Resource sorting method and device, equipment and storage medium

Abstract

The invention provides a resource sorting method and device, equipment and a storage medium, and relates to the technical field of computers, in particular to the technical fields of artificial intelligence, big data, intelligent recommendation and the like. According to the specific implementation scheme, in response to a resource recommendation request of a target object, object features of the target object, resource features of candidate resources and mixed features are obtained; the mixed feature is obtained based on cross feature query of the target object and the candidate resource, and the mixed feature comprises a plurality of information sequences of the target object in a plurality of time windows; performing fusion operation on the plurality of information sequences of the plurality of time windows to obtain sequence characterization; and inputting the object features, the resource features and the sequence features into a sorting model to obtain sorting positions of the candidate resources in the resource sequence.

Description

TECHNICAL FIELD

[0001] The present disclosure relates to the technical field of computer, and particularly relates to the technical field of artificial intelligence, big data, intelligent recommendation, etc. BACKGROUND

[0002] With the rapid development of Internet technology, information is growing explosively, and users often have difficulty in quickly finding content meeting their own needs in the massive information, which makes accurate recommendation a key demand of information service.

[0003] Improving the accuracy of resource recommendation can help users obtain expected information, and thus improve user experience. In related technologies, resource features can be used to recommend resources for users. However, the information provided by the resource features is limited, and the recommendation effect needs to be further optimized. SUMMARY

[0004] The present disclosure provides a resource ranking method, device, equipment and storage medium.

[0005] According to an aspect of the present disclosure, a resource ranking method is provided, comprising: In response to a resource recommendation request of a target object, obtaining an object feature of the target object, a resource feature of a candidate resource, and a mixed feature; the mixed feature is obtained based on a cross-feature query of the target object and the candidate resource, and the mixed feature comprises a plurality of information sequences of the target object in a plurality of time windows; Performing fusion operation on the plurality of information sequences in the plurality of time windows to obtain a sequence feature; Inputting the object feature, the resource feature and the sequence feature into a ranking model to obtain a ranking position of the candidate resource in a resource sequence.

[0006] According to another aspect of the present disclosure, a resource ranking device is provided, comprising: The obtaining module is configured to, in response to a resource recommendation request of a target object, obtain an object feature of the target object, a resource feature of a candidate resource, and a mixed feature; the mixed feature is obtained based on a cross-feature query of the target object and the candidate resource, and the mixed feature comprises a plurality of information sequences of the target object in a plurality of time windows; The fusion module is configured to perform fusion operation on the plurality of information sequences in the plurality of time windows to obtain a sequence feature; The ranking module is configured to input the object feature, the resource feature and the sequence feature into a ranking model to obtain a ranking position of the candidate resource in a resource sequence.

[0007] According to another aspect of the present disclosure, an electronic device is provided, comprising: at least one processor; and a memory connected to the at least one processor in communication; wherein, The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any of the embodiments of the present disclosure.

[0008] According to another aspect of the present disclosure, a non-transitory computer-readable storage medium storing computer instructions is provided, wherein the computer instructions are used to enable the computer to perform the method according to any of the embodiments of the present disclosure.

[0009] According to another aspect of the present disclosure, a computer program product is provided, comprising a computer program which, when executed by a processor, implements the method according to any of the embodiments of the present disclosure.

[0010] It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of the present disclosure, nor is it used to limit the scope of the present disclosure. Other features of the present disclosure will become apparent through the following description. BRIEF DESCRIPTION OF DRAWINGS

[0011] The accompanying drawings are used to better understand the present scheme, and do not constitute a limitation on the present disclosure. Among them: Figure 1 is a flowchart of a resource ranking method according to an embodiment of the present disclosure; Figure 2 is a flowchart of obtaining object features of a target object, resource features of a candidate resource, and mixed features according to an embodiment of the present disclosure; Figure 3 is a flowchart of obtaining sequence features according to an embodiment of the present disclosure; Figure 4 is a flowchart of training a ranking model according to an embodiment of the present disclosure; Figure 5 is a flowchart of a resource ranking method according to an embodiment of the present disclosure; Figure 6 is a structural diagram of a resource ranking device according to an embodiment of the present disclosure; Figure 7 is a block diagram of an electronic device for implementing a resource ranking method according to an embodiment of the present disclosure. DETAILED DESCRIPTION

[0012] Exemplary embodiments of the present disclosure are described below with reference to the accompanying drawings, which include various details of the embodiments of the present disclosure to assist in understanding, and should be considered as merely exemplary. Therefore, those of ordinary skill in the art should recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope of the present disclosure. Also, in order to make the description clear and concise, the description of well-known functions and structures is omitted in the following description.

[0013] The terms “first,” “second,” etc., used in this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion, such as including a series of steps or units. A method, system, product, or apparatus is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses.

[0014] It should be noted that, unless it is explicitly stated that there is a sequential order of execution between different operations, or that there is a sequential order of execution between different operations in terms of technical implementation, the execution order between multiple operations may not be significant, and multiple operations may be executed simultaneously.

[0015] In related technologies, the feature input of resource ranking models generally comes from two independent dimensions: one is user-side features obtained through feature services, and the other is resource-side features also obtained through feature services. These two types of features are completely isolated in the logical definition and computation generation stages. They are only simply concatenated through feature services during model training or online inference to ultimately form the input feature set of the resource ranking model.

[0016] However, user-side features and resource-side features are essentially isolated information silos. Specifically, user-side features are long-term global user profiles built from massive amounts of historical data, reflecting users' overall preferences and behavioral patterns over a longer period; while resource-side features encompass the inherent attributes of resources and statistical information based on historical data, used to characterize the resource's own traits and past performance. Since ranking models only receive a simple union of the two types of features, they cannot effectively uncover the dynamic interaction between target users and candidate resources, nor can they reflect the potential personalized matching relationships between users and resources.

[0017] In view of this, this disclosure provides a resource ranking method that integrates user features, resource features, and dynamic features between the two (hereinafter referred to as hybrid features) to improve the accuracy of the ranking model in judging the matching degree between user interests and resources, thereby improving the accuracy of recommendations.

[0018] The acquisition, storage, and application of any type of information, such as user personal information, involved in the technical solutions disclosed herein comply with relevant laws and regulations and do not violate public order and good morals.

[0019] like Figure 1 The diagram shown is a flowchart illustrating the resource sorting method provided in this disclosure, including the following: S101, in response to a resource recommendation request of a target object, obtaining object features of the target object, resource features of a candidate resource, and mixed features; the mixed features are obtained based on cross-feature queries of the target object and the candidate resource, and the mixed features include multiple information sequences of the target object in multiple time windows.

[0020] The object features of the target object (i.e., the user) are a feature set describing the inherent attributes and long-term behavior preferences of the target object. The feature set can be obtained based on the object identifier (cuid, Customer Unique Identifier) of the target object as an index, and can reflect the inherent attributes and historical behavior patterns of the target object.

[0021] The candidate resource is a resource in a resource set to be recommended to the target object, and each resource in the set can be processed as a candidate resource. The candidate resource can be a resource obtained through coarse sorting, or an original resource. That is, the mixed features provided by the present disclosure can be applied to coarse sorting, and can also be applied to fine sorting.

[0022] The resource features of the candidate resource are a feature set describing the inherent attributes of the candidate resource and associated deep information. The feature set is obtained based on the target information identifier of the candidate resource as an index.

[0023] The mixed features are obtained based on cross-feature queries of the target object and the candidate resource. The cross-feature is a feature associated with the interaction relationship between the target object and the candidate resource. For example, if the target object has clicked on a historical resource, the candidate resource associated with the historical resource has a specific association with the user and the candidate resource, and has an interaction feature between the user and the candidate resource. Based on the interaction feature, the mixed features closely associated with the candidate resource can be queried to construct a unique context scenario of the target object with respect to the candidate resource.

[0024] The time window is a time interval for time sequence division of the operation behavior of the target object. The single information sequence is an ordered behavior record generated by the target object in a single time window, and the ordered behavior record is associated with the target object and the candidate resource. Therefore, the potential association between the target object and the candidate resource can be described.

[0025] S102, performing a fusion operation on the multiple information sequences of the multiple time windows to obtain a sequence feature.

[0026] In implementation, the fusion operation can adopt fusion methods such as splicing and weighted fusion. In implementation, the multiple information sequences of the multiple time windows can be compressed into a single-dimensional sequence feature through a Sum Pooling (sum pooling) operation.

[0027] S103, input the object feature, the resource feature and the sequence feature into the ranking model to obtain a ranking position of the candidate resource in the resource sequence.

[0028] The ranking model can be a machine learning model for prioritizing candidate resources, an artificial intelligence model based on a neural network, or even a large model.

[0029] The large model (LLM, Large Language Models) refers to a deep learning model trained using a large amount of text data, which can generate natural language text or understand the meaning of language text. The large model can handle various natural language tasks such as text classification, question answering, and dialogue.

[0030] In implementation, after inputting the object feature, the resource feature and the sequence feature into the ranking model, the ranking model can output a ranking score of each candidate resource. The ranking position of the candidate resource in the final recommendation list can be determined by the ranking score, and the higher the score, the higher the position, and the higher the corresponding recommendation priority.

[0031] In the embodiments of the present disclosure, on the basis of the object feature and the resource feature, a hybrid feature is additionally introduced, which can make up for the technical defects of simply splicing the user's long-term global portrait and the inherent attributes of the resource in the related art, and effectively mine the personalized matching relationship between the target object and the candidate resource. The fusion operation of multiple information sequences in multiple time windows is performed to obtain sequence features, which can solve the problem that the long-term global portrait cannot identify the user's instantaneous interest and real-time decision mode in the related art based on the historical interaction trajectory of the target object with respect to the candidate resource, and improve the accuracy of the matching degree between the user's interest and the resource in the ranking model. The object feature, the resource feature and the sequence feature are cooperatively input into the ranking model, which realizes the cooperative modeling of the three types of core features of user attributes, resource attributes and potential interaction features between users and resources, so that the ranking model can fully mine the potential association rules between the features, the ranking model can comprehensively judge the recommendation priority of the candidate resource from multiple dimensions, and the output resource ranking result is more in line with the real demand, thereby improving the accuracy of resource recommendation.

[0032] In the embodiments of the present disclosure, the object feature of the target object, the resource feature of the candidate resource, and the hybrid feature are obtained, as shown in the following table: Figure 2 The table includes the following contents: S201, send a first query request based on the resource recommendation request to the feature service to obtain the object feature and the resource feature from the feature service.

[0033] In implementation, the object identifier can be parsed from the resource recommendation request, and the list of potential candidate resources can be obtained synchronously. Based on this, the first query request for the object feature and the resource feature is generated to construct the context information of this resource ranking.

[0034] The characteristic service indexes the cuid in the first query request, queries a target object side storage system (UMS, User Management System), and extracts inherent attributes and historical behaviors of the target object, to obtain corresponding object characteristics. The target object side storage system is a structured storage module that indexes the cuid of the target object as the core and stores inherent attributes and full historical behavior data of the target object.

[0035] The characteristic service indexes the list of nids (Network Identifier, candidate resource identifier) (candidate resource categories) in the first query request in batches, respectively queries the positive index library and the signal library on the resource side, and obtains corresponding resource characteristics. The positive index library is a structured storage module that indexes the nid of the candidate resource as the core and stores the basic attributes and associated entity information of the candidate resource. The signal library is a storage module that indexes the associated entity identifier (such as a commodity ID and an anchor ID) of the candidate resource and stores the deep statistical characteristics and dynamic interaction signals of the resource.

[0036] S202, cross characteristics are constructed based on the target object and the candidate resource.

[0037] In implementation, an identification pair composed of an object identifier of the target object and a target information identifier of the candidate resource can be constructed as the cross characteristics based on the object identifier of the target object and the target information identifier of the candidate resource. The target information identifier is an information identifier of shared information of the candidate resource and other resources.

[0038] The shared information is used to establish the commonality of the candidate resource and the historical resources operated by the target object. For example, information such as the theme, classification, and commodity detail category of a video that can establish association can be used as shared information. Assuming that A commodity, B commodity, and C commodity all belong to a certain brand of foundation liquid category identifier as shared information. For a video, the video type can be divided based on the content, such as the video theme as shared information.

[0039] In implementation, the object identifier of the target object and the target information identifier of the candidate resource are spliced through an underscore “_” to form a globally unique identification pair such as <cuid_nid>, and the identification pair is used as the cross characteristics.

[0040] In the embodiments of the present disclosure, the cross characteristics are constructed based on the target object and the candidate resource, which breaks through the limitation of the traditional one-to-one dimension of the object-resource pair and can effectively cover the interest preference characteristics of the target object for the same type of resource.

[0041] S203, the second query request containing the cross characteristics is sent to the online storage cluster through the built-in client to query the mixed characteristics as a composite key in the online storage cluster.

[0042] The built-in client is a component for communicating with the online storage cluster, responsible for sending the second query request to the online storage cluster and receiving the returned result.

[0043] The online storage cluster is a distributed storage system for storing a large amount of data, such as a cube (Multidimensional Data Cube).

[0044] In implementation, the cross feature can be queried in the online storage cluster as a composite key to obtain a mixed feature containing respective information sequences of the target object in multiple time windows.

[0045] In the embodiments of the present disclosure, by taking the object feature and the resource feature obtained by the feature service query as a query path, and taking the mixed feature based on the cross feature in the online storage cluster as another newly added query path, the precise separation and independent retrieval of different types of features can be realized, the coupling of multi-source feature storage and query is avoided, and the efficiency and flexibility of feature acquisition are greatly improved. By constructing the cross feature based on the target object and the candidate resource and taking it as a composite key, the mixed feature is queried in the online storage cluster, which can effectively filter the interference of irrelevant feature data, guarantee the relevance and accuracy of the retrieved mixed feature to the current recommendation request, and thus improve the recommendation efficiency.

[0046] In some other embodiments, the mixed feature can be generated based on the following steps: Step A1, periodically querying the interaction feature of the target object and at least one historical resource; For example, the interaction log of the target object and the historical resource can be scanned by an offline Spark (Apache Spark, a distributed architecture) / Flink (Apache Flink, a distributed architecture) task every day to obtain an interaction feature set.

[0047] Step A2, for each historical resource, generating an information sequence of the target object in the current time window based on the corresponding interaction feature of the historical resource; In implementation, for each historical resource, the target object corresponding thereto can be divided into a group, and for each group of data, the interaction behavior indicators in multiple time windows are counted to form a structured information sequence.

[0048] Step A3, in the online storage cluster, the information sequence of the target object in the current time window is stored in association under the composite key constructed based on the cross feature of the target object and the historical resource to update the information sequences of multiple time windows corresponding to the composite key, and a fusion feature is obtained.

[0049] For example, the composite key constructed based on the cross features of the target object and the historical resources and the information sequence corresponding to the composite key are stored in the online storage cluster in the format of <combined_key, feature_value>, so that the multiple information sequences of the target object in multiple time windows can be updated in time.

[0050] When online recommendation is performed, the information sequences of the n time windows of the corresponding candidate resources can be obtained in the order from recent to remote, so as to construct fusion features, n being a positive integer. Thus, on the time axis, the fusion features of the same target object are dynamically changed and can be dynamically updated following the operation of the target object.

[0051] In the embodiments of the present disclosure, the information sequences of the target object in multiple time windows are stored in association under the composite key constructed based on the cross features of the target object and the historical resources, so that the maximum amount of information can be retained for model exploration. Meanwhile, the online dynamic is converted into lightweight features, so as to avoid real-time calculation of complex sequence operations.

[0052] In the embodiments of the present disclosure, the multiple time windows corresponding to the mixed features do not overlap in the time range, and the time span of each time window is different; wherein the time window whose starting time is closer to the current time has a smaller time span.

[0053] For example, the time windows can be divided into 1 day, 1-3 days, 3-7 days, 7-30 days, 30-90 days, 90-360 days, and 360-1080 days.

[0054] In the embodiments of the present disclosure, by setting multiple time windows without overlapping in the time range, it can be ensured that any operation behavior of the target object is only attributed to a unique window, so as to avoid repeated statistics of behavior data. By using the window design of closer starting time and smaller time span, the recent behavior of the target object can be captured in fine granularity, so as to obtain the real-time interest preference of the target object, and the long-term behavior can be aggregated in coarse granularity, so as to reduce the over-fine calculation of the long-term behavior and avoid over-emphasizing the features of the long-term, thereby improving the accuracy of sorting and recommendation.

[0055] In the embodiments of the present disclosure, the information sequence of each time window includes statistical information of the target object in the context scene represented by the cross features.

[0056] The context scene is used to describe the instantaneous feature of the target object established for the most relevant or most similar historical resource of the candidate resource. The instantaneous feature can be expressed by statistical information, which can include at least one of the following statistical information: (1) Resource display times; That is, the total number of times that the historical resource associated with the candidate resource within the time window is presented to the target object can be the statistical result of the same type of historical resource.

[0057] (2) the number of clicks; Within the same context scenario and time window, the total number of times that the same type of historical resource is actively clicked by the target object.

[0058] (3) the number of conversion behaviors from the first operation to the second operation.

[0059] Within the time window, the total number of times that the target object further performs the second operation after completing the first operation for the same type of historical resource; wherein the first operation and the second operation are flexibly defined according to the actual business target. For example, the click operation is the first operation; any one of the purchase operation, the collection operation, the comment operation, the forwarding operation, and the download operation can be the second operation.

[0060] In the embodiments of the present disclosure, each time window includes the information sequence of the target object in the context scenario represented by the cross-feature. The information sequence can express the interest intensity of the target object for the corresponding candidate resource and the potential possible operation behavior by statistically analyzing the statistical features of a single time window, thereby improving the accuracy of resource sorting.

[0061] In other embodiments, the instantaneous feature represented by the context environment can also be used to describe the instantaneous interest evolution of the target object before the current acquisition of the recommended resource. For example, the target object is browsing a historical story at the last time point, and may be interested in a certain historical character at the next time point. In order to be able to sensitively capture this interest demand, an instantaneous feature can be constructed to measure the possibility of this demand. In implementation, a browsing resource list of the same type of resource of the historical resource can be constructed according to the historical resource browsed by the target object within the time window in which the historical resource is browsed. Thus, the short-term interest evolution of the user before and after viewing the historical resource can be constructed. The resource list can be used as the information sequence of the time window.

[0062] Of course, the statistical information and the resource list of the same time window can also jointly construct the information sequence of the time window.

[0063] In addition, in other embodiments, the statistical information and / or the feature vector of the resource list of the time window can also be counted as the information sequence of the time window.

[0064] In order to better understand the features of the user before and after viewing the historical resource, the resource content of the resource list can also be encoded to obtain the resource features, so as to construct the information sequence of the time window with at least one of the statistical information and the resource list.

[0065] In the embodiments of the present disclosure, a plurality of information sequences of a plurality of time windows are fused to obtain a sequence representation, as shown in the following formula: Figure 3 The sequence representation includes the following contents: S301, determining window weights of the plurality of time windows.

[0066] In some embodiments, the weight values of the time windows can be calculated by selecting an exponential decay function or a linear decay function in combination with a preset decay coefficient, so that the time window closer to the current time obtains a higher weight.

[0067] In some embodiments, the statistical information (such as the number of resource displays, the number of clicks, and the number of conversion behaviors) in each time window can be input into a preset neural network model, and the weight values corresponding to each time window are output after nonlinear transformation and normalization processing of the neural network model.

[0068] In some embodiments, the unique identifier of each historical resource operated by the target object in each time window can be extracted from the information sequence of each time window, and it is distinguished whether the behavior corresponding to each historical resource is a conversion behavior feature or a non-conversion behavior feature. Then, the two types of fine-grained features are input into a neural network model, and the influence degree of the historical behaviors in each time window on the interest of the candidate resource is learned by the preset neural network model, and then the weights of each time window are adaptively output.

[0069] In some embodiments, for each time window, the non-conversion behavior index (i.e., the proportion of the resource amount performing the first operation and not performing the second operation to the resource amount performing only the first operation) and the conversion behavior index (i.e., the proportion of the resource amount performing the first operation and performing the second operation to the resource amount performing only the first operation) of the specific resource (such as a commodity) can be extracted from the information sequence of the time window, and the time length proportion of the target object operating each historical resource in the time window (i.e., the proportion of the observation total time to the play total time). The above information is input into a weight estimation model (a kind of neural network model) to obtain the weight of each time window.

[0070] S302, based on the window weights of the time windows, performing weighted summation on the plurality of information sequences of the plurality of time windows to obtain a sequence representation.

[0071] In the embodiments of the present disclosure, by assigning different weights to different time windows, the contribution of the behavior of the target object in each time window to the current interest can be distinguished, and the problem of only focusing on the overall preference and behavior pattern in a long period in the related art can be avoided. Based on the weighted sum of the information sequences of the multiple time windows according to the weights of the time windows, the modeling features scattered in multiple time windows can be integrated into structured sequence features, so that the ranking model can more accurately capture the dynamic changes of the interest of the target object, improve the accuracy of resource ranking, and improve the accuracy of resource recommendation.

[0072] In the embodiments of the present disclosure, in order to enable the ranking model to adapt to the mixed features of the target object and perform resource ranking, the ranking model is trained based on the training samples in an offline stage.

[0073] By utilizing a large number of complete training samples in an offline environment, the fine-grained dynamic association rules between the target object and the candidate resources expressed by the mixed features are sufficiently learned, so that the ranking model has accurate resource ranking and recommendation decision-making capabilities, the understanding and processing capabilities of the ranking model for the mixed features are improved, and the resource ranking accuracy of the ranking model is improved.

[0074] In implementation, the training of the ranking model can include the following contents as shown in the following table. Figure 4 S401, constructing a composite key of the initial sample based on the cross features of the sample object and the sample resource in the initial sample.

[0075] In implementation, the initial sample includes the object identifier of the sample object and the historical resource operated by the sample object. The cross features of the sample object and the historical resource can be constructed based on the foregoing manner to form the composite key.

[0076] S402, querying the mixed features of the initial sample in the online storage cluster based on the composite key of the initial sample.

[0077] S403, splicing the mixed features of the initial sample and the sample object features and the sample resource features of the initial sample to obtain a training sample.

[0078] In implementation, the mixed features of the initial sample can be fused by using the foregoing manner to generate corresponding sequence features, and the sequence features and the sample object features and the sample resource features are spliced to obtain the training sample.

[0079] S404, training the ranking model based on the training sample.

[0080] ​In implementation, the model parameters of the ranking model can be optimized in a supervised manner through contrastive learning. For example, the ranking model can generate ranking scores, and then make a binary classification prediction on whether the sample object will perform a preset operation on the sample resource based on the ranking scores. The loss value is determined based on the predicted classification result and the true value, and the model parameters are optimized.

[0081] In the embodiments of the present disclosure, the cross features of the sample object and the sample resource reflect the association relationship between them, and the mixed features supplement more relevant information from the online storage cluster. Combining the mixed features with the sample object features and the sample resource features makes the training samples contain more comprehensive and rich feature dimensions. This helps the ranking model to learn more complex and accurate sample patterns and rules, thereby improving the performance and generalization ability of the ranking model.

[0082] In summary, the overall process of the resource ranking method provided in the embodiments of the present disclosure is as shown in Figure 5 as follows: 1) mixed feature generation and storage: First, the interaction log of the target object and the historical resource can be scanned periodically through Spark / Flink 501 to obtain an interaction feature set. For each historical resource, the information sequence of the target object in the current time window is generated based on the interaction features corresponding to the historical resource.

[0083] Next, the composite key of the cross features of the target object and the historical resource is constructed through Spark / Flink 501, and the multiple information sequences of the corresponding target object in multiple time windows are written in the format of <combined_key, feature_value> to a specific AFS path of the distributed file system (such as HDFS), and are also published to the high-performance online storage cluster 502 at the same time, so as to ensure that the same data snapshot is read for offline training and online service.

[0084] 2) offline training process: First, when training the ranking model, the feature service 503 obtains the sample object features and the sample resource features of the initial samples on the target object side and the resource side, respectively.

[0085] In implementation, the feature service 503 obtains the sample object identifier cui, queries the corresponding sample object features in the cache, and obtains the corresponding sample object features from the storage system in the case of not being able to find. Similarly, the feature service obtains the sample resource identifier nid, and preferentially queries the corresponding sample resource features in the cache, and queries the corresponding features in the corresponding resource index library in the case of not being able to find. The storage system is a structured storage module taking the sample object identifier cui as the core index, and is used for storing the inherent attributes and full history behavior data of the sample object. The index library is a structured storage module taking the nid of the sample resource as the core index, and is used for storing the basic attributes and associated entity information of the sample resource. The signal library (such as a commodity signal library and an anchor signal library) is a storage module taking the sample resource associated entity identifier (such as a commodity ID and an anchor ID) as the index, and is used for storing the resource depth statistical features and dynamic interaction signals.

[0086] Then, the feature service 503 constructs the composite key of the initial sample based on the cross features of the sample object and the sample resource in the initial sample, and then directly initiates a query request to the online storage cluster 502. In order to guarantee the freshness and consistency of the features, this query path adopts real-time penetration query, and does not use the cache of the feature service to query the related features.

[0087] Finally, the feature service 503 splices the mixed features of the initial sample, the sample object features and the sample resource features of the initial sample to obtain the training sample for the sorting model training.

[0088] In implementation, the model parameters of the sorting model can be updated by contrast learning in a supervised manner.

[0089] 3) Online prediction process: Firstly, the online prediction service 504 sends a first query request to the feature service 503 after receiving a recommendation request, so as to obtain the target object features and the candidate resource features from the feature service 503. The process of obtaining the corresponding features through the feature service is consistent with the foregoing offline training process, and will not be described herein.

[0090] Then, the online prediction service 504 constructs cross features based on the target object and the candidate resource. The second query request containing the cross features is sent to the online storage cluster 502 through the built-in client, so as to query the mixed features in the online storage cluster 502 by taking the cross features as the composite key.

[0091] This process also disables any intermediate cache to ensure that the latest information sequence is obtained.

[0092] Finally, the online prediction service 504 inputs the object features, the resource features and the sequence features into the sorting model to obtain the sorting position of the candidate resource in the resource sequence.

[0093] In summary, taking the e-commerce short video recommendation system as an example, the resource sorting method provided in the embodiment of the disclosure is expected to bring significant benefits and effect improvement in multiple dimensions to the e-commerce short video recommendation system.

[0094] Firstly, in terms of core business indicators, this feature can accurately depict the real-time interest and historical behavior patterns of users in the current video context, which is expected to directly improve the relevance and appeal of recommendations. It is expected to achieve significant growth in click-through rate and conversion rate, especially for guiding users from short video content to product detail pages, live streaming rooms or deep conversion paths of purchase behavior.

[0095] Secondly, in terms of user experience, more fine-grained dynamic understanding can greatly improve the personalization and surprise of recommendations. Users will see less generalized recommendations that are irrelevant to the current browsing context, and more content that deeply matches their immediate interests and historical interaction patterns, thereby effectively extending user dwell time and viewing depth, enhancing user stickiness and platform satisfaction.

[0096] Finally, this method enables the model to more quickly and accurately capture emerging hotspots and user interest shifts, helping to improve user experience.

[0097] Through experimental data analysis, in the video e-commerce: the user click volume increased by several percentage points under the condition that the display PV (Page View, page view) and related technical solutions were basically flat.

[0098] From the category data analysis, the resource access volume on each category increased, and in some categories of interest, it increased by more than several percentage points.

[0099] From the user stratification data analysis, the resource exposure and click volume on core users also increased significantly, indicating that the method provided in the embodiment of the disclosure can effectively provide suitable resources for target objects.

[0100] Based on the same technical concept, the embodiment of the disclosure also provides a resource sorting device 600, as shown in Figure 6 comprises: The acquisition module 601 is configured to acquire the object features of the target object, the resource features of the candidate resources, and the mixed features in response to a resource recommendation request of the target object. The mixed features are obtained based on cross-feature queries of the target object and the candidate resources. The mixed features include multiple information sequences of the target object in multiple time windows. The fusion module 602 is configured to perform fusion operations on the multiple information sequences in the multiple time windows to obtain sequence features. The sorting module 603 is configured to input the object feature, the resource feature, and the sequence feature into a sorting model to obtain a sorting position of the candidate resource in the resource sequence.

[0101] In some embodiments, the obtaining module includes: The obtaining unit is configured to send a first query request generated based on the resource recommendation request to a feature service to obtain the object feature and the resource feature from the feature service; and The constructing unit is configured to construct a cross feature based on the target object and the candidate resource. The query unit is configured to send a second query request containing the cross feature to an online storage cluster through an embedded client to query the mixed feature in the online storage cluster as a composite key.

[0102] In some embodiments, the constructing unit is specifically configured to: construct an identification pair composed of an object identification of the target object and a target information identification of the candidate resource as the cross feature; The target information identification is an information identification of shared information of the candidate resource and other resources.

[0103] In some embodiments, the plurality of time windows do not overlap in a time range, and the time span of each time window is different; the time window is closer to the current time, the time span is smaller.

[0104] In some embodiments, the information sequence of each time window includes statistical information about the target object in a context scenario represented by the cross feature; the statistical information includes at least one of: resource display times, click times, and conversion behavior numbers from the first operation to the second operation.

[0105] In some embodiments, the fusion module includes: The determining unit is configured to determine window weights of the plurality of time windows; The weighting unit is configured to weight and sum the plurality of information sequences of the plurality of time windows based on the window weights of the time windows to obtain the sequence feature.

[0106] In some embodiments, the sorting model is trained based on training samples in an offline stage.

[0107] In some embodiments, the training module is further configured to train the sorting model based on the following method: construct a composite key of the initial sample based on the cross feature of the sample object and the sample resource in the initial sample; Query the mixed features of the initial sample based on the composite key of the initial sample in the online storage cluster; The training samples are obtained by concatenating the mixed features of the initial samples, the sample object features of the initial samples, and the sample resource features. The ranking model is trained based on the training samples.

[0108] In some embodiments, a generation module is also included for generating hybrid features based on the following method: Periodically query the interaction characteristics between the target object and at least one historical resource; For each historical resource, generate an information sequence of the target object in the current time window based on the interaction features corresponding to the historical resource; In an online storage cluster, the information sequence of the target object in the current time window is associated and stored under a composite key constructed based on the cross features of the target object and historical resources, so as to update the information sequence of multiple time windows corresponding to the composite key and obtain the fused features.

[0109] The specific functions and examples of each module and submodule of the apparatus in this disclosure can be found in the relevant descriptions of the corresponding steps in the above method embodiments, and will not be repeated here.

[0110] According to embodiments of this disclosure, this disclosure also provides an electronic device, a readable storage medium, and a computer program product.

[0111] Figure 7 A schematic block diagram of an example electronic device 700 that can be used to implement embodiments of the present disclosure is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital assistants, cellular phones, smartphones, wearable devices, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely illustrative and are not intended to limit the implementation of the present disclosure described and / or claimed herein.

[0112] like Figure 7 As shown, device 700 includes a computing unit 701, which can perform various appropriate actions and processes based on a computer program stored in read-only memory (ROM) 702 or a computer program loaded into random access memory (RAM) 703 from storage unit 708. The RAM 703 may also store various programs and data required for the operation of device 700. The computing unit 701, ROM 702, and RAM 703 are interconnected via bus 704. Input / output (I / O) interface 705 is also connected to bus 704.

[0113] The plurality of components in the device 700 are connected to the I / O interface 705, including: an input unit 706, such as a keyboard, a mouse, etc.; an output unit 707, such as various types of displays, speakers, etc.; a storage unit 708, such as a magnetic disk, an optical disk, etc.; and a communication unit 709, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 709 allows the device 700 to exchange information / data with other devices through computer networks, such as the Internet, and / or various telecommunication networks.

[0114] The computing unit 701 can be various general and / or special-purpose processing components with processing and computing capabilities. Some examples of the computing unit 701 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special-purpose artificial intelligence (AI) computing chips, various computing units running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The computing unit 701 performs various methods and processes described above, such as the resource ordering method. For example, in some embodiments, the resource ordering method can be implemented as a computer software program, which is tangibly embodied in a machine-readable medium, such as the storage unit 708. In some embodiments, part or all of the computer program can be loaded and / or installed onto the device 700 via the ROM 702 and / or the communication unit 709. When the computer program is loaded onto the RAM 703 and executed by the computing unit 701, one or more steps of the resource ordering method described above can be performed. Alternatively, in other embodiments, the computing unit 701 can be configured to perform the resource ordering method by any other appropriate means, such as by means of firmware.

[0115] Various implementations of the systems and techniques described above herein can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a programmable logic device (CPLD), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0116] Program code for carrying out methods of the present disclosure can be written in any combination of one or more programming languages. The program code can be provided to a processor or controller of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the program code, when executed by the processor or controller, produces the functions / operations specified in the flowcharts and / or block diagrams. The program code can be executed entirely on a machine, partially on a machine, partially on a machine as a standalone software package, or entirely on a remote machine or server.

[0117] In the context of the present disclosure, a machine-readable medium can be a tangible medium that contains or stores a program for use by or in connection with an instruction execution system, apparatus, or device. The machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can include but is not limited to an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. More specific examples of the machine-readable storage medium will include one or more lines of electrical connections, portable computer disks, hard disk drives, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or Flash memory), optical fibers, portable compact disc read-only memories (CD-ROMs), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0118] To provide for interaction with a user, the systems and techniques described here can be implemented on a computer having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the computer. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0119] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), and the Internet.

[0120] The computer system can include clients and servers. This relationship can be. The servers are typically remote from the clients with the interactions between them occurring over a communication network. The relationship between client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The servers can be cloud servers, servers of a distributed system, or servers incorporating blockchain.

[0121] It should be understood that the various forms of flow shown above can be re-ordered, added to, or deleted from without departing from the scope of the present disclosure. For example, the steps recited in the present disclosure can be performed in parallel, in series, or in a different order, without departing from the desired results of the technology disclosed in the present disclosure, and are not limited herein.

[0122] The specific embodiments described above are not intended to be limiting, and persons skilled in the art will appreciate that various modifications, combinations, sub-combinations and alternatives can be made to the specific embodiments without departing from the principles of the present disclosure. Any further modifications, equivalents, alternatives, and / or improvements made to the specific embodiments described above are intended to fall within the scope of the present disclosure.

Claims

1. A resource ranking method, comprising: obtaining an object feature of a target object, a resource feature of a candidate resource, and a hybrid feature in response to a resource recommendation request of the target object, the hybrid feature being obtained based on a cross-feature query of the target object and the candidate resource, the hybrid feature comprising a plurality of information sequences of the target object in a plurality of time windows; performing fusion operation on the plurality of information sequences of the plurality of time windows to obtain a sequence feature; inputting the object feature, the resource feature, and the sequence feature into a ranking model to obtain a ranking position of the candidate resource in a resource sequence.

2. The method of claim 1, wherein, The obtaining of the object feature of the target object, the resource feature of the candidate resource, and the hybrid feature comprises: sending a first query request generated based on the resource recommendation request to a feature service to obtain the object feature and the resource feature from the feature service; and constructing the cross-feature based on the target object and the candidate resource; sending a second query request containing the cross-feature to an online storage cluster through an embedded client to query the hybrid feature as a composite key in the online storage cluster.

3. The method of claim 2, wherein, The constructing of the cross-feature based on the target object and the candidate resource comprises: constructing an identification pair composed of an object identification of the target object and a target information identification of the candidate resource as the cross-feature based on the object identification and the target information identification; wherein the target information identification is an information identification of shared information of the candidate resource and other resources.

4. The method of claim 1, wherein, The plurality of time windows do not overlap in time range, and the time span of each time window is different; wherein the closer the starting time is to the current time, the smaller the time span is.

5. The method of claim 1 or 4, wherein, The information sequence of each time window comprises statistical information of the target object in a context scenario represented by the cross-feature; the statistical information comprises at least one of: resource display times, click times, and conversion behavior numbers from a first operation to a second operation.

6. The method of any one of claims 1-5, wherein, The fusion operation on the plurality of information sequences of the plurality of time windows to obtain the sequence feature comprises: determining window weights of the plurality of time windows; performing weighted summation on the plurality of information sequences of the plurality of time windows based on the window weights of the time windows to obtain the sequence feature.

7. The method of claim 1, wherein, The ranking model is trained based on training samples in an offline stage.

8. The method of claim 7, further comprising training the ranking model based on the following method: constructing a composite key of the initial sample based on the cross-feature of the sample object and the sample resource in the initial sample; querying the hybrid feature of the initial sample in an online storage cluster based on the composite key of the initial sample; splicing the hybrid feature of the initial sample and the sample object feature and the sample resource feature of the initial sample to obtain a training sample; training the ranking model based on the training sample.

9. The method of claim 2, further comprising generating the hybrid feature based on the following method: periodically query interaction features of the target object and at least one historical resource; for each historical resource, generate an information sequence of the target object in a current time window based on the interaction features corresponding to the historical resource; in the online storage cluster, store the information sequence of the target object in the current time window in association with a composite key constructed based on cross features of the target object and the historical resource, to update information sequences of multiple time windows corresponding to the composite key, and obtain the fusion features.

10. A resource ranking apparatus, comprising: an acquisition module configured to, in response to a resource recommendation request of a target object, acquire object features of the target object, resource features of a candidate resource, and mixed features, the mixed features being obtained based on cross features of the target object and the candidate resource, the mixed features including multiple information sequences of the target object in multiple time windows; a fusion module configured to perform fusion operation on the multiple information sequences in the multiple time windows to obtain sequence features; a ranking module configured to input the object features, the resource features, and the sequence features into a ranking model to obtain a ranking position of the candidate resource in a resource sequence.

11. An electronic device, comprising: at least one processor; and a memory connected in communication with the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method of any one of claims 1-9.

12. A non-transitory computer readable storage medium having stored thereon computer instructions, wherein, The computer instructions are used to enable the computer to perform the method of any one of claims 1-9.

13. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1-9. The computer instructions are used to enable the computer to perform the method of any one of claims 1-9.

13. A computer program product comprising a computer program which, when executed by a processor, implements the method of any one of claims 1-9.

Images