Information recommendation method and device, electronic equipment and medium

Abstract

Embodiments of the present application provide an information recommendation method and device, equipment and medium, the method comprises: in response to the information recommendation request sent by the user, the user behavior sequence modeling and information recall are executed in parallel;The behavior sequence modeling comprises: obtaining the behavior sequence expression corresponding to the user;The information recall comprises: obtaining a plurality of candidate information features corresponding to the information recommendation request;Based on the behavior sequence expression and the plurality of candidate information features, the information corresponding to each of the plurality of candidate information features is sorted;Based on the sorting result of the information corresponding to each of the plurality of candidate information features, the recommendation information is sent to the user, and the recommendation information comprises part or all of the information corresponding to the candidate information features.

Description

Technical Field

[0001] This application relates to the field of server technology, and in particular to an information recommendation method, apparatus, electronic device, and medium. Background Technology

[0002] In recommendation scenarios, information such as store information, product information, or service information is recommended to users based on their search terms. This process is called information recommendation. Information recommendation technologies generally involve two tasks: information retrieval and information ranking. Information ranking is the process of sorting the retrieved candidate information based on information retrieval. In this stage, user behavior data is typically used to model behavioral sequences. The resulting behavioral sequence representation is then used to rank the information, ultimately determining the recommended information to be presented to the user.

[0003] In behavioral sequence modeling, the longer the behavioral sequence, the higher the accuracy of information recommendation. However, due to the limited computing resources, longer behavioral sequences also mean longer recommendation times, resulting in high time consumption and low efficiency. Therefore, to meet the timeliness requirements of information recommendation, behavioral sequence modeling generally focuses on short sequences. However, modeling short sequences, in turn, leads to lower accuracy in information recommendation.

[0004] Therefore, there is an urgent need for an information recommendation scheme that is accurate, timely, and reliable. Summary of the Invention

[0005] To address the aforementioned issues, this application provides an information recommendation method, apparatus, electronic device, and medium, which aims to improve the accuracy of information recommendation by performing long-sequence behavioral sequence modeling while ensuring the timeliness of information recommendation.

[0006] A first aspect of this disclosure provides an information recommendation method, the method comprising:

[0007] In response to a user's information recommendation request, user behavior sequence modeling and information retrieval are performed in parallel; the behavior sequence modeling includes: obtaining the behavior sequence expression corresponding to the user; the information retrieval includes: obtaining multiple candidate information features corresponding to the information recommendation request;

[0008] Based on the behavioral sequence expression and the multiple candidate information features, the information corresponding to each of the multiple candidate information features is sorted.

[0009] Based on the ranking results of the information corresponding to each of the multiple candidate information features, recommendation information is sent to the user, and the recommendation information includes some or all of the information corresponding to the candidate information features.

[0010] Optionally, the method further includes:

[0011] The training model is split into a behavior sequence modeling model and an information ranking model; wherein the training samples for training the ranking model include multiple user behavior feature samples and multiple information feature samples.

[0012] Determining the behavioral sequence representation corresponding to the user includes:

[0013] The user's behavior data is obtained, and the user behavior data is input into the behavior sequence modeling model to obtain the behavior sequence expression;

[0014] Based on the behavioral sequence representation and the multiple candidate information features, the information corresponding to each of the multiple candidate information features is sorted, including:

[0015] The behavioral sequence representation and the multiple candidate information features are input into the information ranking model to obtain the ranking result of the multiple candidate information features.

[0016] Optionally, the steps for training the ranking model include:

[0017] The multiple user behavior data samples and the multiple information feature samples are input into a preset model, which includes a modeling module and a sorting module. The modeling module is used to perform behavior sequence modeling based on the multiple user behavior data samples, and the sorting module is used to output the sorting result of the multiple information feature samples based on the behavior sequence expression after behavior sequence modeling output by the modeling module and the multiple information feature samples.

[0018] Based on the sorting results and sorting labels of the multiple information feature samples, the parameters of the preset model are updated.

[0019] The preset model after multiple updates is determined as the sorting model; wherein, the modeling module in the sorting model is the behavior sequence modeling model, and the sorting module in the sorting model is the information sorting model.

[0020] Optionally, obtaining the behavioral sequence representation corresponding to the user includes:

[0021] Obtain the user's user behavior data; the user behavior data includes the user's real-time behavior data and the user's historical behavior data;

[0022] Based on the user behavior data, behavior sequence modeling is performed to obtain the behavior sequence expression corresponding to the user.

[0023] Optionally, the user's user behavior data may be obtained, including any of the following:

[0024] Acquire user behavior data of the user performing a preset operation within a time period exceeding a preset duration;

[0025] Acquire user behavior data corresponding to multiple preset dimensions within a preset time period; the multiple preset dimensions include at least the operation behavior dimension, user information dimension, and the environment dimension.

[0026] Obtain user behavior data corresponding to the multiple preset dimensions within the time period exceeding the preset duration.

[0027] Optionally, based on the behavioral sequence representation and the plurality of candidate information features, the information corresponding to each of the plurality of candidate information features is sorted, including:

[0028] Based on the number of the multiple candidate information features, the behavior sequence expression is copied into behavior sequence expressions corresponding to the multiple candidate information features respectively;

[0029] Based on the multiple candidate information features and their corresponding behavioral sequence expressions, the information corresponding to each of the multiple candidate information features is sorted.

[0030] Optionally, obtaining multiple candidate information features corresponding to the information recommendation request includes:

[0031] Extract query keywords from the information recommendation request;

[0032] Obtain multiple preliminary information features corresponding to the query keywords;

[0033] The plurality of preliminary information features are sorted to obtain the plurality of candidate information features.

[0034] A second aspect of this application also provides an information recommendation device, the device comprising:

[0035] The parallel execution module is used to respond to information recommendation requests sent by users and perform user behavior sequence modeling and information retrieval in parallel; the behavior sequence modeling includes: obtaining the behavior sequence expression corresponding to the user; the information retrieval includes: obtaining multiple candidate information features corresponding to the information recommendation request;

[0036] The information ranking module is used to rank the information corresponding to each of the multiple candidate information features based on the behavior sequence expression and the multiple candidate information features;

[0037] The information sending module is used to send recommendation information to the user based on the ranking results of the information corresponding to each of the multiple candidate information features. The recommendation information includes some or all of the information corresponding to the candidate information features.

[0038] A third aspect of this disclosure provides an electronic device including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements the information recommendation method as described in one aspect.

[0039] Furthermore, embodiments of this application also provide a computer-readable storage medium storing a computer program that causes a processor to execute the information recommendation method as described in the first aspect.

[0040] In this application, user behavior sequence modeling and information retrieval can be performed in parallel in response to a user's information recommendation request. The behavior sequence modeling includes obtaining a behavior sequence expression corresponding to the user. The information retrieval includes obtaining multiple candidate information features corresponding to the information recommendation request. Then, based on the behavior sequence expression and the multiple candidate information features, the ranking result of the multiple candidate information features is determined. After that, based on the ranking result of the multiple candidate information features, recommendation information is sent to the user.

[0041] The technical solution adopted in this application has the following advantages:

[0042] Because the information recommendation method in this application prioritizes the behavioral sequence modeling required for information ranking, performing it in parallel with information recall, multiple candidate information features can be directly ranked based on the behavioral sequence expressions calculated in parallel with information recall during subsequent information ranking. This shortens the information ranking time and improves the efficiency of information recommendation. Furthermore, prioritizing behavioral sequence modeling in parallel with information recall allows sufficient time for behavioral sequence modeling, thus satisfying the requirements for modeling long sequences and improving the accuracy of information recommendation. Attached Figure Description

[0043] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0044] Figure 1 This is a schematic diagram illustrating the technical concept of an information recommendation method according to an embodiment of this application;

[0045] Figure 2 This is a schematic flowchart illustrating the steps of an information recommendation method according to an embodiment of this application;

[0046] Figure 3This is a conceptual diagram illustrating yet another information recommendation method according to an embodiment of this application;

[0047] Figure 4 This is a schematic diagram illustrating the process of separating a behavior sequence modeling model and an information ranking model from a ranking model, as shown in an embodiment of this application.

[0048] Figure 5 This is a flowchart illustrating the steps of training a ranking model according to an embodiment of this application;

[0049] Figure 6 This is a schematic diagram of the framework of an information recommendation device according to an embodiment of this application. Detailed Implementation

[0050] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0051] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this disclosure are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this disclosure described herein can be implemented in orders other than those illustrated or described herein. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this disclosure. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this disclosure as detailed in the appended claims.

[0052] In related technologies, information recommendation is generally carried out in the following order: full information retrieval, coarse information ranking, and fine information ranking. In the fine information ranking, user behavior sequence modeling is performed, and then the results of coarse information ranking are finely ranked based on the behavior sequence expression obtained from the modeling.

[0053] Therefore, in information ranking within related technologies, although modeling long user behavior sequences can significantly improve recommendation performance, the system's online latency and storage overhead increase linearly with the length of the user's historical behavior sequence, becoming the main bottleneck restricting long behavior sequence modeling. Thus, how to utilize limited resources to improve online computational efficiency and reduce service response time is a crucial issue in long behavior sequence modeling.

[0054] Currently, the main approaches to modeling long behavioral sequences using current technology fall into two categories:

[0055] One approach, A, is user sequence modeling based on random sampling. This approach uses sub-sequences instead of long sequences for modeling. There are two methods for selecting sub-sequences: one based on random sampling and the other based on similarity sampling. Random sampling-based user sequence modeling obtains sub-sequences by randomly sampling from the long sequence, approximating the user's long sequence. Similarity sampling-based user sequence modeling selects behaviors with high similarity to candidate information features from the long sequence as sub-sequences, also approximating the user's long sequence. However, this approach has two drawbacks. First, the sampling ratio may be too low, potentially losing a significant amount of useful information. Second, similarity sampling-based user sequence modeling, if using a complex similarity model, puts a heavy burden on online services; if using a simple similarity model, it can lead to the loss of relevant information.

[0056] One approach, B, decouples user behavior sequence computation from the main model (information recommendation model or information ranking model) to achieve modeling of ultra-long behavior sequences. This framework constructs a separate interest extraction module to store user behavior sequences, alleviating online storage pressure during long behavior sequence modeling. Updating user behavior sequences online independently depends on whether the user has new behavior, not recommendation requests. When a user generates new behavior, the existing user behavior sequence is updated based on the new behavior content. When a new recommendation request arrives, the stored user behavior sequence is directly retrieved, avoiding time-consuming online computation logic. However, in this approach, the user behavior sequence consists of a portion of historical behavior sequences and a portion of real-time user behavior. Since the main model is constantly updated, the historical user behavior sequences and current real-time behavior representations may no longer be in the same semantic space as the main model, leading to reduced recommendation accuracy. Furthermore, when iterating or updating the main model, the user interest representations stored online need to be fully updated synchronously, resulting in high maintenance and iteration costs.

[0057] In view of this, this application proposes the following technical concept: adopting a recommendation method that prioritizes user behavior sequence calculation, the user behavior sequence modeling is performed in advance when a new information recommendation request is received, and is carried out simultaneously with information recall. This provides sufficient time for the calculation of user behavior sequences and lays the foundation for the complex calculation of modeling ultra-long user behavior sequences, thereby improving the accuracy of information recommendation while ensuring the timeliness of information recommendation.

[0058] Meanwhile, the technical concept of this application is based on the information recommendation request sent by the user. When the information recommendation request sent by the user is received, user behavior sequence modeling is started. Compared with the above method A, all behavior data can be used to perform long behavior sequence modeling, thereby avoiding the problem of information loss and improving the accuracy of information recommendation.

[0059] The information recommendation method proposed in this application can be applied to store recommendations as well as product recommendation scenarios.

[0060] Reference Figure 1 The diagram illustrates a schematic representation of the technical concept of an information recommendation method according to an embodiment of this application. Figure 1 As shown, an information recommendation process includes four stages: user behavior sequence modeling, information retrieval, information ranking, and information delivery. Upon receiving a user's information recommendation request, user behavior sequence modeling and information retrieval can be performed in parallel. Information retrieval includes two tasks: full information retrieval and coarse ranking. Then, in the information ranking stage, based on the calculated behavior sequence representation and the coarse ranking results, the coarse ranking results are refined, and finally, the refined information is sent to the user.

[0061] like Figure 1 As shown, by moving the behavioral sequence modeling, which originally belonged to the information ranking process, to be executed in parallel with the information retrieval task, sufficient time is gained for behavioral sequence modeling. This not only reduces the overall time of information recommendation but also allows for long behavioral sequence modeling, thereby ensuring both the timeliness and accuracy of information recommendation.

[0062] Combination Figure 1 As shown, the information recommendation method of this application is introduced, with reference to... Figure 2 As shown, a flowchart illustrating the steps of the information recommendation method of this application is presented, as follows: Figure 2 As shown, it includes the following steps:

[0063] Step S201: In response to the information recommendation request sent by the user, perform user behavior sequence modeling and information retrieval in parallel.

[0064] The behavior sequence modeling includes: obtaining the behavior sequence expression corresponding to the user; the information retrieval includes: obtaining multiple candidate information features corresponding to the information recommendation request.

[0065] In this embodiment, the application is in an information recommendation scenario. In this scenario, the information recommendation request sent by the user may include the query term entered by the user. The query term represents the information that the user expects to search for. For example, if the user enters the query term "hot pot", the system will recommend hot pot restaurant information to the user.

[0066] like Figure 1 As shown, information retrieval and user behavior sequence modeling can be performed in parallel when a user sends an information recommendation request. Parallel execution can be understood as simultaneous execution, meaning that two threads are simultaneously enabled in the computer to perform information retrieval and user behavior sequence modeling at the same time.

[0067] User behavior sequence modeling is related to a user's historical behavior and specifically includes obtaining a user's corresponding behavior sequence representation based on user behavior data. User behavior data refers to data generated by a user performing a certain operation, typically recorded by the server after obtaining permissions granted by the user. In this embodiment, operations may include clicks, searches, and favorites, etc. This behavior data reflects the traces of information queries the user has made in the past. The behavior sequence representation can be understood as data obtained by serializing and then encoding user behavior data. This behavior sequence representation can reflect the user's behavioral characteristics over a period of time from a higher-level semantic perspective, thus characterizing the user's information query preferences in the past.

[0068] Information retrieval can be related to the currently generated information recommendation request, specifically to the query term currently used by the user. This includes retrieving information containing the query term and its synonyms or near-synonyms based on the query term in the recommendation request. Furthermore, the features corresponding to this retrieved information are used as candidate information features, which can be understood as vector representations of the names of the retrieved information.

[0069] For example, if the query term is "hot pot", the retrieved information can be store information whose names include "hot pot", or store information whose names include similar terms such as "soup pot" or "skewers".

[0070] Step S202: Based on the behavioral sequence expression and the multiple candidate information features, sort the information corresponding to each of the multiple candidate information features.

[0071] In this embodiment, since the behavioral sequence expression can characterize the user's preferences for information queries in the historical process, and multiple candidate information features are obtained based on the query terms in the information recommendation request, the degree to which multiple candidate information features are of interest to the user can be determined based on the behavioral sequence expression. Therefore, the information corresponding to multiple candidate information features can be sorted based on the degree to which multiple candidate information features are of interest to the user, thereby obtaining the sorting result.

[0072] In practice, multiple candidate information features can be sorted in descending order of their level of interest to the user.

[0073] Step S203: Based on the ranking results of the information corresponding to each of the multiple candidate information features, send recommendation information to the user, wherein the recommendation information includes some or all of the information corresponding to the candidate information features.

[0074] Each candidate information feature corresponds to one piece of information, and multiple candidate information features correspond to multiple pieces of information. After obtaining the ranking results of multiple candidate information features, a preset number of candidate information features with the highest ranking can be selected from the multiple candidate information features, and the information corresponding to each of the preset number of candidate information features is sent to the user as recommendation information.

[0075] It should be noted that the information in the embodiments of this application can be retrieved based on query terms, and can be multimedia information, store information, product information, etc.

[0076] By adopting the technical solution of this application embodiment, the behavioral sequence modeling that was originally performed in information sorting is moved to be performed in parallel with information recall. In this way, the behavioral sequence expression can start to be calculated from the time of information recall. This provides sufficient time for the behavioral sequence expression, so that the already calculated behavioral sequence expression can be directly used when sorting information later, instead of waiting until the information recall is completed before starting the calculation of the behavioral sequence expression.

[0077] In this way, on the one hand, the time for information ranking is shortened, improving the overall efficiency of information recommendation. On the other hand, by prioritizing behavioral sequence modeling and performing it in parallel with information retrieval, sufficient time is gained for behavioral sequence modeling, allowing for the modeling of long-term behavioral sequences. This enables subsequent information ranking to be based on long-term behavioral sequence expressions, thereby improving the accuracy of information recommendation.

[0078] Combination Figure 1 As shown, the tasks of the three stages—user behavior sequence modeling, information retrieval, and information ranking—are introduced respectively:

[0079] In the information recall phase, such as Figure 1 As shown, a full recall can be performed first, followed by a coarse ranking of the recalled candidate information to obtain multiple candidate information features. Specifically, query keywords can be extracted from the information recommendation request, and multiple preliminary information features corresponding to the query keywords can be obtained; then, the multiple preliminary information features are ranked to obtain the multiple candidate information features.

[0080] In this embodiment, the information recommendation request carries query text, which can refer to the text entered by the user when initiating the information recommendation request. For example, if the user enters the search text "Sichuan hot pot!", then the query keyword can refer to the remaining text after removing meaningless characters from the query text. Meaningless characters can include strings in the query text, such as "!", or auxiliary words such as "at" or "on".

[0081] In this process, obtaining multiple preliminary information features corresponding to the query keywords can be a form of full recall. For example, the query keywords can be segmented to obtain multiple segments, and for each segment, information including that segment and its synonyms or similar words can be obtained. In this way, relatively complete information can be recalled, and the recalled information is called preliminary information. Then, the information features of the preliminary information can be obtained, which can refer to the vector representation of the information.

[0082] After obtaining multiple preliminary information features, the query keywords can be converted into sentence vectors. The similarity between these sentence vectors and each preliminary information feature is then determined. Information features with similarities higher than a preset similarity threshold are selected as candidate information features. In this way, the obtained candidate information features carry the similarity between themselves and the sentence vectors of the query keywords. The ranking of these candidate information features can then be determined based on the similarity. Naturally, higher similarity indicates a closer match between the information and the query keywords, and potentially better meets the user's needs; therefore, the higher the ranking. This process can be called the coarse sorting of information.

[0083] In this implementation method, the initial information feature retrieval is performed based on word segmentation of the query keywords, which can retrieve relatively comprehensive initial information. Then, candidate information features are selected from multiple initial information features based on the similarity between the sentence vector of the query keywords and the initial information features. Thus, the initial information features can be filtered and ranked based on similarity, so that the coarsely ranked candidate information features match the query keywords, thereby improving the accuracy and comprehensiveness of information recommendation.

[0084] In the user behavior sequence modeling stage, user behavior sequence modeling can be performed based on user behavior data. Through user behavior sequence modeling, users' information search preferences and intentions can be obtained. Then, multiple candidate information features can be sorted and filtered based on users' information search preferences and intentions.

[0085] In practice, user behavior data of the user can be obtained, and user behavior sequence modeling can be performed based on the user behavior data to obtain the behavior sequence expression corresponding to the user.

[0086] The user behavior data includes the user's real-time behavior data and the user's historical behavior data.

[0087] In this embodiment, both historical behavior data and real-time behavior data can represent the data recorded when a user performs preset operations during the period from entering to leaving the website. Thus, it can be understood that a user typically enters the website multiple times throughout the historical process, resulting in multiple records of user behavior data.

[0088] Typically, a single user behavior data entry can include multiple operation records, each of which can include the operation time and a description of the operation. For example, for operation record A, the operation time is 2013-09-05-11:03, and the operation description is "search"—a search operation was performed at 11:03 on the date 2013-09-05.

[0089] In practice, the process of modeling user behavior sequences can be as follows:

[0090] First, for each piece of user behavior data, multiple operation records can be serialized according to the operation time in the user behavior data and in chronological order of operation time. For example, different codes can be set for different operation behavior descriptions. In this way, the user behavior sequence can be obtained.

[0091] Next, since there can be multiple user behavior data sets, the user behavior sequences of each set can be concatenated to obtain a representation of the user's behavior sequence.

[0092] Of course, the above process can be carried out through models. Specifically, user behavior sequence modeling can be performed using neural network models such as RNN (Recurrent Neural Network), LSTM (Long Short-Term Memory), and GRU (Gate Recurrent Unit).

[0093] In one specific implementation, when acquiring user behavior data, user behavior data over a relatively long period of time can be acquired, thus enabling long-term user behavior sequence modeling; alternatively, user behavior data across various dimensions over a period of time can be acquired, thereby improving the richness of user behavior data in representing user behavior and enabling extremely rich user behavior sequence modeling.

[0094] In specific implementation, one approach is to acquire user behavior data of the user performing preset operations within a time period exceeding a preset duration. The user behavior data obtained in this way can be referred to as long-term user behavior data.

[0095] Another approach is to obtain user behavior data corresponding to multiple preset dimensions within a preset time period; the multiple preset dimensions include at least the operation behavior dimension, user information dimension, and the environment dimension. The user behavior data obtained in this way can be called multi-dimensional user behavior data.

[0096] Another approach is to obtain user behavior data corresponding to the multiple preset dimensions within the time period exceeding the preset duration. The user behavior data obtained in this way can be called long-term multi-dimensional user behavior data.

[0097] In this embodiment, long-term user behavior data can be behavioral data of users performing preset operations within a period exceeding a preset duration. The preset duration can be pre-set, for example, set to 2 months, or longer. In this case, long-term behavior data refers to data of users performing preset operations within 3 months. Preset operations can be search operations, click operations, order operations, like operations, etc.

[0098] In this case, as mentioned above, a single user behavior data entry can include multiple operation records, and each operation record can include the operation time and a description of the operation behavior.

[0099] Multi-dimensional user behavior data can be user behavior data across multiple preset dimensions within a preset time period. The preset time period can be pre-set, for example, within a month prior to the current time. This would acquire user data across multiple preset dimensions within that month. The preset dimensions include at least the operation behavior dimension, user information dimension, and environment dimension. The operation behavior dimension refers to the dimension of the user's performed actions; the data in this dimension can be the aforementioned user behavior data including multiple operation records. User information dimension behavior data can include the user's account information, registration information, and user identity information, used to describe the user's personal characteristics. Environment dimension behavior data can describe the user's location when performing each preset operation behavior, the weather conditions at that location, etc.

[0100] In this scenario, user behavior data can include information from behavioral, personal, and environmental dimensions. This allows for the modeling of user behavior data to obtain behavioral sequence representations that more accurately describe the information a user is interested in during an information search. Of course, this behavioral data can be obtained with the user's authorization.

[0101] In this embodiment, long-term multi-dimensional user behavior data can refer to behavior data on multiple preset dimensions within a period of time exceeding a preset duration. Compared to long-term user behavior data and multi-dimensional user behavior data, the user behavior data obtained has a longer time span and more dimensions.

[0102] Accordingly, when performing behavior sequence modeling based on the user behavior data to obtain the behavior sequence expression corresponding to the user, behavior sequence modeling can be performed based on long-term user behavior data, multi-dimensional behavior data, or long-term multi-dimensional behavior data to obtain a longer behavior sequence expression.

[0103] Generally, when acquiring user behavior data using the technical solution of this application embodiment, it is possible to model ultra-long behavior sequences. For example, a behavior sequence expression with a length of 1000 can be modeled. This increased length of the behavior sequence expression can significantly improve the accuracy of information recommendation.

[0104] Since the obtained behavioral sequence expression is a set, while the information recall stage will obtain multiple candidate information features, the behavioral sequence expression can be copied in the information ranking stage so that the number of behavioral sequence expressions obtained after copying is consistent with the number of candidate information features.

[0105] In specific implementation, the behavior sequence expression can be copied into behavior sequence expressions corresponding to the multiple candidate information features based on the number of the multiple candidate information features; and the information corresponding to each of the multiple candidate information features can be sorted based on the multiple candidate information features and their respective corresponding behavior sequence expressions.

[0106] In this embodiment, the behavior sequence representation can be copied to obtain multiple copied behavior sequence representations. The number of copied behavior sequence representations is consistent with the number of candidate information features. Thus, each candidate information feature has a corresponding behavior sequence representation, achieving a one-to-one correspondence between candidate information features and behavior sequence representations. Since the behavior sequence representation is copied, different candidate information features correspond to the same behavior sequence representation.

[0107] In this way, based on multiple candidate information features and their corresponding behavioral sequence expressions, the degree of matching between each candidate information feature and the behavioral sequence expression can be determined. The higher the degree of matching, the higher the degree of adaptation between the information corresponding to the candidate information feature and the user. Thus, multiple candidate information features can be sorted.

[0108] By adopting the technical solution of this application embodiment, when performing information ranking, the behavior sequence expression can be directly copied into the behavior sequence expression corresponding to the candidate information features. Therefore, a single information recommendation request only calculates the behavior sequence expression once, which greatly reduces the amount of computation and saves online computing resources.

[0109] Reference Figure 3 As shown, a conceptual diagram of another information recommendation method according to this application is illustrated. Figure 3As shown, a ranking model can be trained in advance using multiple user behavior feature samples and multiple information feature samples. Then, the ranking model is segmented into a behavior sequence modeling model for user behavior sequence modeling and an information ranking model for information ranking.

[0110] like Figure 3 The behavior sequence modeling model can be used to respond to the user's information recommendation request and calculate the behavior sequence expression. The output of the behavior sequence modeling model is connected to the input of the information ranking model, so that the behavior sequence expression is input into the information ranking model. The input of the information ranking model is also used to receive multiple candidate information features obtained from information retrieval. Thus, the information ranking model can perform fine ranking of multiple candidate information features based on the behavior sequence expression and multiple candidate information features.

[0111] In this approach, the ranking model can be trained to closely link behavior sequence modeling and information ranking. In this way, behavior sequence modeling and information ranking can be automatically updated and iterated as the ranking model is updated and iterated, thus ensuring that the behavior sequence expression obtained by behavior sequence modeling and the candidate information features are always in the same semantic space. Therefore, the accuracy of information recommendation can be guaranteed, thereby avoiding the problems of semantic space inconsistency, low information recommendation accuracy and high maintenance cost caused by the decoupling of behavior sequence modeling and information ranking in information recommendation in the above method B.

[0112] Reference Figure 4 The diagram illustrates the process of separating the behavioral sequence modeling model and the information ranking model from the ranking model. Figure 4 As shown, in practical implementation, it is possible to first train and obtain, as Figure 4 The topmost sorting model, refer to Figure 5 As shown, a flowchart illustrating the steps involved in training the ranking model is presented. Figure 5 As shown, the specific steps may include:

[0113] Step S501: Input the multiple user behavior data samples and the multiple information feature samples into a preset model, the preset model including a modeling module and a sorting module.

[0114] The modeling module is used to perform behavior sequence modeling based on the multiple user behavior data samples, and the sorting module is used to output the sorting result of the multiple information feature samples based on the behavior sequence expression after behavior sequence modeling output by the modeling module and the multiple information feature samples.

[0115] In this embodiment, user behavior data samples can refer to the behavior data of multiple users performing preset operations recorded online. Each user's user behavior data sample and multiple information feature samples form a set of training samples, so that different users can correspond to different training samples.

[0116] During training, multiple sets of training samples can be input into a preset model for training. The process of obtaining user behavior data samples can be as described in the previous embodiment, and will not be repeated here. The information feature samples can carry ranking labels, which are used to characterize the true ranking of the information corresponding to each of the multiple information feature samples.

[0117] The output of the modeling module is connected to the input of the sorting module, such as... Figure 4 As shown, the modeling module is Figure 4 As shown in section 401, the sorting module is... Figure 4 As shown in section 402. Multiple user behavior data samples can be input into the modeling module's input terminal, while multiple information feature samples can be input into the sorting module's input terminal.

[0118] The modeling module is used to perform behavior sequence modeling on the multiple user behavior data samples according to the user behavior sequence modeling process described above. Then, the behavior sequence expression after behavior sequence modeling is input to the input end of the sorting module. The input end of the sorting module is also input with multiple information feature samples. Thus, the sorting module is used to output the sorting result of multiple information feature samples based on the behavior sequence expression after behavior sequence modeling output by the modeling module and the multiple information feature samples.

[0119] Step S502: Update the parameters of the preset model based on the sorting results and sorting labels of the multiple information feature samples.

[0120] In this embodiment, a loss function can be constructed based on the ranking results and ranking labels of multiple information feature samples, and the loss value can be obtained. The parameters of the preset model can then be updated based on the loss value.

[0121] The loss value characterizes the difference between the ranking result of multiple information feature samples output by the ranking model and the actual ranking. When updating the preset model based on the loss value, the parameters of both the modeling module and the ranking module can be updated. In this way, the modeling module and the ranking module can be trained and updated together with the training of the preset model, and they always exist in the same semantic space.

[0122] Step S503: Determine the preset model after multiple updates as the sorting model.

[0123] Wherein, the modeling module in the sorting model serves as the behavior sequence modeling model, and the sorting module in the sorting model serves as the information sorting model.

[0124] In this embodiment, after iteratively updating a preset number of times, or until the preset model reaches convergence, the preset model after multiple updates is determined as the ranking model. At this time, it can be sorted according to... Figure 4 As shown, the modeling module and the sorting module are separated at the connection point. The modeling module is separated as a behavior sequence modeling model, while the sorting module is separated as an information sorting model.

[0125] Accordingly, such as Figure 3 As shown, when performing user behavior sequence modeling, user behavior data of the user can be obtained and input into the behavior sequence modeling model to obtain the behavior sequence expression.

[0126] When performing information ranking, the behavioral sequence expression output by the behavioral sequence modeling model and the multiple candidate information features can be input into the information ranking model to obtain the ranking result of the multiple candidate information features.

[0127] In this embodiment, as Figure 4 As shown, the decomposed behavioral sequence modeling model is as follows: Figure 4 As shown in the lower left corner, the input to the behavior sequence modeling model is user behavior data, which can be a vector representation of the user behavior data in practice, and the output is a behavior sequence expression. The input to the information ranking model is the behavior sequence vector expression output by the behavior sequence modeling model, and candidate information features other than the behavior sequence expression. In practice, the information ranking model can concatenate each candidate information feature and the behavior sequence expression together, and then pass them through the main network to obtain a refined ranking score.

[0128] Depend on Figure 4 It can be seen that by separating the modeling module and the ranking module, user behavior sequence modeling and fine ranking scoring are split into two tasks. The user behavior sequence modeling task can be performed in parallel with information retrieval, so that user behavior sequence modeling and information retrieval can be executed in parallel when an information recommendation request is received. In the subsequent fine ranking scoring, the already modeled behavior sequence can be used, thereby improving the overall efficiency of information recommendation.

[0129] Of course, when using the behavior sequence modeling model to calculate the behavior sequence expression, the input user behavior data may include the user's real-time behavior data and the user's historical behavior data. Moreover, the user behavior data input into the behavior sequence modeling model may be the long-term behavior data and / or the multi-dimensional behavior data mentioned above. Thus, the behavior sequence modeling model can perform long behavior sequence modeling.

[0130] Of course, in this case, the user behavior feature samples in the training samples of the ranking model can be long-term behavior data samples and / or the multi-dimensional behavior data samples obtained in the above manner.

[0131] In the process of using the information ranking model for fine information ranking, the behavioral sequence expression output by the behavioral sequence model can be copied first to create multiple behavioral sequence expressions with the same number of candidate information features. Then, the multiple behavioral sequence expressions and multiple candidate information features are input into the information ranking model to obtain the ranking result of multiple candidate information features.

[0132] The technical solution adopted in this application has the following advantages:

[0133] Firstly, it significantly improves computational efficiency.

[0134] By prioritizing the computation of behavioral sequence representations, the parallelism of the entire information recommendation process is improved, allowing ample time to calculate users' long-term interests. Furthermore, behavioral sequence representations are only computed once per request, reducing online computation and saving computing resources.

[0135] Secondly, it helps improve the accuracy of information recommendations.

[0136] Because the computation of behavioral sequence representation is performed beforehand, the response time to information recommendation requests is reduced. Online resources are no longer the main bottleneck restricting the iteration of user sequence models. Long sequence modeling and complex sequence modeling can be performed, which helps to better characterize users' long-term interests and improve the accuracy of information recommendations.

[0137] Based on the same inventive concept as the above embodiments, a second aspect of the present disclosure provides an information recommendation device, referring to... Figure 6 As shown, the device may specifically include the following modules:

[0138] The parallel execution module 601 is used to respond to the information recommendation request sent by the user and perform user behavior sequence modeling and information retrieval in parallel; the behavior sequence modeling includes: obtaining the behavior sequence expression corresponding to the user; the information retrieval includes: obtaining multiple candidate information features corresponding to the information recommendation request;

[0139] The information ranking module 602 is used to rank the information corresponding to each of the multiple candidate information features based on the behavior sequence expression and the multiple candidate information features;

[0140] The information sending module 603 is used to send recommendation information to the user based on the ranking results of the information corresponding to each of the multiple candidate information features, wherein the recommendation information includes some or all of the information corresponding to the candidate information features.

[0141] Optionally, the device further includes:

[0142] The model splitting module is used to split the behavior sequence modeling model and the information ranking model from the trained ranking model; wherein, the training samples for training the ranking model include multiple user behavior feature samples and multiple information feature samples.

[0143] The parallel execution module is specifically used to acquire the user's user behavior data and input the user behavior data into the behavior sequence modeling model to obtain the behavior sequence expression.

[0144] The information ranking module 602 is specifically used to input the behavior sequence expression and the multiple candidate information features into the information ranking model to obtain the ranking result of the multiple candidate information features.

[0145] Optionally, the apparatus further includes a training module for training the ranking model and performing the following steps:

[0146] The multiple user behavior data samples and the multiple information feature samples are input into a preset model, which includes a modeling module and a sorting module. The modeling module is used to perform behavior sequence modeling based on the multiple user behavior data samples, and the sorting module is used to output the sorting result of the multiple information feature samples based on the behavior sequence expression after behavior sequence modeling output by the modeling module and the multiple information feature samples.

[0147] Based on the sorting results and sorting labels of the multiple information feature samples, the parameters of the preset model are updated.

[0148] The preset model after multiple updates is determined as the sorting model; wherein, the modeling module in the sorting model is the behavior sequence modeling model, and the sorting module in the sorting model is the information sorting model.

[0149] Optionally, the parallel execution module 601 includes a behavior sequence modeling unit, which includes:

[0150] The data acquisition subunit is used to acquire the user's user behavior data; the user behavior data includes the user's real-time behavior data and the user's historical behavior data;

[0151] The modeling subunit is used to perform behavior sequence modeling based on the user behavior data to obtain the behavior sequence expression corresponding to the user.

[0152] Optionally, the user's user behavior data may be obtained, including any of the following:

[0153] Acquire user behavior data of the user performing a preset operation within a time period exceeding a preset duration;

[0154] Acquire user behavior data corresponding to multiple preset dimensions within a preset time period; the multiple preset dimensions include at least the operation behavior dimension, user information dimension, and the environment dimension.

[0155] Obtain user behavior data corresponding to the multiple preset dimensions within the time period exceeding the preset duration.

[0156] Optionally, the information sorting module 602 includes:

[0157] The copying unit is used to copy the behavior sequence expression into behavior sequence expressions corresponding to the multiple candidate information features, based on the number of the multiple candidate information features.

[0158] The sorting unit is used to sort the information corresponding to each of the multiple candidate information features based on the multiple candidate information features and their respective corresponding behavioral sequence expressions.

[0159] Optionally, the parallel execution module 601 includes a recall unit, which is used to obtain multiple candidate information features corresponding to the information recommendation request, specifically including:

[0160] The keyword extraction subunit is used to extract query keywords from the information recommendation request;

[0161] The full recall subunit is used to obtain multiple preliminary information features corresponding to the query keywords;

[0162] The coarse sorting subunit is used to sort the multiple preliminary information features to obtain the multiple candidate information features.

[0163] It should be noted that the device embodiments are similar to the method embodiments, so the description is relatively simple. For relevant details, please refer to the method embodiments.

[0164] The information recommendation device provided in this application can be located on a server. Because the behavioral sequence modeling required for information ranking is performed concurrently with information recall, multiple candidate information features can be directly ranked based on the behavioral sequence expressions calculated in parallel with information recall during subsequent information ranking. This shortens the information ranking time and improves information recommendation efficiency. Furthermore, by performing behavioral sequence modeling concurrently with information recall, sufficient time is allocated for behavioral sequence modeling, thus satisfying the requirements for modeling long sequences and improving the accuracy of information recommendation.

[0165] This invention also provides an electronic device that may include a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor is configured to perform the information recommendation method described above.

[0166] This application also provides a non-transitory computer-readable storage medium, wherein when the instructions in the storage medium are executed by a processor, the processor is able to perform an operation to implement the information recommendation method described above in this application.

[0167] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0168] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, apparatus, or computer program products. Therefore, embodiments of the present invention can take the form of entirely hardware embodiments, entirely software embodiments, or embodiments combining software and hardware aspects. Furthermore, embodiments of the present invention can take the form of computer program products implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0169] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, terminal devices (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing terminal device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing terminal device, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0170] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing terminal device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0171] These computer program instructions can also be loaded onto a computer or other programmable data processing terminal equipment, causing a series of operational steps to be performed on the computer or other programmable terminal equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable terminal equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0172] Although preferred embodiments of the present invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of the embodiments of the present invention.

[0173] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal device. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or terminal device that includes said element.

[0174] The above provides a detailed description of the information recommendation method, apparatus, electronic device, and medium provided by the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. An information recommendation method, characterized in that, The method includes: In response to a user's information recommendation request, user behavior sequence modeling and information retrieval are performed in parallel; the behavior sequence modeling includes: obtaining the behavior sequence expression corresponding to the user; the information retrieval includes: obtaining multiple candidate information features corresponding to the information recommendation request; Based on the behavioral sequence expression and the multiple candidate information features, the information corresponding to each of the multiple candidate information features is sorted. Based on the ranking results of the information corresponding to each of the multiple candidate information features, recommendation information is sent to the user, the recommendation information including some or all of the information corresponding to the candidate information features; the method further includes: The training model is split into a behavior sequence modeling model and an information ranking model; wherein the training samples for training the ranking model include multiple user behavior feature samples and multiple information feature samples. The step of obtaining the behavioral sequence representation corresponding to the user includes: The user's behavior data is obtained, and the user behavior data is input into the behavior sequence modeling model to obtain the behavior sequence expression; Based on the behavioral sequence representation and the multiple candidate information features, the information corresponding to each of the multiple candidate information features is sorted, including: The behavioral sequence representation and the multiple candidate information features are input into the information ranking model to obtain the ranking result of the multiple candidate information features.

2. The method according to claim 1, characterized in that, The steps for training the ranking model include: The multiple user behavior data samples and the multiple information feature samples are input into a preset model, which includes a modeling module and a sorting module. The modeling module is used to perform behavior sequence modeling based on the multiple user behavior data samples, and the sorting module is used to output the sorting result of the multiple information feature samples based on the behavior sequence expression after behavior sequence modeling output by the modeling module and the multiple information feature samples. Based on the sorting results and sorting labels of the multiple information feature samples, the parameters of the preset model are updated. The preset model after multiple updates is determined as the sorting model; wherein, the modeling module in the sorting model is the behavior sequence modeling model, and the sorting module in the sorting model is the information sorting model.

3. According to claim 1 2. The method described in any one of the following methods, characterized in that, The step of obtaining the behavioral sequence representation corresponding to the user includes: Obtain the user's user behavior data; the user behavior data includes the user's real-time behavior data and the user's historical behavior data; Based on the user behavior data, behavior sequence modeling is performed to obtain the behavior sequence expression corresponding to the user.

4. The method according to claim 3, characterized in that, Obtaining the user's user behavior data includes any of the following: Acquire user behavior data of the user performing a preset operation within a time period exceeding a preset duration; Acquire user behavior data corresponding to multiple preset dimensions within a preset time period; The multiple preset dimensions include at least the operational behavior dimension, the user information dimension, and the environment dimension. Obtain user behavior data corresponding to the multiple preset dimensions within the time period exceeding the preset duration.

5. According to claim 1 2. The method described in any one of the following methods, characterized in that, Based on the behavioral sequence representation and the multiple candidate information features, the information corresponding to each of the multiple candidate information features is sorted, including: Based on the number of the multiple candidate information features, the behavior sequence expression is copied into behavior sequence expressions corresponding to the multiple candidate information features respectively; Based on the multiple candidate information features and their corresponding behavioral sequence expressions, the information corresponding to each of the multiple candidate information features is sorted.

6. According to claim 1 2. The method described in any one of the following methods, characterized in that, The acquisition of multiple candidate information features corresponding to the information recommendation request includes: Extract query keywords from the information recommendation request; Obtain multiple preliminary information features corresponding to the query keywords; The plurality of preliminary information features are sorted to obtain the plurality of candidate information features.

7. An information recommendation device, characterized in that, The device includes: The parallel execution module is used to respond to information recommendation requests sent by users and perform user behavior sequence modeling and information retrieval in parallel; the behavior sequence modeling includes: obtaining the behavior sequence expression corresponding to the user; the information retrieval includes: obtaining multiple candidate information features corresponding to the information recommendation request; The information ranking module is used to rank the information corresponding to each of the multiple candidate information features based on the behavior sequence expression and the multiple candidate information features; The information sending module is used to send recommendation information to the user based on the ranking results of the information corresponding to each of the multiple candidate information features, wherein the recommendation information includes some or all of the information corresponding to the candidate information features; The device further includes: The training model is split into a behavior sequence modeling model and an information ranking model; wherein the training samples for training the ranking model include multiple user behavior feature samples and multiple information feature samples. The step of obtaining the behavioral sequence representation corresponding to the user includes: The user's behavior data is obtained, and the user behavior data is input into the behavior sequence modeling model to obtain the behavior sequence expression; Based on the behavioral sequence representation and the multiple candidate information features, the information corresponding to each of the multiple candidate information features is sorted, including: The behavioral sequence representation and the multiple candidate information features are input into the information ranking model to obtain the ranking result of the multiple candidate information features.

8. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executed, implements claim 1. The information recommendation method described in any one of the six claims.

9. A computer-readable storage medium, characterized in that, Its stored computer program causes the processor to execute as claimed in claim 1. The information recommendation method described in any one of the six claims.

Images