A method, apparatus, device, and storage medium for content retrieval and recommendation.

Abstract

This invention provides a content retrieval and recommendation method, apparatus, device, and storage medium, applied to a server in a content recommendation system, and relating to the field of data processing technology. The specific implementation scheme is as follows: In response to a content recommendation instruction sent by a client, historical interaction data of the target account logged into the client is obtained; candidate indices are determined based on the correlation between the historical interaction data and the initial indexes of candidate content; wherein the similarity of the content associated with each correlation is greater than a preset similarity threshold; the content corresponding to the candidate index is used as the content retrieval result. Applying the scheme provided by this invention can reduce the computational load required for content retrieval while ensuring a sufficient number of retrieved contents.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method, apparatus, device, and storage medium for content retrieval and recommendation. Background Technology

[0002] Audio and video platforms and other information platforms can provide users with a variety of content. To increase user stickiness on the platform, the platform may recommend content to users so that users can quickly find the content they like from the platform's recommendations.

[0003] The process of initially selecting content from the information platform's content to recommend to users is called content recall.

[0004] In existing technologies, after content recall, the recalled content is typically filtered to identify content that does not conform to the user's browsing habits and exclude it from the subsequent content recommendation process. However, the number of filtered contents is uncertain, resulting in an uncertain number of contents recommended to the user. Furthermore, if a large amount of recalled content is filtered, the actual number of contents that can be recommended to the user may be insufficient.

[0005] Furthermore, since the filtered content is not used in subsequent recommendation processes, the computational resources consumed in the content recall process for this filtered content are wasted, resulting in excessive computational resources being used for content recall. Summary of the Invention

[0006] The purpose of this invention is to provide a method, apparatus, device, and storage medium for content retrieval and recommendation, reducing the computational load required for content retrieval while ensuring a sufficient quantity of retrieved content. The specific technical solution is as follows:

[0007] In a first aspect of the present invention, a content retrieval method is provided, the method being applied to a server in a content recommendation system, comprising:

[0008] In response to a content recommendation instruction sent by the client, historical interaction data of the target account logged into the client is obtained;

[0009] Candidate indices are determined based on the relationships between the historical interaction data and the initial indices of the candidate content; wherein the similarity of the content associated with each relationship is greater than a preset similarity threshold.

[0010] The content corresponding to the candidate index is used as the content retrieval result.

[0011] In a second aspect of this invention, a content recommendation method is also provided, the method being applied to a server in a content recommendation system, comprising:

[0012] The above content recall method is used to recall candidate content and obtain the recalled content;

[0013] Based on the recalled content, determine the content to be recommended;

[0014] The content to be recommended is sent to the client so that the client can display the content to be recommended.

[0015] In a third aspect of the invention, a content recall device is also provided, the device being applied to a server in a content recommendation system, comprising:

[0016] The interaction data acquisition module is used to acquire historical interaction data of the target account logged into the client in response to the content recommendation instruction sent by the client.

[0017] The candidate index determination module is used to determine candidate indexes based on the correlation between the historical interaction data and the initial indexes of the candidate content; wherein the similarity of the content associated with each correlation is greater than a preset similarity threshold.

[0018] The recall result determination module is used to take the content corresponding to the candidate index as the content recall result.

[0019] In a fourth aspect of the invention, a content recommendation apparatus is also provided, the apparatus being applied to a server in a content recommendation system, comprising:

[0020] The recall content acquisition module is used by the aforementioned content recall device to recall candidate content and obtain the recall content.

[0021] The module for determining content to be recommended is used to determine content to be recommended based on the recalled content.

[0022] The module for sending content to be recommended is used to send the content to be recommended to the client so that the client can display the content to be recommended.

[0023] In a fifth aspect of the present invention, an electronic device is also provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0024] Memory, used to store computer programs;

[0025] The processor, when executing a program stored in memory, implements the above-mentioned content retrieval or content recommendation method steps.

[0026] In a sixth aspect of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, the above-described content recall or content recommendation method is implemented.

[0027] In a seventh aspect of the invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform the aforementioned content recall or content recommendation method.

[0028] In the solution provided by this invention, candidate indices are determined from the initial index based on historical interaction data reflecting user behavior during the recall process. This ensures that the content corresponding to the candidate indices conforms to user browsing habits. Thus, the content corresponding to the candidate indices is used as the content recall result, and the recalled content is consistent with user browsing habits. Therefore, filtering based on user behavior is unnecessary after recall, effectively implementing pre-filtering during the recall process. In this case, the number of recalled content can be controlled during the content recall process. Specifically, if the number of determined candidate indices is insufficient, a sufficient number of new candidate indices are searched from the initial index that is related to the determined candidate indices, thereby ensuring a sufficient number of recalled content.

[0029] Under this premise, on the one hand, since content that does not conform to the user's browsing habits has been excluded by the aforementioned pre-filtering process, there will be no additional consumption of computing resources due to the recall of such content, thus reducing the amount of computing power required for content recall; on the other hand, the process of determining candidate indexes can be implemented when the server performs content recall, that is, the server can perform pre-filtering, without having to push the recalled content to the user's client for filtering, saving the client's computing resources. At the same time, avoiding increased latency caused by filtering the recalled content on the client side, it is possible to recommend content to the user more quickly. Attached Figure Description

[0030] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0031] Figure 1 This is a flowchart illustrating a content recall method provided in an embodiment of the present invention.

[0032] Figure 2 This is a schematic diagram of the structure of a content recall system provided in an embodiment of the present invention.

[0033] Figure 3 This is a flowchart illustrating a content recommendation method provided in an embodiment of the present invention.

[0034] Figure 4 This is a schematic diagram of a content recall device provided in an embodiment of the present invention.

[0035] Figure 5 This is a schematic diagram of a content recommendation device provided in an embodiment of the present invention.

[0036] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0037] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.

[0038] In one embodiment of the present invention, a content retrieval method is provided, the method being applied to a server in a content recommendation system, comprising:

[0039] In response to a content recommendation instruction sent by the client, historical interaction data of the target account logged into the client is obtained;

[0040] Candidate indices are determined based on the relationships between the historical interaction data and the initial indices of the candidate content; wherein the similarity of the content associated with each relationship is greater than a preset similarity threshold.

[0041] The content corresponding to the candidate index is used as the content retrieval result.

[0042] As can be seen from the above, in the solution provided by the embodiments of the present invention, candidate indices are determined in the initial index based on historical interaction data reflecting user behavior during the recall process, so that the content corresponding to the candidate indices has characteristics that conform to the user's browsing habits. In this way, the content corresponding to the candidate indices is used as the content recall result, and the recalled content is all content that conforms to the user's browsing habits. Therefore, there is no need to perform filtering based on user behavior after the recall, which is equivalent to implementing filtering in advance during the recall process—that is, implementing a pre-filtering process. In this case, the number of recalled content can be controlled during the content recall process. That is, if the number of determined candidate indices is insufficient, a sufficient number of new candidate indices are searched from the initial index that is associated with the determined candidate indices, thereby ensuring that the number of recalled content is sufficient.

[0043] Under this premise, on the one hand, since content that does not conform to the user's browsing habits has been excluded by the aforementioned pre-filtering process, there will be no additional consumption of computing resources due to the recall of such content, thus reducing the amount of computing power required for content recall; on the other hand, the process of determining candidate indexes can be implemented when the server performs content recall, that is, the server can perform pre-filtering, without having to push the recalled content to the user's client for filtering, saving the client's computing resources. At the same time, avoiding increased latency caused by filtering the recalled content on the client side, it is possible to recommend content to the user more quickly.

[0044] In order to achieve content recall and reduce the amount of computation required for content recall while ensuring a sufficient number of recalled contents, embodiments of the present invention provide a content recall method, apparatus, device, and storage medium, which are described below.

[0045] In one embodiment of the present invention, see Figure 1 A flowchart of a content retrieval method is provided. This method is applied to the server in a content recommendation system and includes the following steps S101-S103.

[0046] Step S101: In response to the content recommendation instruction sent by the client, obtain the historical interaction data of the target account of the logged-in client.

[0047] The aforementioned content recommendation system is a system for implementing content recommendation functions. It may include one or more servers. The servers are used to recommend content to clients on the user's electronic devices. The recommended content may be video, audio, text, etc., and the embodiments of the present invention do not limit this.

[0048] Content recommendation instructions are instructions sent by the client to the server requesting recommendations. Content recommendation instructions can be triggered manually by the user, for example, when the user uses the client's content search function, the content recommendation instruction can be triggered when the user enters search terms into the client interface; or, content recommendation instructions can be triggered automatically, for example, when the user logs into the client or opens the client's homepage, the client automatically sends content recommendation instructions.

[0049] The target account is the identity a user has when using the client.

[0050] Historical interaction data records the interactions between users and content provided by the server. For example, it can record the name of the content that the user interacted with, the duration of the interaction, etc.

[0051] The method for obtaining historical interaction data is related to how the historical interaction data is stored. Historical interaction data can be pre-stored in a local log file on the client's electronic device; in this case, the historical interaction data can be retrieved from the local log file. Alternatively, the client can record historical interaction data each time the user interacts with content provided by the server and send the historical interaction data to the server, which then stores it in a pre-defined storage unit. In this way, the server can directly retrieve the historical interaction data from the aforementioned storage unit.

[0052] Step S102: Determine candidate indexes based on the correlation between historical interaction data and the initial indexes of candidate content.

[0053] Among them, the similarity of the content associated with each relationship is greater than the preset similarity threshold.

[0054] The initial index is a reference to the content, and each initial index can correspond to one piece of content. The initial index can contain relevant information about its corresponding content, such as fields describing the content.

[0055] There is connection information between the initial indexes that are related. Based on the two initial indexes that are connected, the search can be performed from either initial index to the other initial index.

[0056] Connection information can represent the distance between initial indices that have a relationship. The distance is determined based on the similarity of the content corresponding to the initial indices, that is, the content associated with the relationship.

[0057] In this case, the specific relationship can be obtained through the content features of the content corresponding to two or more initial indices. Content features are representations of the attributes of the content, such as the name or summary of a piece of content.

[0058] The following describes two ways to build associations.

[0059] In the first implementation, if the similarity between the content features of two items is greater than a preset similarity threshold, then an association is established between the initial indices of the two items.

[0060] This implementation method calculates the similarity between content features, identifies content with similar features, and establishes a relationship between the initial indices of the identified content.

[0061] In this case, the similarity between the feature vectors representing the content features of the two items can be calculated separately, and this similarity can be used as the similarity between the two items. Specifically, the cosine similarity or inner product between the feature vectors can be calculated as the similarity between the feature vectors.

[0062] In the second implementation, a random projection tree method can be used to divide all content features, and a tree-structured association relationship can be established between the content features contained in each subtree as an index.

[0063] As can be seen from the above, the established associations can indicate that the related content has a high degree of similarity, so similar content can be recommended based on the index.

[0064] Based on the established relationships, the initial index can be used to search for content with similar behavioral characteristics to those recorded in historical interaction data, and candidate indexes corresponding to the searched content can be obtained.

[0065] For details on the specific implementation, please refer to the subsequent embodiments, which will not be described in detail here.

[0066] The behavioral characteristics can be real-time. In one embodiment of the present invention, behavioral characteristics representing the acquired information can be obtained by acquiring information such as click behavior and evaluation behavior related to the content browsed, which are recorded in real time by the client when the user browses content provided by the server.

[0067] Step S103: Use the content corresponding to the candidate index as the content retrieval result.

[0068] Candidate indexes can record the reference positions of corresponding content, thus allowing the corresponding content to be determined based on the candidate indexes, serving as the content retrieval result.

[0069] As can be seen from the above, in the solution provided by the embodiments of the present invention, candidate indices are determined in the initial index based on historical interaction data reflecting user behavior during the recall process, so that the content corresponding to the candidate indices has characteristics that conform to the user's browsing habits. In this way, the content corresponding to the candidate indices is used as the content recall result, and the recalled content is all content that conforms to the user's browsing habits. Therefore, there is no need to perform filtering based on user behavior after the recall, which is equivalent to implementing filtering in advance during the recall process—that is, implementing a pre-filtering process. In this case, the number of recalled content can be controlled during the content recall process. That is, if the number of determined candidate indices is insufficient, a sufficient number of new candidate indices are searched from the initial index that is associated with the determined candidate indices, thereby ensuring that the number of recalled content is sufficient.

[0070] Under this premise, on the one hand, since content that does not conform to the user's browsing habits has been excluded by the aforementioned pre-filtering process, there will be no additional consumption of computing resources due to the recall of such content, thus reducing the amount of computing power required for content recall; on the other hand, the process of determining candidate indexes can be implemented when the server performs content recall, that is, the server can perform pre-filtering, without having to push the recalled content to the user's client for filtering, saving the client's computing resources. At the same time, avoiding increased latency caused by filtering the recalled content on the client side, it is possible to recommend content to the user more quickly.

[0071] The following describes the process of determining candidate indexes in step S102 above, based on the first method of constructing the association relationship between initial indexes.

[0072] In one embodiment of the present invention, determining a candidate index based on the correlation between historical interaction data and the initial index of candidate content includes:

[0073] Based on the similarity between historical interaction data and the initial index, determine the target index; determine whether the content corresponding to the target index is content with the target account tag; if not, determine the target index as a candidate index; if so, re-determine the target index based on the association relationship of the target index, and return to the step of determining whether the content corresponding to the target index is content with the target account tag.

[0074] The similarity between historical interaction data and the initial index can be represented by the similarity between the behavioral features corresponding to the historical interaction data and the content features to which the initial index belongs. The specific similarity calculation method is the same as the similarity calculation method mentioned in the aforementioned step S102 embodiment.

[0075] Specifically, the initial index with high similarity can be obtained as the target index in the following way:

[0076] Based on the similarity between the content features of the candidate content and the behavioral features of historical interaction data, a first feature is determined; the nearest neighbor feature with the highest similarity to the behavioral feature is obtained from the nearest neighbor features of the first feature and used as the second feature, and the initial index of the content corresponding to the second feature is used as the target index; wherein, there is a correlation between the initial index of the content to which the first feature belongs and the initial index of the content to which the nearest neighbor feature belongs.

[0077] In one scenario, historical interaction data can also be compared with all initial indexes.

[0078] In other words, the similarity between each content feature and behavioral feature can be compared to determine the content feature with the highest similarity as the first feature, and then the second feature can be selected from the nearest neighbor features of the first feature.

[0079] In another scenario, historical interaction data can be compared with a portion of the initial index for similarity.

[0080] In this case, the content features of the content belonging to the initial index can be used as representative features, and comparisons can be made in the following manner:

[0081] From the representative features, select the content feature that has the highest similarity to the behavioral features as the first feature;

[0082] Based on the association between the initial index corresponding to the representative feature and the initial index corresponding to the non-representative feature of the candidate content, the nearest neighbor feature with the highest similarity to the behavioral feature is obtained from the nearest neighbor features associated with the first feature through the index, thus obtaining the second feature, and the index corresponding to the second feature is determined as the target index.

[0083] Representative features are some of the features in each content feature. By using the initial index corresponding to all representative features, the content features of any content with the initial index can be found. For example, a representative feature can be a feature located in the first layer of the association relationship formed by the layered network of HNSW (Hierarchical Navigable Small World).

[0084] Since the number of representative features is relatively small compared to the number of content features in all content, it is possible to quickly traverse all representative features and select the representative feature with the highest similarity to the behavioral feature as the first feature.

[0085] Non-representative features are those features other than representative features among all the content features.

[0086] The first feature is a representative feature, and there are indices for non-representative features. This allows us to determine non-representative features that are related to the first feature based on the index of the first feature, and then use them as nearest neighbor features.

[0087] As can be seen from the above, when performing content recall, the nearest neighbor features of the first feature with high similarity to the behavioral feature are always selected for recommendation. The similarity between the nearest neighbor features and the first feature is that they are more similar to the behavioral feature. In this way, the content belonging to the selected second feature is more likely to match the user's behavioral habits reflected by the behavioral feature. Therefore, the recalled content is more likely to be viewed by users, thus improving the accuracy of content recall.

[0088] The target account is tagged with a marker that represents the user's willingness to browse the content provided by the service. Thus, for each identified candidate index, it is both compared with historical interaction data to determine if the content of the identified candidate index matches the user's interactive behavior habits, and further filtered based on the target account tag. This provides more comprehensive filtering information, making the retrieved content more likely to be accepted by the user.

[0089] The following explains how to determine whether the content corresponding to the target index is content marked with the target account.

[0090] In one embodiment of the present invention, the target account tag may be a tag indicating that the content corresponding to the target index has been viewed by the target account.

[0091] The client can automatically tag the content viewed by the user belonging to the target account.

[0092] Furthermore, the target account tag can be related to the user's browsing progress. That is, the target account tag can be a tag indicating that the browsing progress of the content corresponding to the target index has exceeded a preset progress indicator.

[0093] The progress indicators can be set based on the duration of content viewed by the target user, the proportion of the viewed portion to the total content, etc., but this embodiment of the invention does not limit them.

[0094] For content that has already been viewed, if the viewing progress is relatively small, users may revisit it. Therefore, such content does not completely lose the possibility of being viewed again simply because it has already been viewed. Consequently, recalling and recommending such content may still be effective for users.

[0095] In this way, browsing progress can be used to more accurately distinguish content that users have already viewed and are unlikely to view again, thus improving the accuracy of content recall.

[0096] In another embodiment of the present invention, the target account tag may further include a filtering tag set by the target account for the target index. The filtering tag indicates that the user to whom the target account belongs does not intend to browse the content.

[0097] For example, the target account can perform the following judgment and set a flag for content that does not conform to the user's browsing intentions:

[0098] In one method of judgment, content that the user does not intend to browse can be used as reference content, and the content to be marked and the reference content can be compared to determine whether the content to be marked meets the user's browsing intention.

[0099] Users can configure settings by checking preference options or setting tags, etc.

[0100] Since content that has already been viewed or filtered by the target account is more likely to not match the user's viewing habits, candidate content with various target account tags is not used during content recall. This avoids the situation where such content is recommended to users but not viewed, allowing users to access content that better meets their own requirements within a limited space and time, thus improving the user experience.

[0101] The specific implementation method for making the judgment is as follows.

[0102] In one implementation, the attributes of the content to be tagged and the reference content are compared. Specifically, the ID (Identity document), name, keywords, etc. of the content to be tagged and the reference content can be compared.

[0103] In another implementation, besides comparing the content to be labeled and the reference content, content with similar characteristics to the reference content can also be compared with the content to be labeled. Specifically, since the reference content can establish association relationships between indexes and other content as described in the above embodiments, in this case, the associated content obtained through the association relationships between indexes and related to the reference content can also be used to compare with the content to be labeled, thereby referencing the content characteristics of more content during the comparison process and improving the accuracy of the judgment.

[0104] For example, associated content can be determined as follows:

[0105] Calculate the first similarity between the content features of the content to be tagged and the content of the reference content; calculate the second similarity between the content features of the content to be tagged and the third feature, wherein the third feature is: the content features of the associated content, and there is an association between the index of the associated content and the index of the reference content; if there is a first similarity not less than a preset first similarity threshold, and / or there is a second similarity not less than a preset second similarity threshold, then it is determined that the content to be tagged does not meet the user's browsing intention.

[0106] The calculation methods for the first similarity and the second similarity can be referred to the aforementioned embodiments, and will not be described in detail here.

[0107] The indexes for related content and reference content can be part of the initial index in the aforementioned embodiments. Alternatively, a set of indexes with similar structures can be created for related content and reference content according to the construction method of the aforementioned initial index association relationship; this embodiment of the invention does not limit this.

[0108] Following the above method, a high first similarity score and a high second similarity score indicate that the content to be tagged is similar to the reference content and its related content. The reference content and its related content are both likely to be inconsistent with the user's browsing preferences, and the content to be tagged that is similar to it is even more likely to be unviewed by the user.

[0109] Another method of judgment is to use pre-determined trending content and content with a large number of viewers on the information platform as reference content, and compare the content to be marked with the reference content to determine whether the content to be marked meets the user's browsing intentions.

[0110] The following explains how to redetermine the implementation method of the target index based on the relationships between the target indexes.

[0111] In one embodiment of the present invention, if the number of determined candidate indexes does not reach a preset number threshold, the target index is re-determined based on the association between the initial indexes, and the process returns to the step of determining whether the content corresponding to the target index is content not marked by the target account.

[0112] The preset quantity threshold can be specified in advance, such as 10, 50, 100, etc.

[0113] The method for redetermining the target index varies depending on the structure of the relationships between the initial indexes. The following sections will explain each case separately.

[0114] In the first scenario, where there is a direct relationship between the initial indexes, an initial index that is related to the target index can be searched, and the resulting new initial index can be used as the target index.

[0115] For example, in the case where the initial index of the content corresponding to the second feature is used as the target index in the aforementioned embodiment, the nearest neighbor feature with the highest similarity to the behavioral feature can be obtained from the nearest neighbor features corresponding to the initial index that are related to the initial index corresponding to the second feature, based on the association relationship between the initial index corresponding to the representative feature and the initial index corresponding to the non-representative feature. This results in a new second feature, and the initial index corresponding to the new second feature is used as the target index.

[0116] Therefore, the method for re-acquiring the target index is achieved by redetermining the second feature. The method for determining the second feature is the same as in the previous embodiment. In this way, when there are neighboring features for the second feature, new second features can be obtained by continuously searching for neighboring features, and new target indexes can be obtained when the number of recalled contents is insufficient.

[0117] In addition, when obtaining the second feature multiple times to determine multiple target indices, the second features that have been obtained can be recorded each time a new second feature is obtained. This way, when searching for nearest neighbor features, the second features that have been obtained can be excluded first, avoiding the loop between the few nearest neighbor features with the highest similarity to all content when searching for the second feature.

[0118] As can be seen from the above, when redetermining the target index, the search is always conducted among the nearest neighbor features of the content belonging to the initial index which has an association relationship. This ensures that any nearest neighbor feature obtained is always associated with the initially obtained behavioral feature through one or more indexes, which increases the probability that the nearest neighbor feature matches the behavioral feature. Correspondingly, the content belonging to the determined target index is also close to the user's browsing habits, and is therefore more likely to be browsed.

[0119] Since users are more likely to browse the recommended content, the click-through rate of users browsing content on the client will also increase accordingly.

[0120] In the second case, if the association is a tree structure, the node representing the target index in the tree structure association can be obtained, and the initial index of its sibling node can be determined as the new target index, etc.

[0121] If the number of identified candidate indexes reaches the threshold, no new candidate indexes need to be identified, and the content belonging to the identified candidate indexes can be selected as the content retrieval result.

[0122] The quantity threshold can be a pre-set fixed value. This way, the number of contents recalled in each content recall will always be consistent with the quantity threshold.

[0123] See Figure 2 The content retrieval method described above can be implemented by the following system. Here, TF service represents TensorFlow service, and TensorFlow is the software framework for deploying services, used to run the content retrieval method in the above embodiments. The service request module is used to receive requests for content recommendations to users.

[0124] The model management module is used to manage models, such as the processes contained in the TensorFlow runtime within the system, including the ANN (Approximate Nearest Neighbor) model and the DNN (Deep Neural Network) model shown in the diagram.

[0125] The ANN model is used to determine the target index according to the method described in the foregoing embodiments. Taking video as an example of the content to be recalled, the ANN model has an initial index pre-built based on the features of videos in a preset video library.

[0126] exist Figure 2 In the system shown, the DNN model can be used to predict the videos that a user may browse based on the user's real-time behavioral characteristics, or it can be used as a target video with its index as the target index, thereby assisting the ANN model in obtaining the target index.

[0127] The pre-filtering module is used to determine whether the target video to which the target index belongs has a target account tag after obtaining each target index, and to filter out the target videos with the target account tag. The remaining videos are then recommended to the user, i.e., the filtered recall results are returned.

[0128] For the recall methods described above, please refer to [link / reference]. Figure 3 This invention also provides a flowchart of a content recommendation method, which is applied to a server in a content recommendation system and includes the following steps S301-S303:

[0129] Step S301: Use the content recall method to recall candidate content and obtain the recalled content.

[0130] The content recall method used can be any of the content recall methods described in the foregoing embodiments.

[0131] Step S302: Based on the recalled content, determine the content to be recommended;

[0132] The recalled content can be further filtered to determine the final recommendations to users. Filtering methods can include coarse sorting, fine sorting, etc. For example, fine sorting could be used to further filter the recalled content based on its browsability, priority, and other content attributes, then selecting the top predetermined number of items with the highest scores as candidates.

[0133] Step S303: Send the content to be recommended to the client so that the client can display the content to be recommended.

[0134] The client may display a link to obtain the content to be recommended to the user on the user interface, or pop up a new window to display the content interface, etc., but the embodiments of the present invention are not limited to this.

[0135] In this way, by using content recall methods, candidate content that matches the user's browsing behavior can be selected, and recommendations can be made based on the selected candidate content. The recommended content is more likely to be accepted by the user.

[0136] In one embodiment of the present invention, see Figure 4 A schematic diagram of a content recall device is provided.

[0137] The interaction data acquisition module 401 is used to acquire historical interaction data of the target account logged into the client in response to the content recommendation instruction sent by the client.

[0138] The candidate index determination module 402 is used to determine candidate indexes based on the correlation between the historical interaction data and the initial indexes of the candidate content; wherein the similarity of the content associated with each correlation is greater than a preset similarity threshold.

[0139] The recall result determination module 403 is used to take the content corresponding to the candidate index as the content recall result.

[0140] As can be seen from the above, in the solution provided by the embodiments of the present invention, candidate indices are determined in the initial index based on historical interaction data reflecting user behavior during the recall process, so that the content corresponding to the candidate indices has characteristics that conform to the user's browsing habits. In this way, the content corresponding to the candidate indices is used as the content recall result, and the recalled content is all content that conforms to the user's browsing habits. Therefore, there is no need to perform filtering based on user behavior after the recall, which is equivalent to implementing filtering in advance during the recall process—that is, implementing a pre-filtering process. In this case, the number of recalled content can be controlled during the content recall process. That is, if the number of determined candidate indices is insufficient, a sufficient number of new candidate indices are searched from the initial index that is associated with the determined candidate indices, thereby ensuring that the number of recalled content is sufficient.

[0141] Under this premise, on the one hand, since content that does not conform to the user's browsing habits has been excluded by the aforementioned pre-filtering process, there will be no additional consumption of computing resources due to the recall of such content, thus reducing the amount of computing power required for content recall; on the other hand, the process of determining candidate indexes can be implemented when the server performs content recall, that is, the server can perform pre-filtering, without having to push the recalled content to the user's client for filtering, saving the client's computing resources. At the same time, avoiding increased latency caused by filtering the recalled content on the client side, it is possible to recommend content to the user more quickly.

[0142] In one embodiment of the present invention, the candidate index determination module 402 is configured to:

[0143] Based on the similarity between the historical interaction data and the initial index, a target index is determined; it is determined whether the content corresponding to the target index is content with a target account tag; if not, the target index is determined as a candidate index; if so, the target index is re-determined based on the association relationships of the target index, and the process returns to the step of determining whether the content corresponding to the target index is content with a target account tag; wherein, the similarity of the content associated with each association relationship is greater than a preset similarity threshold.

[0144] In this way, for each identified candidate index, the content of the identified candidate index is compared with historical interaction data to determine whether it conforms to the user's interaction behavior habits, and it is further filtered based on the target account tags. The filtering information is more sufficient, and the content to be recalled is more likely to be accepted by the user.

[0145] In one embodiment of the present invention, the target account marker includes at least one of the following:

[0146] A marker indicating that the content corresponding to the target index has been viewed by the target account;

[0147] And / or,

[0148] The target account sets filter tags for the target index.

[0149] Since content that has already been viewed or filtered by the target account is more likely to not match the user's viewing habits, candidate content with various target account tags is not used during content recall. This avoids the situation where such content is recommended to users but not viewed, allowing users to access content that better meets their own requirements within a limited space and time, thus improving the user experience.

[0150] In one embodiment of the present invention, the marker indicating that the content corresponding to the target index has been viewed by the target account includes one of the following:

[0151] A marker indicating that the browsing progress of the content corresponding to the target index has exceeded a preset progress indicator.

[0152] In this way, browsing progress can be used to more accurately distinguish content that users have already viewed and are unlikely to view again, thus improving the accuracy of content recall.

[0153] In one embodiment of the present invention, the apparatus further includes:

[0154] The threshold determination module is used to redetermine the target index based on the correlation between the initial indexes if the number of determined candidate indexes does not reach the preset number threshold, and return to the step of determining whether the content corresponding to the target index is content not marked by the target account.

[0155] As can be seen from the above, when redetermining the target index, the search is always conducted among the nearest neighbor features of the content belonging to the initial index which has an association relationship. This ensures that any nearest neighbor feature obtained is always associated with the initially obtained behavioral feature through one or more indexes, which increases the probability that the nearest neighbor feature matches the behavioral feature. Correspondingly, the content belonging to the determined target index is also close to the user's browsing habits, and is therefore more likely to be browsed.

[0156] In one embodiment of the present invention, the candidate index determination module 402 is configured to:

[0157] Based on the similarity between the content features of the candidate content and the behavioral features of the historical interaction data, a first feature is determined; the nearest neighbor feature with the highest similarity to the behavioral feature is obtained from the nearest neighbor features of the first feature, and is used as the second feature, with the initial index of the content corresponding to the second feature as the target index; wherein, there is an association between the initial index of the content to which the first feature belongs and the initial index of the content to which the nearest neighbor feature belongs; it is determined whether the content corresponding to the target index is content with a target account tag; if not, the target index is determined as a candidate index; if so, the target index is re-determined according to the association relationship of the target index, and the process returns to the step of determining whether the content corresponding to the target index is content with a target account tag; wherein, the similarity of each content associated with each association relationship is greater than a preset similarity threshold.

[0158] As can be seen from the above, when performing content recall, the nearest neighbor features of the first feature with high similarity to the behavioral feature are always selected for recommendation. The similarity between the nearest neighbor features and the first feature is that they are more similar to the behavioral feature. In this way, the content belonging to the selected second feature is more likely to match the user's behavioral habits reflected by the behavioral feature. Therefore, the recalled content is more likely to be viewed by users, thus improving the accuracy of content recall.

[0159] In one embodiment of the present invention, see Figure 5 A schematic diagram of a content recommendation device is provided. The device is applied to a server in a content recommendation system and includes:

[0160] The recall content acquisition module 501 is used to recall candidate content using the above-mentioned content recall device to obtain the recall content.

[0161] The content to be recommended module 502 is used to determine the content to be recommended based on the recalled content;

[0162] The content to be recommended module 503 is used to send the content to be recommended to the client so that the client can display the content to be recommended.

[0163] In this way, by using content recall methods, candidate content that matches the user's browsing behavior can be selected, and recommendations can be made based on the selected candidate content. The recommended content is more likely to be accepted by the user.

[0164] This invention also provides an electronic device, such as... Figure 6 As shown, it includes a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604.

[0165] Memory 603 is used to store computer programs;

[0166] When the processor 601 executes the program stored in the memory 603, it implements the content retrieval or content recommendation method steps described in any of the foregoing embodiments:

[0167] The communication bus mentioned above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of illustration, only one thick line is used to represent it in the diagram, but this does not mean that there is only one bus or one type of bus.

[0168] The communication interface is used for communication between the aforementioned terminal and other devices.

[0169] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0170] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0171] In another embodiment of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements any of the content retrieval or content recommendation methods described in the above embodiments.

[0172] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the content recall or content recommendation methods described in the above embodiments.

[0173] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0174] 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 apparatus 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 apparatus. 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 apparatus that includes said element.

[0175] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, devices, and storage media are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0176] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A content recall method characterized by, The method is applied to a server in a content recommendation system, including: In response to a content recommendation instruction sent by the client, historical interaction data of the target account logged into the client is obtained; Based on the correlation between the historical interaction data and the initial indexes of the candidate content, a candidate index is determined, including: determining a target index based on the similarity between the historical interaction data and the initial index; determining whether the content corresponding to the target index is content with a target account tag; if not, the target index is determined as a candidate index; if so, based on the correlation between the target index, an initial index with a correlation with the target index is searched, and the searched initial index is used as the redefined target index, and the process returns to the step of determining whether the content corresponding to the target index is content with a target account tag; wherein, the correlation refers to the correlation between the initial indexes of each candidate content, and the similarity between the content associated with each correlation is greater than a preset similarity threshold; The content corresponding to the candidate index is used as the content retrieval result.

2. The method according to claim 1, characterized in that, The target account tag includes at least one of the following: A marker indicating that the content corresponding to the target index has been viewed by the target account; And / or, The target account sets filter tags for the target index.

3. The method according to claim 2, characterized in that, The marker indicating that the content corresponding to the target index has been viewed by the target account includes one of the following: A marker indicating that the browsing progress of the content corresponding to the target index has exceeded a preset progress indicator.

4. The method according to claim 1, characterized in that, After determining the target index as a candidate index, the process further includes: If the number of identified candidate indexes does not reach the preset threshold, the target index is re-determined based on the relationships between the initial indexes, and the process returns to the step of determining whether the content corresponding to the target index is content not marked by the target account.

5. The method according to claim 1, characterized in that, The step of determining the target index based on the similarity between the historical interaction data and the initial index includes: The first feature is determined based on the similarity between the content features of the candidate content and the behavioral features of the historical interaction data; The nearest neighbor feature with the highest similarity to the behavioral feature is obtained from the nearest neighbor features of the first feature and used as the second feature. The initial index of the content corresponding to the second feature is used as the target index. There is a correlation between the initial index of the content to which the first feature belongs and the initial index of the content to which the nearest neighbor feature belongs.

6. A content recommendation method, characterized in that, The method is applied to a server in a content recommendation system, including: The candidate content is retrieved using the method described in any one of claims 1-5, and the retrieved content is obtained. Based on the recalled content, determine the content to be recommended; The content to be recommended is sent to the client so that the client can display the content to be recommended.

7. A content recall device, characterized in that, The device is used in a server of a content recommendation system and includes: The interaction data acquisition module is used to acquire historical interaction data of the target account logged into the client in response to the content recommendation instruction sent by the client. The candidate index determination module is used to determine candidate indexes based on the correlation between the historical interaction data and the initial indexes of the candidate content; wherein, the correlation refers to the correlation between the initial indexes of each candidate content, and the similarity between the content associated with each correlation is greater than a preset similarity threshold. The recall result determination module is used to take the content corresponding to the candidate index as the content recall result; The candidate index determination module is specifically used to determine a target index based on the similarity between the historical interaction data and the initial index; determine whether the content corresponding to the target index is content with a target account tag; if not, determine the target index as a candidate index; if so, search for an initial index with a relationship with the target index based on the association relationship of the target index, use the searched initial index as the re-determined target index, and return to execute the step of determining whether the content corresponding to the target index is content with a target account tag.

8. A content recommendation device, characterized in that, The device is used in a server of a content recommendation system and includes: The recall content acquisition module is used to recall candidate content using the apparatus described in claim 7, and obtain the recall content. The module for determining content to be recommended is used to determine content to be recommended based on the recalled content. The module for sending recommended content is used to send the recommended content to the client so that the client can display the candidate content.

9. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method according to any one of claims 1-5 or 6.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-5 or 6.

Images