Model training method, content recommendation method, device, medium, equipment and product

Abstract

The application discloses a model training method, a content recommendation method, a device, a medium, equipment, and a product. The live broadcast content recommendation method comprises the following steps: obtaining a plurality of live broadcast behavior feature encodings of a target object, wherein the plurality of live broadcast behavior feature encodings are used for representing behavior features of the target object for pre-live exposure content and behavior features of the target object for actual live broadcast content; determining a first live broadcast preference degree and a second live broadcast preference degree of the target object by taking the plurality of live broadcast behavior feature encodings as inputs of a preset preference determination model, and determining a comprehensive live broadcast preference degree of the target object according to the first live broadcast preference degree and the second live broadcast preference degree, wherein the first live broadcast preference degree is used for representing a preference degree of the target object for the pre-live exposure content, and the second live broadcast preference degree is used for representing a preference degree of the target object for the actual live broadcast content; and performing live broadcast content recommendation on the target object according to the comprehensive live broadcast preference degree of the target object.

Description

Technical Field

[0001] The present application relates to the field of computer technology, and specifically to a model training method, content recommendation method, device, medium, equipment, and product. Background Art

[0002] With the continuous development of Internet technology, many emerging professions have emerged, such as the live broadcast industry.

[0003] Currently, many software applications are beginning to venture into the live streaming business, such as music software applications. Related technologies typically use deep learning models to calculate user preferences for live streaming and live streamers based on user profiles and live streaming behavior data. Based on this preference, live streaming recommendations are then made. However, on many software platforms, the vast majority of users have never experienced live streaming, resulting in sparse live streaming behavior data. This makes it impossible to accurately predict user preferences for live streaming, making it difficult to make live streaming recommendations. Summary of the Invention

[0004] The embodiments of the present application provide a model training method, content recommendation method, apparatus, medium, equipment, and product that can overcome the problem of sparse user live broadcast behavior data, thereby accurately predicting the user's preference for live broadcast, and further improving the accuracy of live broadcast recommendations to users.

[0005] In a first aspect, a method for training a preference determination model is provided, the method further comprising:

[0006] Obtaining a training sample set, where each piece of training sample data in the training sample set includes at least one training sample object, and a plurality of historical live broadcast behavior features, a first live broadcast preference true value label, and a second live broadcast preference true value label for each training sample object in the at least one training sample object;

[0007] Inputting target training sample data into the preference determination model, wherein the target training sample data is any training sample data in the training sample set;

[0008] Obtaining a first predicted live broadcast preference of a corresponding target training sample object and a second predicted live broadcast preference of the target training sample object determined by the preference determination model according to the target training sample data;

[0009] According to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object, the model parameters of the preference determination model are trained to obtain a trained preference determination model.

[0010] Optionally, before obtaining the training sample set, the method further includes:

[0011] Acquire historical live broadcast behavior data of a target user, wherein the historical live broadcast behavior data is live broadcast behavior data of the target object within a preset historical time period, and the historical live broadcast behavior data includes data of multiple dimensions;

[0012] Determine a first live broadcast preference truth value label of the target object based on data of at least one first dimension preset among the multiple dimensions, and determine a second live broadcast preference truth value label of the target object based on the first live broadcast preference truth value label and data of at least one second dimension preset among the multiple dimensions, wherein the first live broadcast preference truth value label is used to characterize the target object's preference for content exposed before the live broadcast, and the second live broadcast preference truth value label is used to characterize the target object's preference for actual content of the live broadcast;

[0013] Determining multiple historical live broadcast behavior features of the target object based on data of each dimension of the multiple dimensions;

[0014] A training sample set is constructed according to the target object, multiple historical live broadcast behavior features of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object.

[0015] Optionally, the determining of a second live broadcast preference true value label of the target object based on the first live broadcast preference true value label and data of at least one second dimension preset in the multiple dimensions:

[0016] When the first live broadcast preference truth value label is a preset value, the second live broadcast preference truth value label of the target object is determined based on data of at least one second dimension preset in the multiple dimensions.

[0017] Optionally, determining the target guaranteed price for the traffic corresponding to the target object based on multiple bidding records of the demand-side platform for the traffic corresponding to all objects in the target value level within the first preset time period further includes:

[0018] determining an overall average value of all objects in the target value hierarchy based on a third historical bid set of the demand-side platform for all objects in the target value hierarchy within the first preset time period, the third historical bid set including a plurality of bid records of the demand-side platform for traffic corresponding to all objects in the target value hierarchy within the first preset time period;

[0019] The target guaranteed price for the traffic corresponding to the target object is determined based on the overall value average.

[0020] Optionally, constructing a training sample set based on the target object, multiple historical live broadcast behavior features of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object includes:

[0021] Converting continuous features in the plurality of historical live broadcast behavior features of the target object into discrete features;

[0022] Encoding the discrete features in the multiple historical live broadcast behavior features of the target object and the converted discrete features to obtain the historical live broadcast behavior data feature code of the target object;

[0023] A piece of training sample data in the training sample set is determined based on the target object, the historical live broadcast behavior data feature code of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object.

[0024] Optionally, determining a piece of training sample data in the training sample set based on the target object, the feature code of the target object's historical live broadcast behavior data, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object includes:

[0025] Determine the weight of the training sample data corresponding to the target object according to the preset business goal;

[0026] A piece of training sample data in the training sample set is determined based on the target object, the historical live broadcast behavior data feature code of the target object, the first live broadcast preference true value label of the target object, the second live broadcast preference true value label of the target object, and the weight of the training sample data corresponding to the target object.

[0027] Optionally, training model parameters of the preference determination model based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object to obtain a trained preference determination model includes:

[0028] Determining a target loss function according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object;

[0029] The preference determination model is trained according to the target loss function to obtain the trained preference determination model.

[0030] Optionally, determining a target loss function based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object includes:

[0031] The target loss function is determined based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, the second live broadcast preference true value label of the target training sample object and the weight of the training sample data corresponding to the target training sample.

[0032] In a second aspect, a live content recommendation method is provided, the method comprising:

[0033] Acquire multiple live broadcast behavior feature codes of a target object, where the multiple live broadcast behavior feature codes are used to characterize the behavior characteristics of the target object with respect to the content exposed before the live broadcast and the behavior characteristics of the target object with respect to the actual content of the live broadcast;

[0034] Based on a preset preference determination model, the plurality of live broadcast behavior feature codes are used as inputs to the preset preference determination model to determine a first live broadcast preference and a second live broadcast preference of the target object, and a comprehensive live broadcast preference of the target object is determined based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to represent the target object's preference for content exposed before the live broadcast, and the second live broadcast preference is used to represent the target object's preference for the actual content of the live broadcast;

[0035] Recommend live broadcast content to the target object based on the target object's comprehensive live broadcast preference.

[0036] Optionally, the preset preference determination model includes an embedding module, a first preference determination module, and a second preference determination module;

[0037] The method of determining the first live broadcast preference and the second live broadcast preference of the target object based on a preset preference determination model and taking the plurality of live broadcast behavior feature codes as inputs of the preset preference determination model includes:

[0038] Controlling the embedding module to convert the plurality of live broadcast behavior feature codes into a plurality of continuous feature vectors of preset dimensions respectively;

[0039] Controlling the first preference determination module to determine a first live broadcast preference of the target object according to the multiple continuous feature vectors;

[0040] The second preference determination module is controlled to determine a second live broadcast preference of the target object according to the multiple continuous feature vectors.

[0041] Optionally, controlling the first preference determination module to determine the first live broadcast preference of the target object according to the multiple continuous feature vectors includes:

[0042] Control the first preference determination module to calculate the first live broadcast preference of the target object based on the multiple continuous feature vectors and the weight number corresponding to each continuous feature vector in the multiple continuous feature vectors in the first preference determination module determined in the preset preference determination model.

[0043] Optionally, the second determination module can also be used to: determine the overall value average of all objects in the target value level based on the third historical bid set of all objects in the target value level by the demand-side platform within the first preset time period, the third historical bid set including multiple bid records of the demand-side platform for the corresponding traffic of all objects in the target value level within the first preset time period; determine the target floor price of the traffic corresponding to the target object based on the overall value average.

[0044] Optionally, controlling the second preference determination module to determine the second live broadcast preference of the target object according to the multiple continuous feature vectors includes:

[0045] Control the second preference determination module to calculate the second live broadcast preference of the target object based on the multiple continuous feature vectors and the weight number corresponding to each continuous feature vector in the multiple continuous feature vectors in the second preference determination module determined in the preset preference determination model.

[0046] Optionally, determining the comprehensive live broadcast preference of the target object according to the first live broadcast preference and the second live broadcast preference includes:

[0047] Calculating the product of the first live broadcast preference and the second live broadcast preference;

[0048] The product is determined as the comprehensive live broadcast preference of the target object.

[0049] Optionally, recommending live broadcast content to the target object based on the comprehensive live broadcast preference includes:

[0050] determine, based on the preset preference determination model, a comprehensive live broadcast preference degree of all objects containing the target object in the target software;

[0051] sort the all objects according to the comprehensive live broadcast preference degree of the all objects;

[0052] recommend live broadcast content to the target object according to the position of the target object in the sorting.

[0053] In a third aspect, a device for training a preference determination model is provided, and the device includes:

[0054] a first obtaining module, configured to obtain a training sample set, each piece of training sample data in the training sample set including at least one training sample object, a plurality of historical live broadcast behavior features of each training sample object in the at least one training sample object, a first live broadcast preference degree true value label, and a second live broadcast preference degree true value label;

[0055] an input module, configured to input target training sample data into the preference determination model, the target training sample data being any piece of training sample data in the training sample set;

[0056] a second obtaining module, configured to obtain a first predicted live broadcast preference degree of a corresponding target training sample object and a second predicted live broadcast preference degree of the target training sample object, which are determined by the preference determination model based on the target training sample data;

[0057] a training module, configured to train model parameters of the preference determination model based on the first predicted live broadcast preference degree of the target training sample object, the second predicted live broadcast preference degree of the target training sample object, and the first live broadcast preference degree true value label of the target training sample object and the second live broadcast preference degree true value label of the target training sample object, to obtain a trained preference determination model.

[0058] In a fourth aspect, a device for recommending live broadcast content is provided, and the device includes:

[0059] an obtaining module, configured to obtain a plurality of live broadcast behavior feature encodings of a target object, the plurality of live broadcast behavior feature encodings being used to represent behavior features of the target object with respect to pre-live exposure content and behavior features of the target object with respect to actual live broadcast content;

[0060] a determination module for determining, based on a preset preference determination model and using the plurality of live broadcast behavior feature codes as inputs to the preset preference determination model, a first live broadcast preference and a second live broadcast preference of the target subject, and determining a comprehensive live broadcast preference of the target subject based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to represent the target subject's preference for content exposed before the live broadcast, and the second live broadcast preference is used to represent the target subject's preference for actual live broadcast content;

[0061] The recommendation module is used to recommend live broadcast content to the target object based on the comprehensive live broadcast preference of the target object.

[0062] In a fifth aspect, a computer-readable storage medium is provided, wherein the computer-readable storage medium stores a computer program, and the computer program is suitable for loading by a processor to execute the steps in the training method of the preference determination model as described in the first aspect above.

[0063] In a sixth aspect, a computer device is provided, comprising a processor and a memory, wherein a computer program is stored in the memory, and the processor is used to execute the steps in the training method of the preference determination model as described in the first aspect above by calling the computer program stored in the memory.

[0064] In a seventh aspect, a computer program product is provided, comprising computer instructions, which, when executed by a processor, implement the steps in the method for training a preference determination model as described in the first aspect above.

[0065] The embodiment of the present application obtains multiple live broadcast behavior feature codes of the target object, and the multiple live broadcast behavior feature codes are used to characterize the target object's behavior characteristics with respect to the content exposed before the live broadcast and the target object's behavior characteristics with respect to the actual content of the live broadcast; based on a preset preference determination model, the multiple live broadcast behavior feature codes are used as inputs of the preset preference determination model to determine the first live broadcast preference and the second live broadcast preference of the target object, and the comprehensive live broadcast preference of the target object is determined based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to characterize the target object's preference with respect to the content exposed before the live broadcast, and the second live broadcast preference is used to characterize the target object's preference with respect to the actual content of the live broadcast; and live broadcast content is recommended to the target object based on the comprehensive live broadcast preference of the target object. The embodiments of the present application can determine the target object's first live broadcast preference for the pre-live broadcast content and the second live broadcast preference for the actual live broadcast content based on the target object's behavioral characteristics for the pre-live broadcast content and the target object's behavioral characteristics for the actual live broadcast content, and combine the first live broadcast preference and the second live broadcast preference to determine the target object's comprehensive preference for live broadcast, thereby improving the accuracy of predicting the user's preference for live broadcast, and overcoming the problem in related technologies that the user's preference for live broadcast cannot be predicted due to the lack of user behavioral data for the actual live broadcast content, thereby improving the accuracy of live broadcast content recommendation based on the target object's preference for live broadcast. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0067] Figure 1 A flowchart of a method for training a preference determination model provided in an embodiment of the present application.

[0068] Figure 2 A flowchart of the live content recommendation method provided in an embodiment of the present application.

[0069] Figure 3 Schematic diagram of an application scenario of the live content recommendation method provided in an embodiment of the present application.

[0070] Figure 4 A schematic diagram of another application scenario of the live content recommendation method provided in an embodiment of the present application.

[0071] Figure 5 A schematic diagram of the structure of a training device for a preference determination model provided in an embodiment of the present application.

[0072] Figure 6 A schematic diagram of the structure of the live content recommendation device provided in an embodiment of the present application.

[0073] Figure 7 A schematic diagram of the structure of a computer device provided in an embodiment of the present application. DETAILED DESCRIPTION

[0074] The following will be combined with the drawings in the embodiments of this application to clearly and completely describe the technical solutions in the embodiments of this application. Obviously, the embodiments described are only part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without making creative efforts are within the scope of protection of this application.

[0075] The present invention provides a model training method, content recommendation method, apparatus, medium, device, and product. Specifically, the preference determination model training method and content recommendation method of the present invention can be executed by a computer device.

[0076] First, some nouns or terms that appear in the description of the embodiments of this application are explained as follows:

[0077] Artificial Intelligence (AI) is the theory, methods, techniques, and application systems that use digital computers or machines controlled by digital computers to simulate, extend, and expand human intelligence, to perceive the environment, acquire knowledge, and use that knowledge to achieve optimal results. In other words, AI is a comprehensive technology within computer science that seeks to understand the essence of intelligence and produce new intelligent machines that can respond in a manner similar to human intelligence. AI also studies the design principles and implementation methods of various intelligent machines, enabling them to possess the capabilities of perception, reasoning, and decision-making.

[0078] Machine Learning (ML) is a multidisciplinary field that encompasses probability theory, statistics, approximation theory, convex analysis, and algorithmic complexity theory. It specifically studies how computers can simulate or implement human learning behaviors to acquire new knowledge or skills and reorganize existing knowledge structures to continuously improve their performance. Machine learning is at the core of artificial intelligence and the fundamental way to make computers intelligent. Its applications span all areas of AI. Machine learning and deep learning typically include techniques such as artificial neural networks, belief networks, reinforcement learning, transfer learning, inductive learning, and learning through demonstration.

[0079] Deep learning (DL): A branch of machine learning, it is an algorithm that attempts to achieve high-level abstraction of data using multiple processing layers containing complex structures or multiple nonlinear transformations. Deep learning learns the inherent patterns and representational hierarchies of training sample data. The information gained from this learning process is highly helpful in interpreting data such as text, images, and sound. The ultimate goal of deep learning is to enable machines to have human-like analytical learning capabilities and to recognize data such as text, images, and sound. Deep learning is a complex machine learning algorithm that has achieved results in speech and image recognition that far surpass previous related technologies.

[0080] Neural Network (NN): A deep learning model in the field of machine learning and cognitive science that mimics the structure and function of biological neural networks.

[0081] One-hot encoding, also known as one-bit effective encoding, mainly uses an N-bit state register to encode N states. Each state has its own independent register bit, and only one bit is valid at any time.

[0082] Embedding is a technique that converts discrete variables into fixed-length continuous vector representations.

[0083] AvgPooling is a technique for finding the average of multiple vectors of the same length.

[0084] In related technologies, a user's preference for a particular anchor can be obtained by collecting the user's live broadcast behavior data and then calculating the similarity between the user portrait and the live broadcast room portrait. However, this method can only calculate the user's preference for a certain live broadcast room and a certain anchor, and cannot calculate the user's preference for live broadcasting. With the development of artificial intelligence technology, it is now possible to calculate the user's preference for live broadcast rooms and anchors based on deep learning models using user portraits and user live broadcast behavior data. However, based on the deep learning model approach, it is impossible to predict the preference for live broadcast rooms and anchors for non-live broadcast users who have not experienced live broadcasting and have no live broadcast behavior data, nor is it possible to calculate the user's preference for live broadcasting.

[0085] Therefore, the embodiments of the present application propose a training method for a preference determination model and a live broadcast content recommendation method, which can overcome the problem of sparse user live broadcast behavior data, thereby accurately predicting the user's preference for live broadcast, and further improving the accuracy of live broadcast recommendations to users.

[0086] It should be noted that the order of description of the following embodiments does not limit the priority order of the embodiments.

[0087] See also Figure 1 、 Figure 3 and Figure 4 The training method of the preference determination model mainly includes steps 101 to 104, which are described as follows:

[0088] Step 101: Obtain a training sample set, where each piece of training sample data in the training sample set includes at least one training sample object, and multiple historical live broadcast behavior features, a first live broadcast preference true value label, and a second live broadcast preference true value label of each training sample object in the at least one training sample object.

[0089] In some embodiments, before the step of "obtaining a training sample set", it also includes: obtaining historical live broadcast behavior data of the target user, the historical live broadcast behavior data is the live broadcast behavior data of the target object within a preset historical time period, and the historical live broadcast behavior data includes data of multiple dimensions; according to the data of at least one first dimension pre-set in the multiple dimensions, determining the first live broadcast preference true value label of the target object, and according to the first live broadcast preference true value label and the data of at least one second dimension pre-set in the multiple dimensions, determining the second live broadcast preference true value label of the target object, the first live broadcast preference true value label is used to characterize the target object's preference for the content exposed before the live broadcast, and the second live broadcast preference true value label is used to characterize the target object's preference for the actual content of the live broadcast; determining multiple historical live broadcast behavior features of the target object according to the data of each dimension in the multiple dimensions; constructing a training sample set according to the target object, the multiple historical live broadcast behavior features of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object.

[0090] The target user may be a user who uses the target software and authorizes the target software to access the historical usage data of the target software.

[0091] The preset historical time period can be customized as needed, for example, to the past 15 days, past 7 days, etc. Live streaming behavior data can include user behavior data related to the anchor, such as follow data, click data, viewing time data, and consumption data for each anchor. Specifically, the target user's usage log for the target software can be obtained to obtain their historical live streaming behavior data. For example, for each target user, the usage log can record the user's behavior data related to the anchor in the format of <user ID, anchor ID, follow data, ..., consumption data>.

[0092] Specifically, after obtaining the historical live broadcast behavior data of the target user in the anchor dimension, the historical live broadcast behavior data of the target user can be aggregated to calculate the historical live broadcast behavior data of the user dimension. Among them, the historical live broadcast behavior data of the user dimension can mainly include each target user's exposure PV, exposure UV, click PV, click UV, click rate PV, click rate UV, total viewing time, attention UV, fan group PV, fan group UV, number of gifts, gift UV, total gift amount, etc. Specifically, PV refers to the number of times, and UV refers to the number of people. For example, if the target user has clicked on the anchor's homepage three times, then the target user's click PV is 3. These three click data correspond to anchor A and anchor B, and they clicked anchor A once and anchor B twice respectively, so the click UV is 2.

[0093] Next, based on each target user's historical live streaming behavior data in the user dimension, the first live streaming preference truth value label and the second live streaming preference truth value label of each target user are determined. Specifically, the data of the at least one pre-set first dimension may include the target user's click-through volume (PV), and the data of the at least one pre-set second dimension may include data such as follow UV, fan group PC, viewing time, number of gifts sent, and gift amount.

[0094] Next, training sample features can be constructed based on the target user's historical live streaming behavior data, the target user's user profile data, the target user's live streaming-related behavior data using the target software, and user-unrelated live streaming general data. For example, the target user's historical live streaming behavior data can include each target user's exposure PV, exposure UV, click PV, click UV, click rate PV, click rate UV, total viewing time, follow UV, fan group PV, fan group UV, number of gifting, gifting UV, total gift amount, etc. For example, the target user's user profile data can include the target user's age, gender, city, city level, VIP type, VIP level, device type, operating system, etc. For example, the target user's live streaming-related behavior data using the target software can mainly include the target user's recent listening time, the list of song IDs listened to recently, the list of artist IDs listened to recently, etc. For example, user-unrelated live streaming general data can mainly include the number of live streamers on the platform, the average live stream duration of live streamers, the user's live streaming consumption time, the total user consumption time, the average user consumption time, etc.

[0095] It's easy to understand that the training sample features include not only the target user's historical live streaming behavior data, but also the target user's user profile data and the target user's live streaming-related behavior data when using the target software. This allows prediction of the target user's preference for live streaming even if the target user has never experienced live streaming-related services. For example, the target user's live streaming-related behavior data when using the target software includes a list of song IDs that the target user has recently listened to and a list of artist IDs that the target user has recently listened to. Suppose there are target users A and B. Target user A has never experienced live streaming-related services and lacks historical live streaming behavior data. However, the song ID lists and artist ID lists that target user B has recently listened to overlap significantly with those of target user A. Therefore, it can be assumed that target user B and target user A have the same preferences. Based on target user B's preference for anchors and live streaming, target user A's preference for anchors and live streaming can be inferred.

[0096] In this embodiment, the step of "determining the second live broadcast preference truth value label of the target object based on the first live broadcast preference truth value label and the data of at least one second dimension pre-set in multiple dimensions" includes: when the first live broadcast preference truth value label is a preset value, determining the second live broadcast preference truth value label of the target object based on the data of at least one second dimension pre-set in multiple dimensions.

[0097] Among them, the preset value can be 1, and the data of at least one second dimension pre-set in multiple dimensions can be customized as needed, for example, it can be viewing time, attention UV, fan group PV, number of gifts and total gift amount. It is easy to understand that the user's exposure, click behavior, and live consumption behavior have an obvious sequence relationship. Therefore, the target user's second live broadcast preference truth value label can be determined based on the target user's first live broadcast preference truth value label. For example, if the target user's click PV is greater than 0, then its first live broadcast preference is determined to be 1, otherwise it is 0. In the case of the first live broadcast preference being 1, if the target user meets one of the following conditions: viewing time greater than 30 seconds, attention UV greater than 0, fan group PV greater than 0, number of gifts greater than 0 and total gift amount greater than 0, then its second live broadcast preference truth value label can be determined to be 1, otherwise it is 0.

[0098] In this embodiment, the step of "constructing a training sample set based on the target object, multiple historical live broadcast behavior features of the target object, the target object's first live broadcast preference true value label, and the target object's second live broadcast preference true value label" includes: converting the continuous features in the target object's multiple historical live broadcast behavior features into discrete features; encoding the discrete features in the target object's multiple historical live broadcast behavior features and the converted discrete features to obtain the target object's historical live broadcast behavior data feature code; determining a training sample data in the training sample set based on the target object, the target object's historical live broadcast behavior data feature code, the target object's first live broadcast preference true value label, and the target object's second live broadcast preference true value label.

[0099] Discrete features are features with discontinuous, countable values. For example, gender only has three possible values: male, female, and unknown. Continuous features are features with a continuous range of values. For example, the number of gifts sent has a value between 0 and N, and the click-through rate has a decimal value between 0 and 1.

[0100] Specifically, the step of "converting continuous features in multiple historical live broadcast behavior features of the target object into discrete features" may mainly include: sorting each continuous feature according to the value of each continuous feature; bucketing the sorting order according to preset boundaries; and converting each continuous feature into a discrete feature according to the bucket in which each continuous feature is located.

[0101] For example, the number of gifts given is a continuous feature. First, sort the number of gifts given from small to large. Then, take 10, 20, ..., 90 as 9 boundaries and divide the sorting order into 10 ranges (i.e., buckets). Then, according to the range of each feature, the continuous feature can be converted into a discrete feature.

[0102] Specifically, the step of "encoding the discrete features in the multiple historical live broadcast behavior features of the target object, as well as the converted discrete features, to obtain the historical live broadcast behavior data feature encoding of the target object" can specifically include: One-Hot encoding the discrete features in the multiple historical live broadcast behavior features of the target object, as well as the converted discrete features, to obtain the historical live broadcast behavior data feature encoding of the target object.

[0103] Specifically, multiple historical live broadcast behavior features may also include ID sequence features. For the ID sequence features, the ID sequence features may be arranged in reverse chronological order. At the same time, the ID sequence features may be truncated to a length of 10 digits, or the ID sequence features that are less than 10 digits may be padded with the number 0.

[0104] In this embodiment, the step of "determining a piece of training sample data in the training sample set based on the target object, the target object's historical live broadcast behavior data feature code, the target object's first live broadcast preference true value label, and the target object's second live broadcast preference true value label" includes: determining the weight of the training sample data corresponding to the target object according to the preset business goal; determining a piece of training sample data in the training sample set based on the target object, the target object's historical live broadcast behavior data feature code, the target object's first live broadcast preference true value label, the target object's second live broadcast preference true value label, and the weight of the training sample data corresponding to the target object.

[0105] Specifically, if the second live broadcast preference truth value label of the target sample is a preset value, the weight of the training sample data corresponding to the target object can be determined based on the preset business goal and the training sample features corresponding to the target object. For example, if the second live broadcast preference truth value label of the target sample is 1, the weight of the sample can be obtained by weighted summing up the preset business goal and the corresponding viewing time, attention UV, fan group PV and gift amount of the target object. For example, if the preset business goal is to increase the live broadcast viewing time, the weight of the viewing time corresponding to the target object can be appropriately increased, and then the viewing time, attention UV, fan group PV and gift amount corresponding to the target object can be weighted summed to obtain the weight of the sample. For another example, if the preset business goal is to increase the gift amount, the gift amount corresponding to the target object can be appropriately increased, and the viewing time, attention UV, fan group PV and gift amount corresponding to the target object can be weighted summed to obtain the weight of the sample.

[0106] Specifically, if the second live broadcast preference true value label of the target sample is not a preset value, its weight is determined to be 0. For example, if the second live broadcast preference true value label of the target sample is not 1, its weight is determined to be 0.

[0107] Step 102: input target training sample data into the preference determination model, where the target training sample data is any piece of training sample data in the training sample set.

[0108] See also Figure 3 The preference determination model may include a user feature processing module, which may be used to obtain the user's historical live broadcast behavior features and the user's historical live broadcast behavior feature encoding.

[0109] Step 103 : Obtain a first predicted live broadcast preference degree corresponding to the target training sample object and a second predicted live broadcast preference degree of the target training sample object determined by the preference determination model according to the target training sample data.

[0110] Specifically, if Figure 3As shown, the preference determination model includes an embedding module (Embedding Layer), a first preference determination module and a second preference determination module. Among them, the embedding module (Embedding Layer) can convert the One-Hot encoding feature into a 32-dimensional continuous feature vector. Among them, the embedding module also includes a sub-module Avg Pooling, and Avg Pooling is used to convert the ID sequence feature into a 32-dimensional continuous feature vector. The first preference determination module includes LR, FM and DNN. The embedding module inputs each continuous feature vector into LR, FM and DNN respectively. LR, FM and DNN perform feature crossover on the continuous feature vector. Afterwards, the outputs of LR, FM and DNN are weighted and summed, and the first live broadcast preference is obtained through the Sigmoid function. The second preference determination module includes LR, FM and DNN. The embedding module inputs each continuous feature vector into LR, FM and DNN respectively. LR, FM and DNN perform feature crossover on the continuous feature vector. Afterwards, the outputs of LR, FM and DNN are weighted and summed, and the second live broadcast preference is obtained through the Sigmoid function.

[0111] Step 104: Based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object, the model parameters of the preference determination model are trained to obtain a trained preference determination model.

[0112] In some embodiments, step 104 may mainly include: determining a target loss function based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, and the first live broadcast preference true value label of the target training sample object and the second live broadcast preference true value label of the target training sample object; training the preference determination model according to the target loss function to obtain a trained preference determination model.

[0113] The target loss function is used to measure the deviation between the first predicted live broadcast preference and the first live broadcast preference true value label, as well as the deviation between the second predicted live broadcast preference and the second live broadcast preference true value label. Specifically, the preference determination model is trained with the goal of minimizing the target loss function.

[0114] In this embodiment, the step of "determining the target loss function based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, and the first live broadcast preference true value label of the target training sample object and the second live broadcast preference true value label of the target training sample object" includes: determining the target loss function based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, and the first live broadcast preference true value label of the target training sample object, the second live broadcast preference true value label of the target training sample object and the weight of the training sample data corresponding to the target training sample.

[0115] Among them, the following formula can be determined as the target loss function Loss:

[0116]

[0117]

[0118] in, Represents the first live broadcast preference truth label, y label2 Represents the second live broadcast preference truth label, x i represents the input of the preference determination model, represents the parameters of the first preference determination module, represents the parameters of the second preference determination module, Represents the first predicted live broadcast preference, Represents the second predicted live broadcast preference.

[0119] All of the above technical solutions can be combined in any way to form optional embodiments of the present application, and will not be described in detail here.

[0120] The embodiment of the present application obtains a training sample set, wherein each training sample data in the training sample set includes at least one training sample object, and multiple historical live broadcast behavior features, a first live broadcast preference true value label, and a second live broadcast preference true value label for each training sample object in at least one training sample object. Then, the target training sample data is input into the preference determination model, where the target training sample data is any training sample data in the training sample set. Then, the first predicted live broadcast preference of the corresponding target training sample object and the second predicted live broadcast preference of the target training sample object determined by the preference determination model according to the target training sample data are obtained, and the model parameters of the preference determination model are trained according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, and the first live broadcast preference true value label of the target training sample object and the second live broadcast preference true value label of the target training sample object to obtain a trained preference determination model. The embodiment of the present application overcomes the problem of being unable to effectively train the model due to the sparse user live broadcast consumption data by jointly training the first live broadcast preference and the second live broadcast preference, and avoids the problem of the trained model having a large deviation due to the different data distribution of the predicted user and the training user due to the lack of live broadcast data of the user.

[0121] See also Figure 2 、 Figure 3 and Figure 4 The embodiment of the present application further provides a live content recommendation method, including steps 201 to 203, as described below:

[0122] Step 201 : Acquire multiple live broadcast behavior feature codes of a target object, where the multiple live broadcast behavior feature codes are used to represent the behavior features of the target object with respect to the content exposed before the live broadcast and the behavior features of the target object with respect to the actual content of the live broadcast.

[0123] The multiple live streaming behavior feature codes may be one-hot codes. The target user is a user who uses the target software and authorizes the target software to obtain their historical usage data. Pre-live streaming content may refer to promotional text, images, and other content displayed within the target software, such as posters and links, about the live stream or the host. The actual live stream content may refer to the content on the host's live streaming page.

[0124] Specifically, the multiple live broadcast behavior feature codes may also include a target object portrait data code, a target object live broadcast-independent behavior feature code using the target software, and a user-independent live broadcast general feature code. The target object portrait data code, the target object live broadcast-independent behavior feature code using the target software, and the user-independent live broadcast general feature code are obtained by encoding training sample features constructed based on the target user's user portrait data, the target user live broadcast-independent behavior data using the target software, and the user-independent live broadcast general data.

[0125] For example, the target audience's behavioral characteristics regarding pre-live content exposure and the actual live content of the live broadcast may include the target audience's exposure PV, exposure UV, click PV, click UV, click rate PV, click rate UV, total viewing time, follow UV, fan group PV, fan group UV, number of giftings, gifting UV, total gift amount, etc. For example, the target audience's user profile data may include the target user's age, gender, city, city level, VIP type, VIP level, device type, operating system, etc. For example, the target user's non-live broadcast behavior characteristics using the target software may mainly include the target user's recent listening time, the list of song IDs listened to recently, the list of artist IDs listened to recently, etc. For example, user-unrelated live broadcast general data may mainly include the number of live streamers on the platform, the average live broadcast duration of the live streamer, the user's live broadcast consumption time, the total user consumption time, the average user consumption time, etc.

[0126] It's easy to understand that the target subject's multiple live streaming behavior feature codes not only include behavioral features characterizing the target subject's response to pre-live streaming exposure and the target subject's response to the actual live streaming content, but also include user profile features characterizing the target subject and non-live streaming behavior features of the target subject using the target software. This allows prediction of the target subject's preference for live streaming even if the target subject has no experience with live streaming services. For example, non-live streaming behavior features of the target subject using the target software include a list of song IDs the target subject has recently listened to and a list of artist IDs the target subject has recently listened to. Assume there are target subjects A and B. Target subject A has no experience with live streaming services and lacks historical live streaming behavior data. However, if the song ID lists and artist ID lists that target subject B has recently listened to overlap significantly with those of target subject A, it can be assumed that target subject B and target subject A share similar preferences. Target subject A's preference for live streaming can then be inferred based on target subject B's preference for live streamers and live streaming.

[0127] Step 202: Based on a preset preference determination model, a plurality of live broadcast behavior feature codes are used as inputs of the preset preference determination model to determine a first live broadcast preference and a second live broadcast preference of the target object, and a comprehensive live broadcast preference of the target object is determined based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to characterize the target object's preference for the content exposed before the live broadcast, and the second live broadcast preference is used to characterize the target object's preference for the actual content of the live broadcast.

[0128] In related technologies, users' preferences for live broadcasting are usually predicted based on user portrait data and user live broadcasting behavior data, using a deep learning model. This results in the inability to predict the preferences for live broadcasting and anchors for users without live broadcasting behavior data. This embodiment is based on a deep learning model, and based on the behavioral characteristics of the target object for the content exposed before the live broadcast and the behavioral characteristics of the target object for the actual content of the live broadcast, predicts the target object's preference for the content exposed before the live broadcast and the target object's preference for the actual content of the live broadcast, and calculates the target object's preference for the live broadcast by combining the target object's preference for the content exposed before the live broadcast and the target object's preference for the actual content of the live broadcast. This can also predict the preference for live broadcasting for users who have not perceived or experienced live broadcasting, overcoming the problem of being unable to calculate their preference for live broadcasting due to the sparseness of their live broadcasting behavior data. This embodiment can effectively predict the preference for live broadcasting of all users, and then make accurate live broadcast recommendations to users based on the preference of all users for live broadcasting.

[0129] In some embodiments, the preset preference determination model includes an embedding module, a first preference determination module and a second preference determination module, and the step of "based on the preset preference determination model, taking multiple live broadcast behavior feature codes as input of the preset preference determination model, determining the first live broadcast preference and the second live broadcast preference of the target object" includes: controlling the embedding module to convert the multiple live broadcast behavior feature codes into multiple continuous feature vectors of preset dimensions respectively; controlling the first preference determination module to determine the first live broadcast preference of the target object based on the multiple continuous feature vectors; controlling the second preference determination module to determine the second live broadcast preference of the target object based on the multiple continuous feature vectors.

[0130] The preset dimension may be 32. Specifically, the embedding module may convert the One-Hot encoding or ID sequence features into a 32-dimensional continuous feature vector.

[0131] For example, see Figure 3The preset preference determination module includes a user feature processing module, which can be used to obtain multiple live broadcast behavior feature codes of the target object. Alternatively, the user feature processing module can also be used to obtain multiple live broadcast behavior features of the target object and encode the multiple live broadcast behavior features of the target object. The multiple live broadcast behavior features of the target object may include discrete features, continuous features, and ID sequence features. The user feature processing module can perform One-Hot encoding on discrete features, and after bucketing the continuous features, perform One-Hot encoding on the bucketed continuous features.

[0132] Specifically, if Figure 3 As shown in Figure 1, the embedding module (Embedding Layer) can convert the One-Hot encoded features into a 32-dimensional continuous feature vector. The embedding module also includes a submodule Avg Pooling, which is used to convert the ID sequence features into a 32-dimensional continuous feature vector.

[0133] In some embodiments, the step of "controlling the first preference determination module to determine the first live broadcast preference of the target object based on multiple continuous feature vectors" includes: controlling the first preference determination module to calculate the first live broadcast preference of the target object based on multiple continuous feature vectors, and the weight number corresponding to each continuous feature vector in the multiple continuous feature vectors in the first preference determination module determined in a preset preference determination model.

[0134] Among them, the weight number corresponding to each of the multiple continuous feature vectors in the first preference determination module is obtained by training the preset preference determination model according to the business objectives during the training phase. Specifically, the preset preference determination model can be optimized and trained according to different business objectives, such as DAU, attention rate, payment rate, etc., and then the target object's preference for live broadcast is predicted based on the optimized model.

[0135] Please continue reading Figure 3 and Figure 4 The first preference determination module includes LR, FM and DNN. The embedding module inputs each continuous feature vector into LR, FM and DNN respectively. LR, FM and DNN perform feature cross-pollination on the continuous feature vectors. After that, the outputs of LR, FM and DNN are weighted and summed, and the first live broadcast preference is obtained through the Sigmoid function.

[0136] In some embodiments, the step of "controlling the second preference determination module to determine the second live broadcast preference of the target object based on multiple continuous feature vectors" includes: controlling the second preference determination module to calculate the second live broadcast preference of the target object based on multiple continuous feature vectors, and the weight number corresponding to each continuous feature vector in the multiple continuous feature vectors in the second preference determination module determined in a preset preference determination model.

[0137] Similarly, the weight number corresponding to each of the multiple continuous feature vectors in the second preference determination module is obtained by training the preset preference determination model according to the business objectives during the training phase. Specifically, the preset preference determination model can be optimized and trained according to different business objectives, such as DAU, attention rate, payment rate, etc., and then the target object's preference for live broadcast can be predicted based on the optimized model.

[0138] Please continue reading Figure 3 and Figure 4 The second preference determination module includes LR, FM and DNN. The embedding module inputs each continuous feature vector into LR, FM and DNN respectively. LR, FM and DNN perform feature cross-pollination on the continuous feature vectors. After that, the outputs of LR, FM and DNN are weighted and summed, and the second live broadcast preference is obtained through the Sigmoid function.

[0139] In some embodiments, the step of "determining the comprehensive live broadcast preference of the target object based on the first live broadcast preference and the second live broadcast preference" includes: calculating the product between the first live broadcast preference and the second live broadcast preference; and determining the product as the comprehensive live broadcast preference of the target object.

[0140] For example, see Figure 4 The first preference determination module determines that the first live broadcast preference is 0.7, and the second preference determination module determines that the second live broadcast preference is 0.3. Then, the product of the first live broadcast preference and the second live broadcast preference, 0.21, can be determined as the comprehensive live broadcast preference of the target object.

[0141] Step 203: recommend live broadcast content to the target object based on the target object's comprehensive live broadcast preference.

[0142] For example, based on the comprehensive preferences of the target objects, more live broadcast exposure resources can be recommended on the pages of target objects with higher comprehensive preferences, and more exposure resources, such as music, podcasts, etc., can be recommended on the pages of target objects with lower comprehensive preferences.

[0143] In some embodiments, step 203 may mainly include: determining the comprehensive live broadcast preference of all objects including the target object in the target software based on a preset preference determination model; sorting all objects according to the comprehensive live broadcast preference of all objects; and recommending live broadcast content to the target object according to the position of the target object in the sorting.

[0144] For example, we can rank the target audiences based on their overall live streaming preference from high to low, and then make live streaming recommendations based on their position in the ranking. For example, we can recommend more different types of live streamers to the top 10% of the ranked audiences, and more newcomer-friendly live streamers to the bottom 20%.

[0145] The embodiment of the present application obtains multiple live broadcast behavior feature codes of the target object, and the multiple live broadcast behavior feature codes are used to characterize the target object's behavior characteristics with respect to the content exposed before the live broadcast and the target object's behavior characteristics with respect to the actual content of the live broadcast. Then, based on a preset preference determination model, the multiple live broadcast behavior feature codes are used as inputs of the preset preference determination model to determine the first live broadcast preference and the second live broadcast preference of the target object, and the comprehensive live broadcast preference of the target object is determined based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to characterize the target object's preference with respect to the content exposed before the live broadcast, and the second live broadcast preference is used to characterize the target object's preference with respect to the actual content of the live broadcast. Thereafter, live broadcast content is recommended to the target object based on the comprehensive live broadcast preference of the target object. The embodiments of the present application can determine the target object's first live broadcast preference for the pre-live broadcast content and the second live broadcast preference for the actual live broadcast content based on the target object's behavioral characteristics for the pre-live broadcast content and the target object's behavioral characteristics for the actual live broadcast content, and combine the first live broadcast preference and the second live broadcast preference to determine the target object's comprehensive preference for live broadcast, thereby improving the accuracy of predicting the user's preference for live broadcast, and overcoming the problem in related technologies that the user's preference for live broadcast cannot be predicted due to the lack of user behavioral data for the actual live broadcast content, thereby improving the accuracy of live broadcast content recommendation based on the target object's preference for live broadcast.

[0146] In order to better implement the training method of the preference determination model of the embodiment of the present application, the embodiment of the present application also provides a training device for the preference determination model. Figure 5 , Figure 5 A schematic diagram of the structure of a training device for a preference determination model provided in an embodiment of the present application. The training device 10 for a preference determination model may include:

[0147] A first acquisition module 11 is configured to acquire a training sample set, wherein each piece of training sample data in the training sample set includes at least one training sample object, and a plurality of historical live broadcast behavior features, a first live broadcast preference true value label, and a second live broadcast preference true value label for each training sample object in the at least one training sample object;

[0148] An input module 12 is used to input target training sample data into the preference determination model, where the target training sample data is any training sample data in the training sample set;

[0149] The second acquisition module 13 is used to obtain the first predicted live broadcast preference of the target training sample object and the second predicted live broadcast preference of the target training sample object determined by the preference determination model according to the target training sample data;

[0150] The training module 14 is used to train the model parameters of the preference determination model based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object to obtain a trained preference determination model.

[0151] Optionally, the training device 10 of the preference determination model may further include a processing module, which is used to: before obtaining the training sample set, obtain the historical live broadcast behavior data of the target user, the historical live broadcast behavior data is the live broadcast behavior data of the target object within a preset historical time period, and the historical live broadcast behavior data includes data of multiple dimensions; determine the first live broadcast preference true value label of the target object based on the data of at least one first dimension pre-set in the multiple dimensions, and determine the second live broadcast preference true value label of the target object based on the first live broadcast preference true value label and the data of at least one second dimension pre-set in the multiple dimensions, the first live broadcast preference true value label is used to characterize the target object's preference for the content exposed before the live broadcast, and the second live broadcast preference true value label is used to characterize the target object's preference for the actual content of the live broadcast; determine multiple historical live broadcast behavior features of the target object based on the data of each dimension in the multiple dimensions; construct a training sample set based on the target object, the multiple historical live broadcast behavior features of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object.

[0152] Optionally, the processing module can be specifically used to: when the first live broadcast preference truth value label is a preset value, determine the second live broadcast preference truth value label of the target object based on data of at least one second dimension preset in multiple dimensions.

[0153] Optionally, the processing module can be specifically configured to: convert continuous features in the plurality of historical live broadcast behavior features of the target object into discrete features; encode the discrete features in the plurality of historical live broadcast behavior features of the target object and the converted discrete features to obtain a historical live broadcast behavior data feature code of the target object; and determine a piece of training sample data in the training sample set according to the target object, the historical live broadcast behavior data feature code of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object.

[0154] Optionally, the processing module can be specifically configured to: determine a weight of the training sample data corresponding to the target object according to the preset business target; and determine a piece of training sample data in the training sample set according to the target object, the historical live broadcast behavior data feature code of the target object, the first live broadcast preference true value label of the target object, the second live broadcast preference true value label of the target object, and the weight of the training sample data corresponding to the target object.

[0155] Optionally, the training module 14 can be specifically configured to: determine a target loss function according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, and the first live broadcast preference true value label of the target training sample object and the second live broadcast preference true value label of the target training sample object; and train the preference determination model according to the target loss function to obtain the trained preference determination model.

[0156] Optionally, the training module 14 can be specifically configured to: determine a target loss function according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, and the first live broadcast preference true value label of the target training sample object, the second live broadcast preference true value label of the target training sample object, and the weight of the training sample data corresponding to the target training sample.

[0157] The various units in the training device of the above preference determination model can be realized by software, hardware, and combinations thereof, in whole or in part. The various units can be embedded in or independent of the processor in the computer device in the form of hardware, or stored in the memory in the computer device in the form of software, so as to be called and executed by the processor to perform the operations corresponding to the various units.

[0158] The training device 10 of the preference determination model can be integrated in a terminal or a server with a storage and a processor, or the training device 10 of the preference determination model is the terminal or the server.

[0159] The training device 10 of the preference determination model provided in the embodiment of the present application obtains a training sample set through a first acquisition module 11, and each training sample data in the training sample set includes at least one training sample object, and multiple historical live broadcast behavior features, a first live broadcast preference true value label and a second live broadcast preference true value label of each training sample object in at least one training sample object. Afterwards, the input module 12 inputs the target training sample data into the preference determination model, and the target training sample data is any training sample data in the training sample set. Then, the second acquisition module 13 obtains the first predicted live broadcast preference of the corresponding target training sample object and the second predicted live broadcast preference of the target training sample object determined by the preference determination model according to the target training sample data. Then, the training module 14 trains the model parameters of the preference determination model according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, and the first live broadcast preference true value label and the second live broadcast preference true value label of the target training sample object to obtain a trained preference determination model. The embodiment of the present application overcomes the problem of being unable to effectively train the model due to sparse user live broadcast consumption data by jointly training the first live broadcast preference and the second live broadcast preference, and avoids the problem of different data distributions between the predicted user and the training user due to the lack of user live broadcast data, which in turn leads to a large deviation in the trained model.

[0160] The present application also provides a live content recommendation device. Figure 6 , Figure 6 This is a schematic diagram of the structure of the live content recommendation device provided in an embodiment of the present application. The live content recommendation device 20 may include:

[0161] An acquisition module 21 is configured to acquire multiple live broadcast behavior feature codes of a target object, wherein the multiple live broadcast behavior feature codes are used to represent the behavior characteristics of the target object with respect to the content exposed before the live broadcast and the behavior characteristics of the target object with respect to the actual content of the live broadcast;

[0162] Determination module 22 is configured to determine a first live broadcast preference and a second live broadcast preference of a target subject based on a preset preference determination model and using a plurality of live broadcast behavior feature codes as inputs to the preset preference determination model, and to determine a comprehensive live broadcast preference of the target subject based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to represent the target subject's preference for content exposed before the live broadcast, and the second live broadcast preference is used to represent the target subject's preference for the actual content of the live broadcast;

[0163] The recommendation module 23 is used to recommend live broadcast content to the target object based on the target object's comprehensive live broadcast preference.

[0164] Optionally, the preset preference determination model includes an embedding module, a first preference determination module and a second preference determination module; the determination module 22 can be specifically used to: control the embedding module to convert multiple live broadcast behavior feature codes into multiple continuous feature vectors of preset dimensions; control the first preference determination module to determine the first live broadcast preference of the target object based on the multiple continuous feature vectors; control the second preference determination module to determine the second live broadcast preference of the target object based on the multiple continuous feature vectors.

[0165] Optionally, the determination module 22 can be specifically used to: control the first preference determination module to calculate the first live broadcast preference of the target object based on multiple continuous feature vectors and the weight number corresponding to each continuous feature vector in the multiple continuous feature vectors in the first preference determination module determined in the preset preference determination model.

[0166] Optionally, the determination module 22 can be specifically used to: control the second preference determination module to calculate the second live broadcast preference of the target object based on multiple continuous feature vectors and the weight number corresponding to each continuous feature vector in the multiple continuous feature vectors in the second preference determination module determined in the preset preference determination model.

[0167] Optionally, the determination module 22 may also be configured to: calculate the product of the first live broadcast preference and the second live broadcast preference; and determine the product as the comprehensive live broadcast preference of the target object.

[0168] Optionally, the recommendation module 23 can be specifically used to: determine the comprehensive live broadcast preference of all objects including the target object in the target software based on a preset preference determination model; sort all objects according to the comprehensive live broadcast preference of all objects; and recommend live broadcast content to the target object according to the position of the target object in the sorting.

[0169] The live content recommendation device 20 provided in the embodiment of the present application obtains multiple live behavior feature codes of the target object through the acquisition module 21, and the multiple live behavior feature codes are used to characterize the behavioral characteristics of the target object with respect to the content exposed before the live broadcast and the behavioral characteristics of the target object with respect to the actual content of the live broadcast. Then, the determination module 22 determines the first live broadcast preference and the second live broadcast preference of the target object based on the preset preference determination model, using the multiple live broadcast behavior feature codes as inputs of the preset preference determination model, and determines the comprehensive live broadcast preference of the target object based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to characterize the target object's preference for the content exposed before the live broadcast, and the second live broadcast preference is used to characterize the target object's preference for the actual content of the live broadcast. Then, the recommendation module 13 recommends live content to the target object based on the comprehensive live broadcast preference of the target object. The embodiment of the present application determines the target object's first live broadcast preference for the pre-live broadcast content and the second live broadcast preference for the actual live broadcast content based on the target object's behavioral characteristics for the pre-live broadcast content and the target object's behavioral characteristics for the actual live broadcast content, and combines the first live broadcast preference and the second live broadcast preference to determine the target object's comprehensive preference for live broadcast, thereby improving the accuracy of predicting the user's preference for live broadcast, and overcoming the problem in related technologies that is unable to predict the user's preference for live broadcast due to a lack of user behavioral data for the actual live broadcast content, thereby improving the accuracy of live broadcast content recommendation based on the target object's preference for live broadcast.

[0170] Optionally, the present application also provides a computer device, including a memory and a processor, wherein a computer program is stored in the memory, and the processor implements the steps in the above-mentioned method embodiments when executing the computer program.

[0171] Figure 7 A schematic diagram of the structure of a computer device provided in an embodiment of the present application is shown in FIG. Figure 7 As shown, the computer device 700 may include: a communication interface 701, a memory 702, a processor 703, and a communication bus 704. The communication interface 701, the memory 702, and the processor 703 communicate with each other via the communication bus 704. The communication interface 701 is used for data communication between the computer device 700 and external devices. The memory 702 can be used to store software programs and modules. The processor 703 executes the software programs and modules stored in the memory 702, such as the software programs for the corresponding operations in the aforementioned method embodiments.

[0172] Optionally, the processor 703 can call the software program and module stored in the memory 702 to perform the following operations: obtain a training sample set, each training sample data in the training sample set includes at least one training sample object, and multiple historical live broadcast behavior characteristics, a first live broadcast preference true value label and a second live broadcast preference true value label of each training sample object in at least one training sample object; input the target training sample data into the preference determination model, the target training sample data is any training sample data in the training sample set; obtain the first predicted live broadcast preference of the corresponding target training sample object and the second predicted live broadcast preference of the target training sample object determined by the preference determination model according to the target training sample data; train the model parameters of the preference determination model according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, and the first live broadcast preference true value label and the second live broadcast preference true value label of the target training sample object to obtain a trained preference determination model.

[0173] Optionally, the processor 703 can call the software program and module stored in the memory 702 to perform the following operations: obtain multiple live broadcast behavior feature codes of the target object, and the multiple live broadcast behavior feature codes are used to characterize the behavior characteristics of the target object for the content exposed before the live broadcast and the behavior characteristics of the target object for the actual content of the live broadcast; based on a preset preference determination model, use the multiple live broadcast behavior feature codes as inputs of the preset preference determination model to determine the first live broadcast preference and the second live broadcast preference of the target object, and determine the comprehensive live broadcast preference of the target object based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to characterize the target object's preference for the content exposed before the live broadcast, and the second live broadcast preference is used to characterize the target object's preference for the actual content of the live broadcast; recommend live broadcast content to the target object based on the comprehensive live broadcast preference of the target object.

[0174] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding process of the training method for the preference determination model in the embodiment of this application. For the sake of brevity, it is not further described here.

[0175] This application also provides a computer-readable storage medium for storing a computer program. This computer-readable storage medium can be applied to a computer device, and the computer program causes the computer device to execute the corresponding process of the live content recommendation method in the embodiment of this application. For the sake of brevity, it is not further described here.

[0176] This application also provides a computer program product, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding processes of the method for training a preference determination model in the embodiments of this application. For the sake of brevity, these processes are not further described here.

[0177] This application also provides a computer program product, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding process of the live content recommendation method in the embodiment of this application. For the sake of brevity, the details are not repeated here.

[0178] This application also provides a computer program comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding processes of the method for training a preference determination model in the embodiments of this application. For the sake of brevity, these processes are not further described here.

[0179] This application also provides a computer program, which includes computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions, causing the computer device to perform the corresponding process of the live content recommendation method in the embodiment of this application. For the sake of brevity, the details are not repeated here.

[0180] It should be understood that the processor of the embodiments of the present application may be an integrated circuit chip with signal processing capabilities. During implementation, each step of the above method embodiment can be completed by hardware integrated logic circuits in the processor or software instructions. The above processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. The various methods, steps, and logic block diagrams disclosed in the embodiments of the present application can be implemented or executed. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in the embodiments of the present application can be directly implemented as a hardware decoding processor, or can be implemented by a combination of hardware and software modules in the decoding processor. The software module can be located in a storage medium mature in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, registers, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware.

[0181] It is understood that the memory in the embodiments of the present application may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memories. Among them, the non-volatile memory may be a read-only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable programmable read-only memory (EEPROM), or a flash memory. The volatile memory may be a random access memory (RAM), which is used as an external cache. By way of example and not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDR SDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous link dynamic random access memory (SLDRAM), and direct RAM bus random access memory (DR RAM). It should be noted that the memory of the systems and methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0182] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0183] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the systems, devices and units described above can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0184] In the several embodiments provided in this application, it should be understood that the disclosed systems, devices and methods can be implemented in other ways. For example, the device embodiments described above are merely schematic. For example, the division of the units is merely a logical function division. In actual implementation, there may be other division methods, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0185] The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of these units may be selected to achieve the purpose of this embodiment according to actual needs.

[0186] In addition, each functional unit in the embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit.

[0187] If the functions are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling a computer device (which can be a personal computer, a server) to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes various media that can store program codes, such as a USB flash drive, a mobile hard disk, a ROM, a RAM, a magnetic disk, or an optical disk.

[0188] The above description is merely a specific embodiment of the present application, but the scope of protection of the present application is not limited thereto. Any changes or substitutions that can be easily conceived by a person skilled in the art within the technical scope disclosed in the present application should be included in the scope of protection of the present application. Therefore, the scope of protection of the present application should be based on the scope of protection of the claims.

Claims

1. A method for training a preference determination model, characterized in that: The method further comprises: Obtaining a training sample set, where each piece of training sample data in the training sample set includes at least one training sample object, and a plurality of historical live broadcast behavior features, a first live broadcast preference true value label, and a second live broadcast preference true value label for each training sample object in the at least one training sample object; Inputting target training sample data into the preference determination model, wherein the target training sample data is any training sample data in the training sample set; Obtaining a first predicted live broadcast preference of a corresponding target training sample object and a second predicted live broadcast preference of the target training sample object determined by the preference determination model according to the target training sample data; Training the model parameters of the preference determination model according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object to obtain a trained preference determination model; Based on the preference determination model, the comprehensive live broadcast preference of all objects including the target object in the target software is determined; according to the comprehensive live broadcast preference of all objects, the objects are sorted; and according to the position of the target object in the sorting, live broadcast content is recommended for the target object.

2. The method for training a preference determination model according to claim 1, wherein: Before obtaining the training sample set, the method further includes: Acquire historical live broadcast behavior data of a target user, wherein the historical live broadcast behavior data is live broadcast behavior data of the target object within a preset historical time period, and the historical live broadcast behavior data includes data of multiple dimensions; Determine a first live broadcast preference truth value label of the target object based on data of at least one first dimension preset among the multiple dimensions, and determine a second live broadcast preference truth value label of the target object based on the first live broadcast preference truth value label and data of at least one second dimension preset among the multiple dimensions, wherein the first live broadcast preference truth value label is used to characterize the target object's preference for content exposed before the live broadcast, and the second live broadcast preference truth value label is used to characterize the target object's preference for actual content of the live broadcast; Determining multiple historical live broadcast behavior features of the target object based on data of each dimension of the multiple dimensions; A training sample set is constructed according to the target object, multiple historical live broadcast behavior features of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object.

3. The method for training a preference determination model according to claim 2, wherein: The determining the second live broadcast preference true value label of the target object according to the first live broadcast preference true value label and data of at least one second dimension preset in the multiple dimensions includes: When the first live broadcast preference truth value label is a preset value, the second live broadcast preference truth value label of the target object is determined based on data of at least one second dimension preset in the multiple dimensions.

4. The method for training a preference determination model according to claim 2, wherein: The constructing of a training sample set according to the target object, the multiple historical live broadcast behavior features of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object includes: Converting continuous features in the plurality of historical live broadcast behavior features of the target object into discrete features; Encoding the discrete features in the multiple historical live broadcast behavior features of the target object and the converted discrete features to obtain the historical live broadcast behavior data feature code of the target object; A piece of training sample data in the training sample set is determined based on the target object, the historical live broadcast behavior data feature code of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object.

5. The method for training a preference determination model according to claim 4, wherein: The determining of a piece of training sample data in the training sample set based on the target object, the feature code of the historical live broadcast behavior data of the target object, the first live broadcast preference true value label of the target object, and the second live broadcast preference true value label of the target object includes: Determine the weight of the training sample data corresponding to the target object according to the preset business goal; A piece of training sample data in the training sample set is determined based on the target object, the historical live broadcast behavior data feature code of the target object, the first live broadcast preference true value label of the target object, the second live broadcast preference true value label of the target object, and the weight of the training sample data corresponding to the target object.

6. The method for training a preference determination model according to claim 5, wherein: Training model parameters of the preference determination model based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object to obtain a trained preference determination model, including: Determining a target loss function according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object; The preference determination model is trained according to the target loss function to obtain the trained preference determination model.

7. The method for training a preference determination model according to claim 6, wherein: The determining of the target loss function according to the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object includes: The target loss function is determined based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, the second live broadcast preference true value label of the target training sample object and the weight of the training sample data corresponding to the target training sample.

8. A live content recommendation method, characterized in that: The method comprises: Acquire multiple live broadcast behavior feature codes of a target object, where the multiple live broadcast behavior feature codes are used to characterize the behavior characteristics of the target object with respect to the content exposed before the live broadcast and the behavior characteristics of the target object with respect to the actual content of the live broadcast; Based on a preset preference determination model, the plurality of live broadcast behavior feature codes are used as inputs to the preset preference determination model to determine a first live broadcast preference and a second live broadcast preference of the target object, and a comprehensive live broadcast preference of the target object is determined based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to represent the target object's preference for content exposed before the live broadcast, and the second live broadcast preference is used to represent the target object's preference for the actual content of the live broadcast; Recommending live broadcast content to the target object based on the target object's comprehensive live broadcast preference; The recommending live broadcast content to the target object according to the comprehensive live broadcast preference includes: Determining the comprehensive live broadcast preference of all objects including the target object in the target software based on the preset preference determination model; Sorting all the objects according to their comprehensive live broadcast preferences; According to the position of the target object in the ranking, live content is recommended to the target object.

9. The live broadcast content recommendation method according to claim 8, wherein: The preset preference determination model includes an embedding module, a first preference determination module, and a second preference determination module; The method of determining the first live broadcast preference and the second live broadcast preference of the target object based on a preset preference determination model and taking the plurality of live broadcast behavior feature codes as inputs of the preset preference determination model includes: Controlling the embedding module to convert the plurality of live broadcast behavior feature codes into a plurality of continuous feature vectors of preset dimensions respectively; Controlling the first preference determination module to determine a first live broadcast preference of the target object according to the multiple continuous feature vectors; The second preference determination module is controlled to determine a second live broadcast preference of the target object according to the multiple continuous feature vectors.

10. The live broadcast content recommendation method according to claim 9, wherein: The controlling the first preference determination module to determine the first live broadcast preference of the target object according to the multiple continuous feature vectors includes: Control the first preference determination module to calculate the first live broadcast preference of the target object based on the multiple continuous feature vectors and the weight number corresponding to each continuous feature vector in the multiple continuous feature vectors in the first preference determination module determined in the preset preference determination model.

11. The live broadcast content recommendation method according to claim 9, wherein: The controlling the second preference determination module to determine the second live broadcast preference of the target object according to the multiple continuous feature vectors includes: Control the second preference determination module to calculate the second live broadcast preference of the target object based on the multiple continuous feature vectors and the weight number corresponding to each continuous feature vector in the multiple continuous feature vectors in the second preference determination module determined in the preset preference determination model.

12. The live broadcast content recommendation method according to claim 8, wherein: The determining the comprehensive live broadcast preference of the target object according to the first live broadcast preference and the second live broadcast preference includes: Calculating the product of the first live broadcast preference and the second live broadcast preference; The product is determined as the comprehensive live broadcast preference of the target object.

13. A training device for a preference determination model, characterized in that: A method for training a preference determination model according to any one of claims 1 to 7, wherein the apparatus comprises: A first acquisition module is configured to acquire a training sample set, wherein each piece of training sample data in the training sample set includes at least one training sample object, and a plurality of historical live broadcast behavior features, a first live broadcast preference true value label, and a second live broadcast preference true value label for each training sample object in the at least one training sample object; An input module, configured to input target training sample data into a preference determination model, wherein the target training sample data is any training sample data in the training sample set; A second acquisition module is configured to acquire a first predicted live broadcast preference of a corresponding target training sample object and a second predicted live broadcast preference of the target training sample object determined by the preference determination model according to the target training sample data; a training module, configured to train model parameters of the preference determination model based on the first predicted live broadcast preference of the target training sample object, the second predicted live broadcast preference of the target training sample object, the first live broadcast preference true value label of the target training sample object, and the second live broadcast preference true value label of the target training sample object, so as to obtain a trained preference determination model; The device is also used to determine the comprehensive live broadcast preference of all objects including the target object in the target software based on the preference determination model; sort all the objects according to the comprehensive live broadcast preference of all the objects; and recommend live broadcast content to the target object according to the position of the target object in the sorting.

14. A live content recommendation device, characterized in that: A method for training a preference determination model according to any one of claims 8 to 12; the apparatus comprising: An acquisition module is configured to acquire a plurality of live broadcast behavior feature codes of a target object, wherein the plurality of live broadcast behavior feature codes are used to characterize the behavior features of the target object with respect to the content exposed before the live broadcast and the behavior features of the target object with respect to the actual content of the live broadcast; a determination module for determining, based on a preset preference determination model and using the plurality of live broadcast behavior feature codes as inputs to the preset preference determination model, a first live broadcast preference and a second live broadcast preference of the target subject, and determining a comprehensive live broadcast preference of the target subject based on the first live broadcast preference and the second live broadcast preference, wherein the first live broadcast preference is used to represent the target subject's preference for content exposed before the live broadcast, and the second live broadcast preference is used to represent the target subject's preference for actual live broadcast content; A recommendation module, configured to recommend live broadcast content to the target object based on the target object's comprehensive live broadcast preference; The recommendation module is also used to determine the comprehensive live broadcast preference of all objects containing the target object in the target software based on the preset preference determination model; sort all the objects according to the comprehensive live broadcast preference of all the objects; and recommend live broadcast content to the target object according to the position of the target object in the sorting.

15. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which is suitable for being loaded by a processor to execute the training method of the preference determination model as described in any one of claims 1-7 or the live content recommendation method as described in any one of claims 8-12.

16. A computer device, characterized in that: The computer device includes a processor and a memory, wherein a computer program is stored in the memory. The processor calls the computer program stored in the memory to execute the training method of the preference determination model as described in any one of claims 1 to 7 or the live content recommendation method as described in any one of claims 8 to 12.

17. A computer program product comprising computer instructions, characterized in that When the computer instructions are executed by the processor, the training method of the preference determination model as described in any one of claims 1 to 7 or the live content recommendation method as described in any one of claims 8 to 12 is implemented.

Images