The microphone is used to collect sound waves from the user and the environment, and convert the sound waves into electrical signals and input them into the processor 102 for processing, or input them into the RF component 108 to achieve voice communication. For the purpose of stereo acquisition or noise reduction, there can be multiple microphones, which are respectively arranged at different parts of the electronic device 100. The microphone can also be an array microphone or an omnidirectional acquisition microphone. The speaker is used to convert the electrical signal from the processor 102 or the RF component 108 into sound waves. The speaker can be a traditional film speaker or a piezoelectric ceramic speaker. When the speaker is a piezoelectric ceramic speaker, it can not only convert the electrical signal into sound waves audible to humans, but also convert the electrical signal into sound waves inaudible to humans for purposes such as ranging. In some embodiments, the audio component 114 may also include a headphone jack.

[0152] The display screen 118 is used to display the UI (User Interface). The UI may include graphics, text, icons, videos, and any combination thereof. When the display screen 118 is a touch display screen, the display screen 118 also has the ability to collect touch signals on the surface or above the surface of the display screen 118. The touch signal can be input to the processor 102 as a control signal for processing. At this time, the display screen 118 can also be used to provide virtual buttons and / or virtual keyboards, also known as soft buttons and / or soft keyboards. In some embodiments, the display screen 118 can be one, and the front panel of the electronic device 100 is set; in other embodiments, the display screen 118 can be at least two, which are respectively set on different surfaces of the electronic device 100 or are folded; in some other embodiments, the display screen 118 can be a flexible display screen, which is set on the curved surface or folded surface of the electronic device 100. Even, the display screen 118 can also be set as a non-rectangular irregular shape, that is, a special-shaped screen. The display screen 118 can be made of materials such as LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode, machine light-emitting diode).

[0153] The power supply 122 is used to power various components in the electronic device 100. The power supply 122 can be an alternating current, a direct current, a disposable battery, or a rechargeable battery. When the power supply 122 includes a rechargeable battery, the rechargeable battery can be a wired rechargeable battery or a wireless rechargeable battery. A wired rechargeable battery is a battery that is charged through a wired line, and a wireless rechargeable battery is a battery that is charged through a wireless coil. The rechargeable battery can also be used to support fast charging technology.

[0154] The embodiment of the present application also provides a structural block diagram of a computer-readable storage medium. The computer-readable medium stores program codes, which can be called by a processor to execute the method described in the above method embodiment.

[0155] The computer-readable storage medium may be an electronic memory such as a flash memory, an EEPROM (electrically erasable programmable read-only memory), an EPROM, a hard disk, or a ROM. Optionally, the computer-readable storage medium includes a non-transitory computer-readable storage medium. The computer-readable storage medium has storage space for program codes that execute any of the method steps in the above method. These program codes can be read from or written to one or more computer program products. The program code can be compressed, for example, in an appropriate form.

[0156] The embodiment of the present application also provides a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. The processor of the computer device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the computer device performs the method described in the above various optional implementations.

[0157] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present application, rather than to limit it. Although the present application has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present application.< / pad>

Claims

1. A content recall method, characterized in that: The method comprises: Acquire candidate content features corresponding to the candidate content and a target feature sequence corresponding to the target content sequence, wherein the target content sequence includes content attribute information corresponding to each of a plurality of historical browsing contents that trigger a specified behavior of the target object; Fusing the candidate content features with the target feature sequence to obtain a fused feature sequence; Determining, according to the fused feature sequence, a correlation weight between each target content feature in the target feature sequence and the candidate content feature; Determining a target browsing content feature according to a correlation weight between each target content feature in the target feature sequence and the candidate content feature and the target feature sequence; The candidate content is recalled according to the target browsing content feature, the object feature corresponding to the target object, and the candidate content feature corresponding to the candidate content.

2. The method according to claim 1, characterized in that The recalling of the candidate content according to the target browsing content feature, the object feature corresponding to the target object, and the candidate content feature corresponding to the candidate content includes: fusing the target browsing content feature, the object feature corresponding to the target object, and the candidate content feature corresponding to the candidate content to obtain a target fusion feature; Based on the fusion feature, predicting the probability that the target object performs the specified behavior on the candidate content; If the probability is greater than the probability threshold, the candidate content is recalled.

3. The method according to claim 1, characterized in that: The fusion feature sequence includes a plurality of fusion features, each of which is obtained by fusing the candidate content with a historical browsing content feature; The step of determining the correlation weight between each target content feature in the target feature sequence and the candidate content feature according to the fused feature sequence includes: Performing compression and dimensionality reduction processing on each fused feature in the fused feature sequence to obtain a fused feature sequence after dimensionality reduction; The fused feature sequence after dimension reduction is stretched to obtain the correlation weight between each target content feature in the target feature sequence and the candidate content feature.

4. The method according to claim 3, characterized in that The step of compressing and reducing the dimension of each fused feature in the fused feature sequence to obtain a fused feature sequence after dimension reduction includes: Compressing each fused feature in the fused feature sequence through a compression parameter matrix to obtain a first intermediate feature sequence; The first intermediate feature sequence is activated to obtain the fusion feature sequence after dimension reduction.

5. The method according to claim 3, characterized in that: The stretching process is performed on the fusion feature sequence after dimension reduction to obtain the correlation weight between each target content feature in the target feature sequence and the candidate content feature, including: Performing dimension-up processing on the fusion feature sequence after dimension reduction by stretching the parameter matrix to obtain a second intermediate feature sequence; The second intermediate feature sequence is activated to obtain the correlation weight between each target content feature and the candidate content feature.

6. The method according to claim 1, characterized in that The step of determining the target browsing content feature according to the correlation weight between each target content feature in the target feature sequence and the candidate content feature and the target feature sequence includes: According to the correlation weight between each target content feature in the target feature sequence and the candidate content feature, weighted summing the target content features in the target feature sequence to obtain a target browsing content feature vector; Pooling is performed on the target browsing content feature vector to obtain target browsing content features.

7. The method according to claim 1, characterized in that The candidate content feature and the target feature sequence are fused to obtain a fused feature sequence, including: The candidate content features are subjected to dot product processing with the target feature sequence to obtain the fused feature sequence.

8. The method according to claim 1, characterized in that Obtain the target feature sequence corresponding to the target content sequence, including: Extracting a target content sequence based on the information browsing record of the target object; Feature extraction is performed on the target content sequence to obtain a target feature sequence.

9. The method according to claim 1, characterized in that: The method further comprises: Acquire target portrait information of the target object and context information of information acquisition by the target object, wherein the context information includes at least one of information acquisition time, information acquisition location, and mood of the target object; Feature extraction is performed on the target portrait information and the context information respectively to obtain object features including portrait features corresponding to the target portrait information and context features corresponding to the context information.

10. A content recall device, characterized in that: The device comprises: A data acquisition module, used to acquire candidate content features corresponding to the candidate content and a target feature sequence corresponding to the target content sequence, wherein the target content sequence includes content attribute information corresponding to each of a plurality of historical browsing contents that trigger a specified behavior of a target object; A feature fusion module, used to fuse the candidate content features with the target feature sequence to obtain a fused feature sequence; A weight determination module, used to determine the correlation weight between each target content feature in the target feature sequence and the candidate content feature according to the fused feature sequence; A feature determination module, configured to determine a target browsing content feature according to a correlation weight between each target content feature in the target feature sequence and the candidate content feature and the target feature sequence; The recall processing module is used to perform recall processing on the candidate content according to the target browsing content features, the object features corresponding to the target object and the candidate content features corresponding to the candidate content.

11. An electronic device, characterized in that: include: one or more processors; Memory; One or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, and the one or more programs are configured to execute the method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores program codes, and the program codes can be called by a processor to execute the method according to any one of claims 1 to 9.

13. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method described in any one of claims 1 to 9 are implemented.

Images