CONTENT RECOMMENDATION METHOD, DEVICE, ELECTRONIC DEVICE, STORAGE MEDIUM, AND COMPUTER PROGRAM
By determining user structural preferences through deep learning, the method enhances content recommendation by matching user interests across multiple structures, addressing the limitations of conventional hierarchical screening mechanisms.
Patent Information
- Application Number
- JP2021049336
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2020-09-01
- Filing Date
- 2021-03-24
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2041-03-24
AI Technical Summary
Conventional content recommendation methods based on hierarchical screening mechanisms fail to adequately match content provision with user preferences, leading to issues such as information cocoons and difficulty in balancing multiple business goals.
A content recommendation method that determines a user's structural preferences based on past behavior data and content characteristics, using deep learning to classify and recommend content across multiple structures, enabling structured recall, sorting, and fusion to enhance matching with user interests.
Improves the matching between content provision and user preferences by providing diverse and relevant recommendations, breaking information cocoons, and balancing business goals effectively.
Smart Images

Figure 0007584340000001 
Figure 0007584340000002 
Figure 0007584340000003
Abstract
Description
[Technical field]
[0001] The present application relates to the Internet field, in particular to smart recommendation technology, and specifically to a content recommendation method, device, electronic device, storage medium and computer program. [Background technology]
[0002] In the current content ecology-based Internet products, personalized content recommendation policies are widely applied. The traditional method of content recommendation mainly includes a three-stage hierarchical screening mechanism of recall, sorting and fusion, which is composed of recall, sorting and fusion modules. Among them, the recall module performs primary screening of the contents that the user may be interested in from the entire resource library based on the user's basic information and past user behavior, the sorting module uses a machine learning algorithm to score the contents returned by the recall module one by one, and the fusion module performs inverted indexing on the scores of the sorting module, realizes business logic such as diversity, and combines the final recommendation results.
[0003] Although the content recommendation method based on the hierarchical screening mechanism has achieved a relatively good recommendation effect, it still has shortcomings in matching content provision with user preferences. Summary of the Invention
[0004] The present application provides a content recommendation method, device, electronic device, storage medium and computer program for improving the matching degree between content offerings and user preferences.
[0005] In a first aspect, the present application provides a content recommendation method, including: determining a target structure preference of a user based on past behavior data of the user using a product and structural characteristics of recommended content, where the structure is determined by classifying the recommended content based on a classification method of an arbitrary content tag system; and determining a recommendation result for each user, including at least two types of structures and recommended contents corresponding to the various structures, based on the target structure preference.
[0006] In a second aspect, the present application further provides a content recommendation device, including: a structure preference determination module for determining a target structure preference of a user based on past behavior data of the user using a product and structural features of the recommended content, where the structure is determined by classifying the recommended content based on a classification method of an arbitrary content tag system; and a recommendation result determination module for determining a recommendation result for each user, including at least two types of structures and recommended content corresponding to each type of structure, based on the target structure preference.
[0007] In a third aspect, the present application provides an electronic device comprising at least one processor and a memory communicatively connected to the at least one processor, the memory storing instructions executable by the at least one processor, the instructions, when executed by the at least one processor, causing the at least one processor to perform a content recommendation method according to any of the embodiments of the present application.
[0008] In a fourth aspect, the present application provides a non-transitory computer readable storage medium having computer instructions stored thereon, the computer instructions being for causing the computer to perform a content recommendation method according to any of the embodiments of the present application.
[0009] According to a fifth aspect, there is provided a computer program product which, when executed by a processor, realises the content recommendation method according to any of the embodiments of the present application.
[0010] Based on the technical solutions of the present application, a global structured display and interest inference can be realized by determining a user's structural preferences, and then the recommendation results are determined based on the structural preferences, so that the recommendation results have corresponding recommended content for various different structures, thereby providing recommended content that is more in line with the user's preferences, and improving the matching degree between content provision and user preferences.
[0011] It should be noted that the contents described in the Summary of the Invention are not intended to limit the key or important features of the embodiments of the present application, and do not limit the scope of the present application. Other features of the present application will be easily understood from the following description. Other effects of the above-mentioned optional embodiments will be described below based on specific examples. [Brief description of the drawings]
[0012] The drawings are used for a better understanding of the present application and do not constitute limitations to the present application. [Figure 1] 1 is a schematic flowchart of a content recommendation method according to an embodiment of the present application; [Figure 2A] 1 is a schematic flowchart of a content recommendation method according to an embodiment of the present application; [Figure 2B] FIG. 2 is a network structure diagram of a conversation level structure preference model according to an embodiment of the present application. [Figure 3A] 1 is a schematic flowchart of a content recommendation method according to an embodiment of the present application; [Figure 3B] FIG. 2 is a schematic diagram of a refresh level structure preference model according to an embodiment of the present application; [Figure 4] FIG. 2 is a structural schematic diagram of a content recommendation device according to an embodiment of the present application; [Diagram 5]FIG. 2 is a block diagram of an electronic device for implementing a content recommendation method according to an embodiment of the present application. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0013] The following describes exemplary embodiments of the present application with reference to the drawings, and various details of the embodiments of the present application are described here for ease of understanding, but these are merely illustrative. Therefore, it is obvious that those skilled in the art can make various changes and modifications to the embodiments of the present specification without departing from the scope and gist of the present application. In the following description, for clarity and simplicity, the description of known functions and configurations is omitted.
[0014] FIG. 1 is a schematic flow chart of a content recommendation method according to an embodiment of the present application, which can be applied to recommending content to a user, for example, recommending content to a user through a feed system, and relates to smart recommendation, deep learning and big data technology. The method can be performed by a content recommendation device, which is realized in a software and / or hardware manner, and is preferably located in an electronic device, such as a server or a computing device. As shown in FIG. 1, the method specifically includes the following steps:
[0015] S101: Determine a user's target structure preference based on the user's past behavior data of using a product and the structural characteristics of the recommended content, and the structure is determined by classifying the recommended content based on a classification method of an arbitrary content tag system.
[0016] In the recommendable resource content of Internet products based on content ecology, it can be divided into different structures according to the classification method of any content tag system, and various structures correspond to corresponding resource content. It should be noted that the structure in this application may be an explicit structure or an implicit structure. Among them, the explicit structure is a category structure with a real meaning classified according to a certain classification method, such as different structures such as video and emoji, sports and entertainment, international and domestic. The implicit structure does not necessarily have a real meaning, and may be an abstract classification method, which may be specifically defined according to business needs. The embodiments of this application do not limit the classification method of the structure, and the technical solution of this application is suitable for structures classified according to any method.
[0017] The past behavioral data of the user using the product may include, for example, which recommended content the user clicked on, or which recommended content the user did not click on, or the user spent a long time viewing some recommended content, and the like. By analyzing these user behaviors, it is possible to express which structure's recommended content the user is relatively interested in, or which structure's recommended content the user is not very interested in, and such a preference relationship of the user is also called structure preference. Therefore, based on the past behavioral data of the user using the product, and further combining the structure characteristics of the structure itself, the target structure preference of the user can be determined. Among them, the structure characteristics of the structure itself may be, for example, the name of the structure, or the popularity or interest level of a specific structure among many structures, or a combination of both, and either can be the structure characteristics of the structure itself, thereby providing a basis for personalized scoring of different structures.
[0018] In particular, a prediction model is pre-constructed and trained based on deep learning methods, and a user's target structure preference is determined based on the prediction model based on past behavior data and structure features. For example, a user's interest level is predicted for various structures based on a scoring method based on the prediction model, and the score or percentage is used as the predicted value. Then, a predetermined number of structures that rank high in all the predicted values are selected according to the score to obtain the final structure preference. The embodiments of the present application do not limit the specific model structure and algorithm used.
[0019] S102: Determine, for each user, a recommendation result including at least two kinds of structures and recommended content corresponding to each kind of structure based on the target structure preference.
[0020] After the target structure preference is determined, at least two structures that the user is most interested in are obtained, and then corresponding recommended contents are obtained based on the various structures, and these several structures and their recommended contents become the final recommendation result.
[0021] Specifically, the recommendation result of each user can be determined according to the target structure preference through three stages: recall, sorting and fusion. That is, in the recall and sorting stage, structured recall and sorting are realized, and recall and sorting are performed based on different structures, and in the fusion stage, structured fusion is also realized, and fusion is performed in various structures, and corresponding recommended content is determined based on various structures. The specifically adopted recall algorithm, sorting algorithm and fusion algorithm are not limited to the embodiments of the present application.
[0022] Furthermore, the recommendation policy framework based on hierarchical screening of recall, sorting and fusion in the prior art is lacking in matching content offerings with user preferences.
[0023] First, in the prior art, the fusion module composes the final delivery result in the reverse order of the sorting scores, so for example, if a user clicks on video-based content more than picture-based content, more video-based content will be provided in the recommendation policy framework based on hierarchical screening, and the top 100 recommended contents may be video-based, which means that there is no explicit control on the delivered content structure, and the delivered structure may not match the user's preferences. Although the user has clicked on a lot of video-based content in the past, it does not mean that the user is only interested in video-based content. In the embodiment of the present application, content recommendation can be made based on the actual structuring needs of the user, and the final recommendation result includes at least two types of structures and recommended contents corresponding to the various types of structures, thereby realizing the explicit control of structuring, and contents in multiple types of structures that the user may be interested in can be recommended to the user, and at the same time, the variety and richness of the recommended content can be improved, which is helpful in inducing immersive experience and consumption for the user.
[0024] Then, for a user with less popular interest points, in the process of hierarchical screening based on the prior art, such interest points are generally easily overwhelmed by the universal interest points with high scores, and cannot fully meet the needs of the user, so the hierarchical screening mechanism in the prior art is easy to overwhelm the less popular interest points and generate information cocoon. In the embodiment of the present application, in order to perform global structured display and interest inference, first determine the user's structured preference, and then perform structured content recommendation, in addition to recommending these popular universal interest points, less popular interest points that match the user are also recommended, thereby breaking the information cocoon and providing recommended content that meets the user's actual structured needs.
[0025] In addition, the content recommendation method based on hierarchical screening in the prior art often adopts the method of adjusting weights at each level when a certain business goal needs to be achieved, and when the business goal gradually increases, the recommendation system tends to become a black box, and while satisfying one business goal, it affects another business goal, and it is difficult to balance between each business goal, which restricts the development of the whole system. The structured content recommendation based on user interest inference realized in the embodiment of the present application can avoid this problem. For example, the resource can be structured and classified according to different business goals, and then interest inference and structured recommendation can be performed, thus avoiding the problem of mutual influence between each business goal.
[0026] In addition, before determining the recommendation result of each user based on the target structure preference, the method can further include searching the structure dimension based on the target structure preference. For example, the inferred structure preference represents the true structured preference inferred based on the past user behavior, but as a recommendation system, the structure that the user is not interested in may be because the recommendation system has not been provided to the user in the past. Therefore, the structure preference is adjusted based on business needs, and interference and exploration are performed on the structure preference, so that the new interest points of the user can be effectively discovered, the information cocoon can be broken, and the user needs can be further satisfied.
[0027] The technical solution of the embodiments of the present application realizes global structured display and interest inference by determining a user's structural preference, and then determines the recommendation result based on the structural preference, so that the recommendation result all has corresponding recommended content based on various different structures, thereby providing recommended content that is more in line with the user's preference, enhancing the matching degree between content provision and user preferences, promoting users to have an immersive experience using the recommendation system, and significantly improving indicators such as product time, delivery, etc.
[0028] 2A is a schematic flow chart of a content recommendation method according to an embodiment of the present application. This embodiment further performs optimization based on the above embodiment. As shown in FIG. 2A, the method specifically includes S201 to S204.
[0029] S201: In response to a conversation start request, determine a user's current conversation level structure preference based on past behavior data in the user's past conversation process and the structural characteristics of the content to be recommended, and set the conversation level structure preference as the target structure preference.
[0030] When a user uses a content recommendation product, from opening a page and starting to use it to closing the page and stopping use, the user always goes through a certain browsing or consumption process, and during that time, a series of actions and feedback are performed, and this process is called a session. The longer the user's session, the more persistent the user's immersion in the content recommendation, which indirectly means that the recommendation effect is good and the degree of matching with the user's preferences is high.
[0031] A user's preferences do not remain consistent in every conversation, so the present application responds to a conversation start request by determining the user's current conversation level structure preference based on the user's past behavior data in the user's past conversation process and the structural features of the content to be recommended, and the conversation level structure preference is taken as the target structure preference. That is, every time a conversation starts, the user's conversation level structure preference of the current conversation is predicted, thereby realizing more suitable recommendation. Wherein, the past behavior data may be, for example, data generated by the user in the previous conversation process, or at least includes data generated in the conversation process occurring within a predetermined period including the previous data.
[0032] Specifically, the past behavior data includes previously selected structure content, previously selected entry content, previously unselected structure content, and previously unselected entry content. The previously selected structure content and previously unselected structure content refer to which structures of the structures related to the recommended content were selected by the user (e.g., selected by clicking) and which structures were not selected, and the previously selected entry content and previously unselected entry content refer to which entries in the recommended content were selected by the user and which were not selected. The previously selected structure content and previously selected entry content may indicate content that is of interest to the user, and the previously unselected structure content and previously unselected entry content may indicate content that is not of interest to the user, and these structure content and entry content may all be displayed in a list format.
[0033] The feature of the structure is represented by the name of the structure and the attribute of the structure, specifically, the name of the structure and the attribute of the structure can be connected to form the feature of the structure, among which, the attribute of the structure is used to display the past click-view information of different structures, such as the number of clicks and the number of views, and can reflect the average advantage of different structures as a whole according to the past click-view information, which can provide information for personalized scoring to a certain extent.
[0034] In addition to considering the past behavior data and the structure features, the first scene features are also considered, and the user's current conversation level structure preference is determined based on the past behavior data, the first scene features and the structure features. That is, the user's current conversation level structure preference is determined based on the past behavior data in the user's past conversation process, the first scene features and the structure features of the content to be recommended. Among them, the first scene features are used to represent the scene to which each conversation belongs. The scenes can be classified according to business needs in advance, for example, the scenes can be classified according to time periods into morning rush hour, morning, noon, afternoon and night rush hour, or the scenes can be classified according to geographical locations, for example, home, office or park. This application does not limit the specific scene classification method. By combining the user's scene features to determine the conversation level structure preference, the inference result of the structure preference is more accurate and more comprehensive.
[0035] In one embodiment, determining a current conversation level structure preference of the user based on past behavior data in the past conversation process of the user and the structure characteristics of the recommended content includes: The method includes using a pre-trained conversation level structure preference model to determine the user's current conversation level structure preference based on the user's past behavior data in the user's past conversation process and the structural characteristics of the content to be recommended, wherein the learning goal of the conversation level structure preference model includes the user's preference for any structure in the conversation.
[0036] That is, the prediction of the conversation level structure preference is achieved by combining deep learning technology and using a pre-trained deep learning model, where the model can be updated daily and is used to predict the user's conversation level structure preference in real time. Specifically, the training process of the conversation level structure preference model includes: Input a plurality of training samples as a model, each training sample being user behavior data generated in one conversation process, and each training sample includes N sub-samples, each sub-sample being user behavior data of various structures displayed in the conversation process, where N is a natural number; training the discourse-level structure preference model using the click rates of various structures for each pre-annotated training sample as model output.
[0037] Among them, since each training sample contains multiple types of structures, in order to avoid confusion, the present application divides the training sample into N subsamples, and each subsample is user behavior data for various structures displayed in the conversation process of the conversation. In addition, the model can be realized by adopting DNN (Deep Neural Networks), for example, and the training can learn the model parameters by adopting error backpropagation. In addition, the present application is not limited to the algorithm and training method of the model.
[0038] Figure 2B is a network structure diagram of a conversation-level structure preference model according to an embodiment of the present application. As an example, as shown in the figure, the model extracts a user's interest indication based on the embedding of input past click structure list and aggregated past click entry list, extracts a user's non-interest indication based on the embedding of past non-click structure list and aggregated past non-click entry list, and further combines the embedding of user's scene features to obtain a user indication, and finally combines the embedding of structure features to predict the user's structure preference.
[0039] S202: Based on a recall algorithm, directionally recall the contents in the corresponding structures according to various structures and their weights in the target structure preferences.
[0040] In the prior art, in the content recommendation method based on the hierarchical screening mechanism, the recall module policy performs recall by matching the user vector with the content vector in the full resource library based on the past user behavior and the deep learning model. In the embodiment of the present application, the user's target structure preference is predicted and calculated in advance before recall, so that the recall policy can be upgraded to structured recall, that is, the system directionally recalls the content in the corresponding structure based on the target structure preference and its weight predicted by the system. Among them, the target structure preference includes a weighted indication of the user's preference for various structures, and based on the weight indication and a predetermined threshold, a predetermined number of top structures can be selected as the target of recall, so that the structured recall can focus more computing power on the structures that the user is really interested in, provide sufficient candidate content for each structure in the target structure preference, and recall better resource content in the user's preferred structure. Note that the embodiment of the present application does not limit the recall algorithm.
[0041] S203: Sorting the recalled contents in the various structures based on a sorting algorithm.
[0042] That is, the recalled contents are sorted according to various structures in the target structure preference, where the embodiment of the present application does not limit the sorting algorithm.
[0043] S204: Based on the weights of various structures, determine the positions of various structures in the recommendation results, and according to the sorting results of the contents recalled by various structures, fill the corresponding contents in various structures into the corresponding structures, and obtain the recommendation results after filling.
[0044] The structured fusion achieved by S204 allows content to be delivered to the user using the structure that best matches the user's preferences, and allows the optimal resources to which each structure belongs to to be displayed.
[0045] Based on the technical solutions of the embodiments of the present application, a global structured display and interest inference can be realized by determining a user's conversation-level structure preferences, and each time a conversation is started, the recommendation results are determined based on the conversation-level structure preferences, so that the recommendation results have corresponding recommended content for various different structures, thereby providing recommended content that is more in line with the user's preferences, and improving the matching degree between content provision and user preferences.
[0046] 3A is a schematic flow chart of a content recommendation method according to an embodiment of the present application. This embodiment further performs optimization based on the above embodiment. As shown in FIG. 3A, the method specifically includes S301 to S305.
[0047] S301: In response to a conversation start request, a current conversation level structure preference of a user is determined based on past behavior data of the user in the past conversation process and the structure characteristics of the content to be recommended.
[0048] S302: In response to every refresh request in the conversation, determine a user's current refresh level structural preference based on the conversation level structural preference and the user's structured feedback information, and set the refresh level structural preference as a target structural preference.
[0049] In one conversation, the server sends the recommendation results to the client in multiple times, and the user's operations such as scrolling up and down on the client trigger interactions with the server, which is called one refresh. Generally, one conversation is composed of multiple consecutive refreshes, and the next refresh is adjusted in real time based on the user's feedback on the previous refresh. That is, based on the conversation level structure preference predicted in this conversation, the conversation level structure preference and the user's structured feedback information are combined to determine the refresh level structure preference for each refresh, thereby realizing a more fine-grained prediction of the user's structure preference.
[0050] Among them, the structured feedback information is used to represent feedback information of different structures obtained by statistically analyzing user behavior data within a past set period including the previous refresh process. That is, based on behavior data in the previous refresh process and behavior data in other refresh processes within a set period (e.g., one week), feedback information of different structures is statistically analyzed, and examples of the statistical values include the number of views or click rates of different structures.
[0051] In one embodiment, an evolutionary learning method can be used to predict the refresh level structure preference. For example, the conversation level structure preference and the structured feedback information are taken as model inputs, and the evolutionary learning model is used to adjust the conversation level structure preference and output the refresh level structure preference. The examples in this application do not limit the structure and algorithm of the model.
[0052] In addition, the second scene feature can be further taken into consideration, that is, the user's current refresh level structure preference is determined according to the dialogue level structure preference, the second scene feature, and the user's structured feedback information, where the second scene feature is used to represent the scene to which each refresh belongs. The combination of the scene features can predict a more accurate and comprehensive structure preference. The embodiment of the present application does not limit the way of dividing the scenes.
[0053] Figure 3B is a schematic diagram of the refresh level structure preference model according to the embodiment of the present application, where after obtaining the refresh level structure score, the structure dimension can further be post-processed, searched and normalized according to business needs, and the final refresh level structure preference is obtained.
[0054] S303: Based on a recall algorithm, directionally recall content in the corresponding structure based on various structures and their weights in the target structure preferences.
[0055] S304: Sorting the contents in the various recalled structures based on a sorting algorithm.
[0056] S305: Based on the weights of the various structures, determine the positions of the various structures in the recommendation results; based on the sorting results of the contents recalled in the various structures, fill the corresponding contents in the various structures into the corresponding structures, and obtain the recommendation results after filling.
[0057] The technical solution of the embodiment of the present application determines a conversation level structure preference by setting up a two-level structured display, and then further determines a refresh level structure preference at each refresh during the conversation based on the conversation level structure preference, and estimates the user's preferences at a finer granularity, so that recommended content that is more in line with the user's current structure preference can be obtained during each refresh process, and the matching degree between content provision and user preferences is further improved.
[0058] FIG. 4 is a structural schematic diagram of a content recommendation device according to an embodiment of the present application, which can be applied to recommending content to a user, such as recommending content to a user through a feed system, and relates to smart recommendation, deep learning and big data technologies. The device can realize the content recommendation method described in any embodiment of the present application. As shown in FIG. 4, the device 400 specifically includes: A structure preference determining module 401 for determining a target structure preference of a user based on the past behavior data of the user using a product and the structure characteristics of the recommended content, the structure being determined by classifying the recommended content based on a classification method of an arbitrary content tag system; and a recommendation result determining module 402 for determining a recommendation result for each user according to the target structure preference, the recommendation result including at least two kinds of structures and recommended content corresponding to each kind of structure.
[0059] Optionally, the structural preference determination module further comprises: The method includes a conversation level structure preference determining unit for determining a current conversation level structure preference of the user based on past behavior data in the user's past conversation process and the structural characteristics of the content to be recommended in response to a conversation start request, and setting the conversation level structure preference as the target structure preference.
[0060] Optionally, the past behavior data includes previously selected structure content, previously selected entry content, previously non-selected structure content, and previously non-selected entry content.
[0061] Optionally, the structure characteristics are represented by a structure name and a structure attribute for representing past click-view information of different structures.
[0062] Optionally, the conversation level structure preference determining unit specifically comprises: The system is configured to determine the user's current conversation level structure preference based on past behavioral data in the user's past conversation process, a first scene feature for representing the scene to which each conversation belongs, and the structural features of the content to be recommended.
[0063] Optionally, the conversation level structure preference determining unit specifically further comprises: The method is configured to use a pre-trained conversation level structure preference model to determine the user's current conversation level structure preference based on the past behavior data of the user's past conversation process and the structure characteristics of the recommended content; The learning goal of the conversation-level structure preference model includes a user's preference for any structure in a conversation.
[0064] Optionally, the process of training the conversation level structure preference model comprises: Input a plurality of training samples as a model, each training sample being user behavior data generated in one conversation process, and each training sample includes N sub-samples, each sub-sample being user behavior data of various structures displayed in the conversation process, where N is a natural number; training the discourse-level structure preference model using the click rates of various structures for each pre-annotated training sample as model output.
[0065] Optionally, the structural preference determination module further comprises: a refresh level structure preference determining unit configured to determine a current refresh level structure preference of a user based on the conversation level structure preference and the user's structured feedback information in response to each refresh request in the conversation, and set the refresh level structure preference as the target structure preference; The structured feedback information is used to represent feedback information of different structures obtained statistically based on user behavior data within a past set period including the previous refresh process.
[0066] Optionally, the refresh level structure preference determining unit specifically comprises: The conversation level structure preference and the structured feedback information are used as model inputs, and an evolutionary learning model is used to adjust the conversation level structure preference and output the refreshment level structure preference.
[0067] Optionally, the refresh level structure preference determining unit specifically further comprises: The apparatus is configured to determine a current refresh level structure preference of the user based on the dialogue level structure preference, a second scene feature for representing a scene to which each refresh belongs, and the user's structured feedback information.
[0068] Optionally, the apparatus further comprises: The recommendation result determining module further includes a searching module configured to perform searching of structure dimensions based on the target structure preferences before determining the recommendation result for each user based on the target structure preferences.
[0069] Optionally, the recommendation result determining module is specifically Based on the target structure preferences, the recommendation result for each user is determined by recall, sorting and fusion.
[0070] Optionally, said target structure preferences include weighted representations of user preferences for various structures; Accordingly, the recommendation result determination module: A recall unit configured to directionally recall content in a corresponding structure according to various structures and their weights in the target structure preferences based on a recall algorithm; A sorting unit configured to sort the contents in the various structures of the recall based on a sorting algorithm; and a fusion unit configured to determine the positions of the various structures in the recommendation results based on the weights of the various structures, and to fill the corresponding contents in the various structures into the corresponding structures based on the sorting results of the recalled contents in the various structures, and obtain the recommendation results after filling.
[0071] The content recommendation device 400 provided in the embodiment of the present application can execute the content recommendation method provided in any embodiment of the present application, and has corresponding functional modules and beneficial effects for executing the method. For contents not described in detail in this embodiment, please refer to the description in any method embodiment of the present application.
[0072] According to an embodiment of the present application, the present application further provides an electronic device and a readable storage medium.
[0073] As shown in FIG. 5, a block diagram of an electronic device for implementing a content recommendation method according to an embodiment of the present application is shown. It represents various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers and other suitable computers. The electronic device may also represent various forms of mobile devices, such as personal digital processing, mobile phones, smartphones, wearable devices and other similar computing devices. Note that the components shown here, their connections and their functions are merely exemplary and are not intended to limit the embodiments of the present application described and / or claimed herein.
[0074] As shown in FIG. 5, the electronic device includes one or more processors 501, memory 502, and interfaces (including high-speed and low-speed interfaces) for connecting each component. Each component is connected to each other by different buses, and may be implemented on a common motherboard, or may be implemented in other ways as necessary. The processor can process instructions executed in the electronic device, including storing instructions in or on a memory to display graphic information of a graphical user interface (GUI) on an external input / output device such as a display device coupled to the interface. In other embodiments, multiple processors and / or multiple buses and multiple memories can be used along with multiple memories as necessary. Also, multiple electronic devices may be connected, with each device providing some required operations, such as a server array, a blade server group, or a multiprocessor system. In FIG. 5, one processor 501 is taken as an example.
[0075] The memory 502 is a non-transitory computer-readable storage medium provided by the present application, where the memory stores instructions executable by at least one processor, thereby causing the at least one processor to perform the content recommendation method provided by the present application. The non-transitory computer-readable storage medium of the present application stores computer instructions, which are used to cause a computer to perform the content recommendation method provided by the present application.
[0076] The memory 502, as a non-transitory computer-readable storage medium, can be used to store non-transitory software programs, non-transitory computer-executable programs and modules, such as program instructions / modules corresponding to the content recommendation method in the embodiment of the present application (for example, the structure preference determining module 401 and the recommendation result determining module 402 shown in FIG. 4). The processor 501 executes the non-transitory software programs, instructions and modules stored in the memory 502 to perform various functional applications and data processing of the server, that is, to realize the content recommendation method in the embodiment of the method.
[0077] The memory 502 may include a program storage area capable of storing an operating system, an application required for at least one function, and a data storage area capable of storing data generated in response to use of the electronic device that executes the content recommendation method of the embodiment of the present application. The memory 502 may also include a high-speed random access memory, and may also include a non-transitory memory (e.g., at least one magnetic disk storage device, a flash memory device, or other non-transitory solid-state storage device). In some embodiments, the memory 502 may optionally include a memory that is remotely located with respect to the processor 501, and these remote memories may be connected to the electronic device for implementing the content recommendation method of the embodiment of the present application via a network. Examples of the network include, but are not limited to, the Internet, a corporate intranet, a local area network, a mobile communication network, and combinations thereof.
[0078] The electronic device for implementing the content recommendation method of the embodiment of the present application may further include an input device 503 and an output device 504. The processor 501, the memory 502, the input device 503, and the output device 504 may be connected by a bus or other method, and FIG. 5 shows an example in which they are connected by a bus.
[0079] The input device 503, such as a touch panel, keypad, mouse, track pad, touch pad, pointing device, one or more mouse buttons, track ball, joystick, etc., can receive input numeric and character information and generate key signal inputs related to user settings and function control of the electronic device for the content recommendation method of the present application embodiment. The output device 504 can include a display device, auxiliary lighting devices (e.g., LEDs), and haptic feedback devices (e.g., vibration motors), etc. The display device can include, but is not limited to, a liquid crystal display (LCD), a light emitting diode (LED) display, and a plasma display. In some embodiments, the display device can be a touch panel.
[0080] Various embodiments of the systems and techniques described herein may be realized in digital electronic circuitry systems, integrated circuit systems, application specific integrated circuits (ASICs), computer hardware, firmware, software, and / or combinations thereof. These various embodiments may be implemented in one or more computer programs that may be executed and / or interpreted in a programmable system that includes at least one programmable processor, which may be a special purpose or general purpose programmable processor, and may include receiving data and instructions from a storage system, at least one input device, and at least one output device, and transmitting data and instructions to the storage system, the at least one input device, and the at least one output device.
[0081] These computer programs, which may also be referred to as programs, software, software applications or code, include machine instructions for a programmable processor and may be implemented using high level process and / or object oriented programming languages and / or assembly / machine language. As used herein, the terms "machine-readable medium" and "computer-readable medium" refer to any computer program product, apparatus, and / or device (e.g., magnetic disk, optical disk, memory, programmable logic device (PLD)) for providing machine instructions and / or data to a programmable processor, including a machine-readable medium for receiving machine instructions that are machine-readable signals. The term "machine-readable signal" refers to any signal for providing machine instructions and / or data to a programmable processor.
[0082] To provide interaction with a user, the systems and techniques described herein can be implemented on a computer that includes a display device (e.g., a Cathode Ray Tube (CRT) or LCD (Liquid Crystal Display) monitor) for displaying information to a user, and a keyboard and pointing device (e.g., a mouse or trackball) through which a user can provide input to the computer. Other types of devices can also be used to provide interaction with a user. For example, the feedback provided to the user can be any form of sensing feedback, e.g., visual feedback, auditory feedback, or haptic feedback, and can receive input from the user in any form, including sound input, speech input, or tactile input.
[0083] The systems and techniques described herein may be implemented in a computing system including background components (e.g., a data server), or may be implemented in a computing system including middleware components (e.g., an application server), or may be implemented in a computing system including front-end components (e.g., a user computer having a graphical user interface or web browser) through which a user may interact with the embodiments of the systems and techniques described herein, or may be implemented in a computing system including any combination of such background components, middleware components, or front-end components. The components of the system may also be connected by digital data communication through any form or medium, such as a communication network. Examples of communication networks include a local area network (LAN), a wide area network (WAN), the Internet, a blockchain network, and the like.
[0084] The computer system may include a client and a server. The client and the server are usually remote from each other and interact with each other via a communication network. The relationship between the client and the server is generated by running computer programs having a client-server relationship with each other on the respective computers. The server may be a cloud server, also called a cloud computing server or cloud host, which is a host product in a cloud computing service system and solves the defects of the traditional physical host and VPS service, such as high management difficulty and low business scalability.
[0085] Based on the technical solutions of the embodiments of the present application, a global structured display and interest inference can be realized by determining a user's structural preferences, and then the recommendation results are determined based on the structural preferences, so that the recommendation results have corresponding recommended content for various different structures, thereby providing recommended content that is more in line with the user's preferences, and improving the matching degree between content provision and user preferences.
[0086] It should be noted that the above-mentioned various forms of flow may be used to rearrange, add, or delete steps. For example, the steps described in this application may be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solution disclosed in this application are achieved. This specification is not limited thereto.
[0087] The above specific embodiments do not limit the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, subcombinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and principle of the present application should be included within the scope of protection of the present application.
Claims
1. determining a target structure preference of a user based on past behavioral data of the user's use of a product, structural characteristics of the recommended content, and the user's business needs, wherein the structure of the recommended content is determined by classifying the recommended content based on a classification method of an arbitrary content tag system, and the structural characteristics of the recommended content are represented by the name of the structure and structural attributes for representing the structure's past click / display information, and the target structure preference is determined by using a score obtained by predicting the user's interest in various structures as a predicted value and selecting a predetermined number of structures that are at the top of all predicted values; For each user, determining a recommendation result for the user based on the target structure preference of the user, in which the determination of the recommendation result for the user is performed such that, for each structure in the target structure preference of the user, recall, sorting and merging are performed to obtain recommended content for the structure, and the recommended content for all structures in the target structure preference is set as the recommendation result for the user. A method performed by a content recommendation device.
2. Determining the user's target structure preference based on the user's past product use behavior data, the structure characteristics of the recommended content, and the user's business needs; A method executed by the content recommendation device of claim 1, in response to a conversation start request, determining the user's current conversation level structure preference based on past behavior data in the user's past conversation process and the structural characteristics of the content to be recommended, and setting the conversation level structure preference as the target structure preference.
3. The method performed by the content recommendation device of claim 2, wherein the past behavioral data includes previously selected structural content, previously selected entry content, previously not selected structural content, and previously not selected entry content.
4. The method of determining the user's current conversation level structure preference based on the past behavior data of the user's past conversation process and the structure characteristics of the recommended content includes: A method performed by the content recommendation device of claim 2, comprising determining the user's current conversation level structure preference based on past behavioral data in the user's past conversation process, a first scene feature for representing the scene to which each conversation belongs, and structural features of the content to be recommended.
5. The method of determining the user's current conversation level structure preference based on the past behavior data of the user's past conversation process and the structure characteristics of the recommended content includes: Using a pre-trained conversation level structure preference model, determining the user's current conversation level structure preference according to the user's past behavior data in the past conversation process and the structure characteristics of the recommended content; The method performed by the content recommendation device of claim 2 , wherein the learning goal of the conversation-level structure preference model includes a user's preference for any structure in a conversation.
6. The training process of the conversation level structure preference model includes: Input a plurality of training samples as a model, each training sample being user behavior data generated in one conversation process, and each training sample includes N sub-samples, each sub-sample being user behavior data of various structures displayed in one conversation process corresponding to the training sample, where N is a natural number; 6. The method performed by the content recommendation device of claim 5, further comprising: training the conversation-level structure preference model using click rates of various structures for each pre-annotated training sample as output of the model.
7. Determining the user's target structure preference based on the user's past product use behavior data, the structure characteristics of the recommended content, and the user's business needs; and determining a current refresh level structure preference of the user based on the conversation level structure preference and the user's structured feedback information in response to each refresh request within the user's current conversation, and setting the refresh level structure preference as the target structure preference; The method performed by the content recommendation device according to claim 2, wherein the structured feedback information is used to represent feedback information of different structures obtained statistically based on user behavior data within a past set period including a previous refresh process.
8. determining a current refresh level structure preference of a user based on the conversation level structure preference and the user's structured feedback information, comprising: The method performed by the content recommendation device of claim 7, further comprising: taking the conversation level structure preference and structured feedback information as model inputs, utilizing an evolutionary learning model to adjust the conversation level structure preference, and outputting the refresh level structure preference.
9. determining a current refresh level structure preference of a user based on the conversation level structure preference and the user's structured feedback information, comprising: The method performed by the content recommendation device of claim 7, further comprising determining a user's current refresh level structure preference based on the conversation level structure preference, a second scene feature for representing the scene to which each refresh belongs, and the user's structured feedback information.
10. Before determining the recommendation result for each user based on the target structure preference, The method performed by the content recommendation device of claim 1 , further comprising: performing a search of structural dimensions based on the target structural preferences.
11. the target structure preferences include weighted representations of user preferences for various structures; determining, for each user, a recommendation result for each user based on the target structure preference of the user; Directionally recalling content in the corresponding structure according to various structures and their weights in the target structure preferences based on a recall algorithm; Sorting the contents of the various structures of the recall based on a sorting algorithm; The method performed by the content recommendation device of claim 1, including: determining the positions of various structures in the recommendation results based on the weights of the various structures; and filling the corresponding structures with the content in the various structures based on the sorting results of the content recalled in the various structures, and obtaining the recommendation results after filling.
12. a structure preference determination module for determining a target structure preference of a user based on past behavioral data of the user using a product, structural characteristics of the recommended content, and the user's business needs, the structure of the recommended content being determined by classifying the recommended content based on a classification method of an arbitrary content tag system, the structural characteristics of the recommended content being represented by a structure name and a structural attribute for representing the structure's past click / display information, and the target structure preference being determined by using a score obtained by predicting the user's interest level for various structures as a predicted value and selecting a predetermined number of structures that are at the top of all predicted values; A content recommendation device including: a recommendation result determination module that determines, for each user, a recommendation result for the user based on the user's target structure preference, wherein the determination of the recommendation result for the user is performed by performing recall, sorting, and merging for each structure in the user's target structure preference to obtain recommended content for that structure, and the recommended content for all structures in the target structure preference is set as the recommendation result for the user.
13. The structural preference determination module includes: The content recommendation device according to claim 12, further comprising a conversation level structure preference determining unit for determining a current conversation level structure preference of the user based on past behavior data in the user's past conversation process and the structural characteristics of the content to be recommended in response to a conversation start request, and making the conversation level structure preference the target structure preference.
14. The content recommendation device of claim 13 , wherein the past behavior data includes previously selected structure content, previously selected entry content, previously not selected structure content, and previously not selected entry content.
15. The conversation level structure preference determining unit includes: The content recommendation device of claim 13, configured to determine the user's current conversation level structure preference based on past behavioral data in the user's past conversation process, a first scene feature for representing the scene to which each conversation belongs, and structural features of the content to be recommended.
16. The conversation level structure preference determining unit further comprises: The method is configured to use a pre-trained conversation level structure preference model to determine the user's current conversation level structure preference based on the past behavior data of the user's past conversation process and the structure characteristics of the recommended content; The content recommendation device according to claim 13 , wherein the learning goal of the conversation-level structure preference model includes a preference for any structure in a user's conversation.
17. The training process of the conversation level structure preference model includes: Input a plurality of training samples as a model, each training sample being user behavior data generated in one conversation process, and each training sample includes N sub-samples, each sub-sample being user behavior data of various structures displayed in one conversation process for the training sample, where N is a natural number; and training the conversation-level structure preference model using click-through rates of various structures for each pre-annotated training sample as output of the model.
18. The structural preference determination module includes: a refresh level structure preference determining unit configured to determine a current refresh level structure preference of the user based on the conversation level structure preference and the user's structured feedback information in response to each refresh request in the user's current conversation, and set the refresh level structure preference as the target structure preference; The content recommendation device according to claim 13 , wherein the structured feedback information is used to represent feedback information of different structures obtained statistically based on user behavior data within a past set period including a previous refresh process.
19. The refresh level structure preference determining unit includes: The content recommendation device of claim 18 , configured to take the conversation level structure preference and structured feedback information as model inputs, utilize an evolutionary learning model, adjust the conversation level structure preference, and output the refresh level structure preference.
20. The refresh level structure preference determining unit further comprises: The content recommendation device of claim 18, configured to determine a user's current refresh level structure preference based on the conversation level structure preference, a second scene feature for representing a scene to which each refresh belongs, and the user's structured feedback information.
21. The content recommendation device includes: The content recommendation device of claim 12 , further comprising a search module configured to perform a search of a structural dimension based on the target structural preference before the recommendation result determination module determines the recommendation result for each user based on the target structural preference.
22. the target structure preferences include weighted representations of user preferences for various structures; The recommendation result determination module includes: A recall unit configured to directionally recall content in a corresponding structure according to various structures and their weights in the target structure preferences based on a recall algorithm; A sorting unit configured to sort the contents in the various structures of the recall based on a sorting algorithm; The content recommendation device of claim 13, further comprising: a fusion unit configured to determine the positions of the various structures in the recommendation result based on the weights of the various structures; and to fill the corresponding contents in the various structures into the corresponding structures based on the sorting results of the contents recalled in the various structures, and obtain the recommendation result after filling.
23. At least one processor; and a memory communicatively coupled to the at least one processor, An electronic device, wherein the memory stores instructions executable by the at least one processor, and when the instructions are executed by the at least one processor, the at least one processor is caused to execute a method performed by a content recommendation device described in any one of claims 1 to 11.
24. A non-transitory computer-readable storage medium having computer instructions stored thereon, comprising: A non-transitory computer readable storage medium in which the computer instructions are used to cause a computer to perform the method performed by the content recommendation device according to any one of claims 1 to 11.
25. A computer program product which, when executed by a processor, implements the method performed by a content recommendation device according to any one of claims 1 to 11.
Citation Information
Patent Citations
Personalized content recommendation method and device and storage medium
CN111209470A
Electronic program chart display device, medium and information assembly
JP2001285744A
Program information providing device, program information providing method and its program
JP2007060398A
Recommendation information evaluation device and recommendation information evaluation method
JP2009252179A
Display device, remote operating set, and display program
JP2010041163A
