Method for constructing multi-scene recommendation model, scene recommendation method and related device

Abstract

The application discloses a kind of multi-scene recommendation model construction method, scene recommendation method and related equipment, obtain basic feature vector and scene sensitive feature vector, basic feature vector is processed by multilayer perceptron to obtain depth feature vector, basic feature vector is cross-processed to obtain target feature cross representation vector, scene sensitive feature vector is constructed dynamic weight network based on, to scene sensitive feature vector and target feature cross representation vector are cross-processed to obtain high-order cross feature vector, high-order cross feature vector and depth feature vector are spliced, and are constructed by fully connected layer neural network neural network to obtain multi-scene recommendation model.The application is cross-processed to scene sensitive feature vector and target feature cross representation vector by dynamic weight network, to adjust scene recommendation result and scene ordering weight, to have good recommendation effect to each scene.

Description

Technical Field

[0001] This invention relates to the field of computer technology, and more specifically, to a method for constructing a multi-scenario recommendation model, a scenario recommendation method, and related equipment. Background Technology

[0002] Personalized recommendations utilize user-specific attributes (historical behavior, related preferences, etc.) to mine information that users may be interested in and push it in real time. In personalized recommendation scenarios, different channel pages (such as TV channels and movie channels) belong to different recommendation scenarios, but at the same time, there are similar content recommendation formats. Based on this, a common approach for multi-scenario recommendations is to train a single-scenario recommendation model for scenario-based recommendations. Specifically, for each recommendation scenario, a recommendation model for that scenario is maintained and iterated separately. However, this approach leads to high resource consumption, and subsequent upgrades and maintenance involve a lot of repetitive work, thus significantly reducing the efficiency of model iteration. To address the above problems of single-scenario recommendation models, existing technologies provide another solution: training a multi-scenario recommendation model by combining data from different scenarios, and then using the multi-scenario recommendation model to recommend to multiple scenarios. However, this approach is prone to imbalances in the amount of data across different scenarios, with the entire multi-scenario recommendation model being dominated by scenarios with abundant data, making it difficult to recommend scenarios with less data.

[0003] Therefore, how to provide a method for constructing a multi-scenario recommendation model that can achieve good recommendation results in various scenarios has become a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0004] In view of this, the present invention discloses a method for constructing a multi-scenario recommendation model, a scenario recommendation method, and related equipment, so as to achieve good recommendation results for various scenarios and effectively avoid the problem that scenarios with less data are difficult to recommend due to the multi-scenario recommendation model being dominated by scenarios with abundant data.

[0005] A method for constructing a multi-scenario recommendation model, comprising:

[0006] Obtain basic feature vectors and scene-sensitive feature vectors. The basic feature vectors contain the basic features required to build a multi-scene recommendation model, and the scene-sensitive feature vectors contain the distinguishing features of different scenes.

[0007] The basic feature vectors are processed by a multilayer perceptron to obtain deep feature vectors;

[0008] The basic feature vectors are subjected to feature cross processing to obtain the target feature cross representation vector;

[0009] A dynamic weighted network is constructed based on the scene-sensitive feature vectors.

[0010] Based on the dynamic weight network, feature cross processing is performed on the scene-sensitive feature vector and the target feature cross representation vector to obtain a higher-order cross feature vector;

[0011] The higher-order cross feature vector and the deep feature vector are concatenated together, and a multi-scenario recommendation model is obtained by constructing a neural network.

[0012] Optionally, obtaining the basic feature vector and the scene-sensitive feature vector includes:

[0013] Obtain the basic features of the recommendation model, which include the basic features required to build a multi-scenario model;

[0014] The basic features of the recommendation model are transformed into vector form to obtain the basic feature vector;

[0015] Obtain scene-sensitive features, which include distinguishing features for different scenes;

[0016] The scene-sensitive features are converted into vector form to obtain the scene-sensitive feature vector.

[0017] Optionally, the step of performing feature cross processing on the basic feature vector to obtain the target feature cross representation vector includes:

[0018] The current layer of the cross-network performs a vector transpose operation on the feature cross representation vector of the previous layer to obtain the feature cross transpose vector of the previous layer;

[0019] The target feature cross representation vector is obtained by performing feature cross between the transpose vector of the previous layer and the basic feature vector;

[0020] Wherein, when the current layer is the first layer of the cross network, the feature cross representation vector of the previous layer is the basic feature vector.

[0021] Optionally, constructing a dynamic weight network based on the scene-sensitive feature vector includes:

[0022] The scene-sensitive feature vector is passed through a fully connected neural network to obtain a fixed-dimensional scene-sensitive feature vector;

[0023] The fixed-dimensional scene-sensitive feature vector is reshaped to obtain the dynamic weight network;

[0024] The dynamic weight network includes a multilayer perceptron, the number of layers of which is the same as the number of layers of the cross network, and each layer of the multilayer perceptron has a corresponding scene-sensitive feature weight vector and bias vector.

[0025] Optionally, the step of performing feature cross processing on the scene-sensitive feature vector and the target feature cross representation vector based on the dynamic weight network to obtain a higher-order cross feature vector includes:

[0026] The vectors of each layer of the multilayer perceptron are cross-processed with the feature cross-representation vectors of the corresponding layers of the cross network to extract high-level feature information, and the feature vector obtained by crossing the target feature cross-representation vector of the cross network with the dynamic weight network is determined as the high-order cross feature vector.

[0027] The vectors of each layer of the multilayer perceptron include: the scene-sensitive feature weight vector and the bias vector.

[0028] Optionally, it also includes: mapping the constructed multi-scenario recommendation model from multi-dimensional to one-dimensional to obtain a one-dimensional output vector;

[0029] The one-dimensional output vector is processed using the Sigmoid function to obtain the recommendation model score.

[0030] A scenario recommendation method, using the multi-scenario recommendation model obtained by the construction method of the multi-scenario recommendation model described above, the method comprising:

[0031] Obtain user data;

[0032] Extract user features from the user data;

[0033] The user features are input into the multi-scenario recommendation model to obtain scenario recommendation results.

[0034] An apparatus for constructing a multi-scenario recommendation model, comprising:

[0035] The vector acquisition unit is used to acquire basic feature vectors and scene-sensitive feature vectors. The basic feature vectors contain the basic features required to build a multi-scene recommendation model, and the scene-sensitive feature vectors contain the distinguishing features of different scenes.

[0036] The first processing unit is used to process the basic feature vector through a multilayer perceptron to obtain a deep feature vector;

[0037] The second processing unit is used to perform feature cross processing on the basic feature vector to obtain the target feature cross representation vector.

[0038] Network construction unit, used to construct a dynamic weight network based on the scene-sensitive feature vector;

[0039] The third processing unit is used to perform feature cross processing on the scene-sensitive feature vector and the target feature cross representation vector based on the dynamic weight network to obtain a higher-order cross feature vector.

[0040] The model building unit is used to concatenate the high-order cross feature vector and the deep feature vector together, and construct a multi-scenario recommendation model through a neural network.

[0041] A scene recommendation device, using the multi-scene recommendation model obtained by the construction method of the multi-scene recommendation model described above, the device comprising:

[0042] The data acquisition unit is used to acquire user data;

[0043] A feature extraction unit is used to extract user features from the user data;

[0044] The recommendation unit is used to input the user features into the multi-scenario recommendation model to obtain scenario recommendation results.

[0045] An electronic device, comprising: a memory and a processor;

[0046] The memory is used to store at least one instruction;

[0047] The processor is used to execute the at least one instruction to implement the above-described method for constructing a multi-scenario recommendation model.

[0048] A computer-readable storage medium storing at least one instruction, which, when executed by a processor, implements the method for constructing a multi-scenario recommendation model as described above.

[0049] As can be seen from the above technical solution, this invention discloses a method for constructing a multi-scenario recommendation model, a scenario recommendation method, and related equipment. It obtains basic feature vectors and scenario-sensitive feature vectors. The basic feature vectors are processed using a multilayer perceptron to obtain deep feature vectors. Feature cross processing is then performed on the basic feature vectors to obtain a target feature cross representation vector. After constructing a dynamic weight network based on the scenario-sensitive feature vectors, feature cross processing is performed on the scenario-sensitive feature vectors and the target feature cross representation vectors based on the dynamic weight network to obtain a higher-order cross feature vector. The higher-order cross feature vector and the deep feature vector are concatenated together, and a multi-scenario recommendation model is constructed using a neural network. This invention first processes the basic feature vectors separately to obtain the target feature cross representation vector. The dynamic weight network then performs cross processing on the target feature cross representation vector and the scenario-sensitive feature vector to amplify the importance of the basic features of the recommendation model, allowing for better adjustment of subsequent scenario recommendation results and scenario ranking weights. This effectively avoids the problem of scenarios with limited data being difficult to recommend due to the multi-scenario recommendation model being dominated by data-rich scenarios. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the published drawings without creative effort.

[0051] Figure 1 This is a flowchart of a method for constructing a multi-scenario recommendation model according to an embodiment of the present invention;

[0052] Figure 2 This is a flowchart of a method for obtaining basic feature vectors and scene-sensitive feature vectors disclosed in an embodiment of the present invention;

[0053] Figure 3 This is a system block diagram of a multi-scenario recommendation model disclosed in an embodiment of the present invention;

[0054] Figure 4 This is a schematic diagram of the structure of a multi-scenario recommendation model construction device disclosed in an embodiment of the present invention;

[0055] Figure 5 This is a schematic diagram of the structure of an electronic device disclosed in an embodiment of the present invention. Detailed Implementation

[0056] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] This invention discloses a method for constructing a multi-scenario recommendation model, a scenario recommendation method, and related equipment. It involves obtaining basic feature vectors and scenario-sensitive feature vectors, processing the basic feature vectors to obtain deep feature vectors, and performing feature cross processing on the basic feature vectors to obtain target feature cross representation vectors. After constructing a dynamic weight network based on the scenario-sensitive feature vectors, it performs feature cross processing on the scenario-sensitive feature vectors and target feature cross representation vectors based on the dynamic weight network to obtain higher-order cross feature vectors. By concatenating the higher-order cross feature vectors and deep feature vectors, and constructing a multi-scenario recommendation model through a neural network, this invention first processes the basic feature vectors separately to obtain the target feature cross representation vector. The dynamic weight network then performs cross processing on the target feature cross representation vector and the scenario-sensitive feature vector to amplify the importance of the basic features of the recommendation model, allowing for better adjustment of subsequent scenario recommendation results and scenario ranking weights. This effectively avoids the problem of scenarios with limited data being difficult to recommend due to the multi-scenario recommendation model being dominated by data-rich scenarios.

[0058] See Figure 1 The present invention discloses a flowchart of a method for constructing a multi-scenario recommendation model, the method comprising:

[0059] Step S101: Obtain the basic feature vector and the scene-sensitive feature vector.

[0060] The basic feature vector contains the basic features required to build a multi-scenario recommendation model, such as discrete or continuous features like user age, user gender, geographical location, tags and types of videos watched by the user.

[0061] Scene-sensitive feature vectors contain distinguishing features for different scenes, which are representative features that differentiate different scenes, such as video channels.

[0062] It should be noted that in this embodiment, the basic feature vector is used as input to the deep network and cross network.

[0063] Step S102: Process the basic feature vectors using a multilayer perceptron to obtain deep feature vectors.

[0064] Specifically, deep feature vectors are obtained by processing the basic feature vectors using a multilayer perceptron based on deep networks.

[0065] In this embodiment, the multilayer perceptron is preferably a 3-layer multilayer perceptron (MLP).

[0066] Deep networks typically consist of three layers. They process the basic feature vectors using a multilayer perceptron, transforming the basic feature vectors into a deep representation of the features, resulting in deep feature vectors.

[0067] Deep feature vectors contain non-linear relationships between features and are used for subsequent model fusion with cross-networks.

[0068] Ideally, deep networks can be ESMM (Entire Space Multi-task Model), ESCM2 (ESourcing Capability Model), etc.

[0069] Step S103: Perform feature cross processing on the basic feature vector to obtain the target feature cross representation vector.

[0070] In practical applications, feature crossing processing can be performed on the basic feature vectors based on cross-networks to obtain the target feature cross-representation vector.

[0071] In this embodiment, the cross network can be the cross network in the DCN (Deep & Cross Network) model.

[0072] The target feature cross-representation vectors obtained from the cross-network are mainly used to capture nonlinear relationships between features. By performing feature cross-processing on the basic feature vectors, the cross-network can determine higher-order relationships between features, thereby improving the expressive power of the model.

[0073] Step S104: Construct a dynamic weight network based on scene-sensitive feature vectors.

[0074] Specifically, the scene-sensitive feature vector is passed through a fully connected neural network to obtain a fixed-dimensional scene-sensitive feature vector; the fixed-dimensional scene-sensitive feature vector is then reshaped to construct a dynamic weighted network (DWN).

[0075] For example, the fixed-dimensional scene-sensitive feature vector has a total of 768 dimensions, 768 = (128 + 128) × 3. Therefore, the scene-sensitive feature weights and biases in the dynamic weight network can be set to 128 dimensions, thus reshaping it into a 3-layer multilayer perceptron with 128→128→128.

[0076] In other words, the dynamic weight network includes: a multilayer perceptron, where each layer of the multilayer perceptron has a corresponding scene-sensitive feature weight vector and bias vector.

[0077] It should be noted that the number of layers in a multilayer perceptron is the same as the number of layers in a cross-connect network. That is, when the cross-connect network has 3 layers, the multilayer perceptron also has 3 layers.

[0078] Step S105: Based on the dynamic weight network, perform feature cross processing on the scene-sensitive feature vector and the target feature cross representation vector to obtain the high-order cross feature vector.

[0079] By improving the recommendation results and ranking weights, the dynamic weight network enables the multi-scenario recommendation model obtained in this embodiment to recommend more interesting content to users.

[0080] Step S106: Concatenate the high-order cross feature vector and the deep feature vector together, and construct a multi-scenario recommendation model through a neural network.

[0081] In ModelFusion, a concatenated feature vector is obtained by concatenating the high-order cross-feature vector obtained from the dynamic weight network with the deep feature vector obtained from the deep network. This concatenated feature vector is then used to construct a multi-scenario recommendation model through a neural network. The neural network can be a fully connected layer neural network.

[0082] In summary, this invention discloses a method for constructing a multi-scenario recommendation model. It obtains basic feature vectors and scene-sensitive feature vectors, processes the basic feature vectors using a multilayer perceptron to obtain deep feature vectors, performs feature cross-processing on the basic feature vectors to obtain target feature cross-representation vectors, constructs a dynamic weight network based on the scene-sensitive feature vectors, and then performs feature cross-processing on the scene-sensitive feature vectors and target feature cross-representation vectors using the dynamic weight network to obtain higher-order cross-feature vectors. By concatenating the higher-order cross-feature vectors and deep feature vectors, and constructing a multi-scenario recommendation model through a neural network, this invention first processes the basic feature vectors separately to obtain the target feature cross-representation vectors. The dynamic weight network then performs cross-processing on the target feature cross-representation vectors and scene-sensitive feature vectors to amplify the importance of the basic features of the recommendation model, allowing for better adjustment of subsequent scene recommendation results and scene ranking weights. This effectively avoids the problem of multi-scenario recommendation models being dominated by data-rich scenes, leading to difficulties in recommending scenes with limited data.

[0083] To further optimize the above embodiments, see [link to relevant documentation]. Figure 2The flowchart of a method for obtaining basic feature vectors and scene-sensitive feature vectors disclosed in this embodiment of the invention, specifically step S101, may include:

[0084] Step S201: Obtain the basic features of the recommendation model.

[0085] Among them, the basic features of the recommendation model include the basic features required to build multi-scenario models, such as discrete or continuous features such as user age, user gender, geographical location, tags and types of videos watched by the user.

[0086] Step S202: Convert the basic features of the recommendation model into vector form to obtain the basic feature vector.

[0087] Specifically, the basic features of the recommendation model can be embedded to convert them into vector form, thus obtaining the basic feature vector.

[0088] The basic feature vectors are used as inputs to deep networks and cross-networks.

[0089] Step S203: Obtain scene-sensitive features.

[0090] In practical applications, representative features that can distinguish different scenarios are selected as scenario-sensitive features. That is, scenario-sensitive features include distinguishing characteristics of different scenarios, such as video channels.

[0091] Step S204: Convert the scene-sensitive features into vector form to obtain the scene-sensitive feature vector.

[0092] Specifically, the scene-sensitive features are embedded and transformed into vector form to obtain the scene-sensitive feature vector, which is used as the input to the subsequent dynamic weight network.

[0093] The processes of obtaining the basic feature vectors in steps S201 and S202 are independent of the processes of obtaining the scene-sensitive feature vectors in steps S203 and S204. In practical applications, the execution order of the two processes can be... Figure 2 In the embodiments shown, steps S203 and S204 may be executed first, followed by steps S201 and S202, or both processes may be executed simultaneously, depending on the actual needs. This invention does not impose any limitations on these processes.

[0094] In summary, this invention transforms the basic features of the recommendation model into vector form of basic feature vectors, and scene-sensitive features into vector form of scene-sensitive feature vectors, to obtain the basic vectors for constructing a multi-scene recommendation model. These basic vectors include basic feature vectors and scene-sensitive feature vectors. When constructing the multi-scene recommendation model, the basic feature vectors are first processed separately to obtain the target feature cross-representation vector. A dynamic weight network is then used to cross-process the target feature cross-representation vector and the scene-sensitive feature vector to amplify the importance of the basic features of the recommendation model. This allows for better adjustment of subsequent scene recommendation results and scene ranking weights, effectively avoiding the problem of multi-scene recommendation models being dominated by data-rich scenes, making it difficult to recommend scenes with limited data.

[0095] To further optimize the above embodiments, step S103 may specifically include:

[0096] The current layer of the cross-network performs a vector transpose operation on the feature cross representation vector of the previous layer to obtain the feature cross transpose vector of the previous layer;

[0097] The target feature cross representation vector is obtained by performing feature cross between the transpose vector of the previous layer and the basic feature vector;

[0098] In this case, when the current layer is the first layer of the cross network, the feature cross representation vector of the previous layer is the basic feature vector.

[0099] It should be noted that cross-networks are multi-layered in practical applications, typically three layers. Taking a three-layer cross-network as an example, the following explains the specific process of performing feature crossing on the basic feature vectors to obtain the target feature cross-representation vector. The specific operations for each layer are as follows:

[0100] The first layer of the cross-network performs a vector transpose operation on the basic feature vector to obtain the basic feature transpose vector;

[0101] The transpose of the basic feature vector is combined with the basic feature vector through feature crossing (e.g., multiplication and addition) to obtain the first-layer feature crossing representation vector;

[0102] Based on the second layer of the cross network, the first layer feature cross representation vector is transposed to obtain the first layer feature cross transpose vector.

[0103] The first layer feature cross transpose vector is cross-referenced with the base feature vector (e.g., multiplied and added) to obtain the second layer feature cross representation vector;

[0104] Based on the third layer of the cross network, the second layer feature cross representation vector is transposed to obtain the second layer feature cross transpose vector;

[0105] The second-layer feature cross transpose vector is cross-referenced with the base feature vector (e.g., multiplied and added) to obtain the target feature cross representation vector.

[0106] It should be noted that the feature cross representation vector of each layer of the cross network is obtained by transposing the feature cross representation vector of the previous layer, multiplying it with the base feature vector, and then adding the vectors to complete the feature cross.

[0107] The cross-representation vectors of target features obtained by cross-networks are mainly used to capture non-linear relationships between features. By performing cross-operations on feature vectors, cross-networks can determine higher-order relationships between features, thereby improving the expressive power of the model.

[0108] To further optimize the above embodiments, step S105 may specifically include:

[0109] The vectors of each layer of the multilayer perceptron are cross-processed with the feature cross-representation vectors of the corresponding layers of the cross-network to extract high-level feature information. The feature vectors obtained by crossing the target feature cross-representation vector of the cross-network with the dynamic weight network are determined as high-order cross feature vectors.

[0110] The vectors for each layer of the multilayer perceptron include: scene-sensitive feature weight vector and bias vector.

[0111] Assuming that both the cross-network and the dynamic weight network of the multilayer perceptron have 3 layers, the scene-sensitive feature weight vector and bias vector of each layer of the 3-layer multilayer perceptron will be connected to the cross-network. That is, the scene-sensitive feature weight vector and bias vector of each layer will be multiplied and added with the feature cross-representation vector of each layer of the cross-network to complete the extraction of high-level feature information. The feature after the cross-representation vector of the 3rd layer of the cross-network is crossed with the dynamic weight network is taken as the high-order cross feature vector.

[0112] In summary, the present invention employs a dynamic weight network and a cross network for feature crossing, which can fit the Cartesian product feature crossing and solve the sparsity problem. It is also similar to the DNN model in achieving automatic feature crossing and preventing overgeneralization.

[0113] To further optimize the above embodiments, after constructing the multi-scenario recommendation model in step S106, the following may also be included:

[0114] The constructed multi-scenario recommendation model is mapped from multi-dimensional to one-dimensional to obtain a one-dimensional output vector.

[0115] The one-dimensional output vector is processed using the Sigmoid function to obtain the recommendation model score.

[0116] In summary, this embodiment can determine the ranking of each scenario in multiple scenarios based on the recommendation model score, thereby making it easier to recommend content that users are more interested in.

[0117] To better understand the construction process of multi-scenario recommendation models, see [link / reference]. Figure 3 This invention discloses a system block diagram of a multi-scenario recommendation model. This embodiment mainly focuses on the cross-network in the DCN network, adding a dynamic weight network on the left. By integrating the dynamic weight network into the DCN network, the expressive power of scene-sensitive features is enhanced. The construction process of the multi-scenario recommendation model is as follows:

[0118] 1) After the basic features of the recommendation model are converted into basic feature vectors, they are input into the cross network and the deep network respectively.

[0119] 2) The deep network processes the basic feature vectors through a multilayer perceptron to obtain the deep feature vectors; the cross network performs feature cross processing on the basic feature vectors to obtain the target feature cross representation vector.

[0120] 3) The scene-sensitive features are converted into scene-sensitive feature vectors, and a dynamic weight network is constructed based on the scene-sensitive feature vectors. The dynamic weight network includes a multilayer perceptron, and each layer of the multilayer perceptron has a corresponding scene-sensitive feature weight vector and bias vector.

[0121] 4) The scene-sensitive feature weight vector and bias vector of each layer of the multilayer perceptron are cross-processed with the feature cross-representation vector of the corresponding layer of the cross network to extract high-level feature information, and the feature vector after the cross-representation vector of the target feature of the cross network is crossed with the dynamic weight network is determined as the high-order cross feature vector.

[0122] 5) The high-order cross feature vectors obtained from the dynamic weight network are concatenated with the deep feature vectors of the deep network, and a multi-scenario recommendation model is constructed through a neural network.

[0123] In summary, this invention first processes the basic feature vector separately to obtain the target feature cross-representation vector. Then, it uses a dynamic weight network to cross-process the target feature cross-representation vector and the scene-sensitive feature vector to amplify the importance of the basic features of the recommendation model. This allows for better adjustment of the subsequent scene recommendation results and scene ranking weights, thereby effectively avoiding the problem that multi-scene recommendation models are dominated by scenes with abundant data, making it difficult to recommend scenes with limited data.

[0124] Corresponding to the above embodiments, the present invention also discloses a scene recommendation method. This scene recommendation method uses the multi-scene recommendation model obtained by the construction method of the multi-scene recommendation model described in the above embodiments, and the recommendation method includes:

[0125] Obtain user data;

[0126] Extract user features from the user data;

[0127] The user features are input into the multi-scenario recommendation model to obtain scenario recommendation results.

[0128] User data includes, but is not limited to, user age, user gender, geographical location, and data related to the videos watched (such as video channels).

[0129] User characteristics may include: user age, user gender, geographical location, tags and types of videos watched, and other discrete or continuous features, as well as the video channels watched.

[0130] In summary, when making scenario recommendations to users, this invention adopts a multi-scenario recommendation model that better adjusts the scenario recommendation results and scenario ranking weights. Therefore, it can recommend better scenarios to users based on their characteristics, thereby achieving personalized recommendations.

[0131] Corresponding to the above method embodiments, the present invention also discloses a device for constructing a multi-scenario recommendation model.

[0132] See Figure 4 The present invention discloses a structural schematic diagram of a device for constructing a multi-scenario recommendation model, the device comprising:

[0133] Vector acquisition unit 301 is used to acquire basic feature vectors and scene-sensitive feature vectors;

[0134] The basic feature vector contains the basic features required to build a multi-scenario recommendation model, such as discrete or continuous features like user age, user gender, geographical location, tags and types of videos watched by the user.

[0135] Scene-sensitive feature vectors contain distinguishing features for different scenes, which are representative features that differentiate different scenes, such as video channels.

[0136] It should be noted that in this embodiment, the basic feature vector is used as input to the deep network and cross network.

[0137] The first processing unit 302 is used to process the basic feature vector through a multilayer perceptron to obtain a deep feature vector;

[0138] The first processing unit 302 can be specifically used to: process the basic feature vector based on a deep network through a multilayer perceptron to obtain a deep feature vector;

[0139] Deep networks typically consist of three layers. They process the basic feature vectors using a multilayer perceptron, transforming the basic feature vectors into a deep representation of the features, resulting in deep feature vectors.

[0140] Deep feature vectors contain non-linear relationships between features and are used for subsequent model fusion with cross-networks.

[0141] The second processing unit 303 is used to perform feature crossing processing on the basic feature vector to obtain the target feature crossing representation vector.

[0142] The target feature cross representation vector can be obtained by performing feature cross processing on the basic feature vector based on the cross network.

[0143] In this embodiment, the cross network can be the cross network in the DCN (Deep & Cross Network) model.

[0144] The target feature cross-representation vectors obtained from the cross-network are mainly used to capture nonlinear relationships between features. By performing feature cross-processing on the basic feature vectors, the cross-network can determine higher-order relationships between features, thereby improving the expressive power of the model.

[0145] Network construction unit 304 is used to construct a dynamic weight network based on the scene-sensitive feature vector;

[0146] It should be noted that the number of layers in a multilayer perceptron is the same as the number of layers in a cross-connect network. That is, when the cross-connect network has 3 layers, the multilayer perceptron also has 3 layers.

[0147] The third processing unit 305 is used to perform feature cross processing on the scene-sensitive feature vector and the target feature cross representation vector based on the dynamic weight network to obtain a higher-order cross feature vector.

[0148] By improving the recommendation results and ranking weights, the dynamic weight network enables the multi-scenario recommendation model obtained in this embodiment to recommend more interesting content to users.

[0149] The model building unit 306 is used to concatenate the high-order cross feature vector and the deep feature vector together, and construct a multi-scenario recommendation model through a neural network.

[0150] In model fusion, a concatenated feature vector is obtained by concatenating the high-order cross feature vector obtained by the dynamic weight network with the deep feature vector obtained by the deep network. The concatenated feature vector can then be used to construct a multi-scenario recommendation model through a neural network.

[0151] In summary, this invention discloses a device for constructing a multi-scene recommendation model. It obtains basic feature vectors and scene-sensitive feature vectors, processes the basic feature vectors using a multilayer perceptron to obtain deep feature vectors, performs feature cross processing on the basic feature vectors to obtain target feature cross representation vectors, constructs a dynamic weight network based on the scene-sensitive feature vectors, and then performs feature cross processing on the scene-sensitive feature vectors and target feature cross representation vectors based on the dynamic weight network to obtain higher-order cross feature vectors. By concatenating the higher-order cross feature vectors and deep feature vectors, and constructing a multi-scene recommendation model through a neural network, this invention first processes the basic feature vectors separately to obtain the target feature cross representation vectors. The dynamic weight network then performs cross processing on the target feature cross representation vectors and scene-sensitive feature vectors to amplify the importance of the basic features of the recommendation model, allowing for better adjustment of subsequent scene recommendation results and scene ranking weights. This effectively avoids the problem of multi-scene recommendation models being dominated by data-rich scenes, making it difficult to recommend scenes with limited data.

[0152] To further optimize the above embodiments, the vector acquisition unit 301 may include:

[0153] The first acquisition subunit is used to acquire the basic features of the recommendation model;

[0154] The first transformation subunit is used to transform the basic features of the recommendation model into vector form to obtain the basic feature vector;

[0155] The second acquisition subunit is used to acquire scene-sensitive features, which include distinguishing features of different scenes;

[0156] The second transformation subunit is used to transform the scene-sensitive features into vector form to obtain the scene-sensitive feature vector.

[0157] In summary, this invention transforms the basic features of the recommendation model into vector form of basic feature vectors, and scene-sensitive features into vector form of scene-sensitive feature vectors, to obtain the basic vectors for constructing a multi-scene recommendation model. These basic vectors include basic feature vectors and scene-sensitive feature vectors. When constructing the multi-scene recommendation model, the basic feature vectors are first processed separately to obtain the target feature cross-representation vector. A dynamic weight network is then used to cross-process the target feature cross-representation vector and the scene-sensitive feature vector to amplify the importance of the basic features of the recommendation model. This allows for better adjustment of subsequent scene recommendation results and scene ranking weights, effectively avoiding the problem of multi-scene recommendation models being dominated by data-rich scenes, making it difficult to recommend scenes with limited data.

[0158] To further optimize the above embodiments, the second processing unit 303 may specifically be used for:

[0159] Based on the current layer of the cross network, the feature cross representation vector of the previous layer is transposed to obtain the feature cross transpose vector of the previous layer;

[0160] The target feature cross representation vector is obtained by performing feature cross between the transpose vector of the previous layer and the basic feature vector;

[0161] Wherein, when the current layer is the first layer of the cross network, the feature cross representation vector of the previous layer is the basic feature vector.

[0162] It should be noted that cross-networks are multi-layered in practical applications, typically three layers. Taking a three-layer cross-network as an example, the following explains the specific process of performing feature crossing on the basic feature vectors to obtain the target feature cross-representation vector. The specific operations for each layer are as follows:

[0163] The first layer of the cross-network performs a vector transpose operation on the basic feature vector to obtain the basic feature transpose vector;

[0164] The transpose of the basic feature vector is combined with the basic feature vector through feature crossing (e.g., multiplication and addition) to obtain the first-layer feature crossing representation vector;

[0165] Based on the second layer of the cross network, the first layer feature cross representation vector is transposed to obtain the first layer feature cross transpose vector.

[0166] The first layer feature cross transpose vector is cross-referenced with the base feature vector (e.g., multiplied and added) to obtain the second layer feature cross representation vector;

[0167] Based on the third layer of the cross network, the second layer feature cross representation vector is transposed to obtain the second layer feature cross transpose vector;

[0168] The second-layer feature cross transpose vector is cross-referenced with the base feature vector (e.g., multiplied and added) to obtain the target feature cross representation vector.

[0169] To further optimize the above embodiments, the network construction unit 304 can specifically be used for:

[0170] The scene-sensitive feature vector is passed through a fully connected neural network to obtain a fixed-dimensional scene-sensitive feature vector;

[0171] The fixed-dimensional scene-sensitive feature vector is reshaped to obtain the dynamic weight network.

[0172] The dynamic weight network includes a multilayer perceptron, the number of layers of which is the same as the number of layers of the cross network, and each layer of the multilayer perceptron has a corresponding scene-sensitive feature weight vector and bias vector.

[0173] To further optimize the above embodiments, the third processing unit 305 may specifically be used for:

[0174] The vectors of each layer of the multilayer perceptron are cross-processed with the feature cross-representation vectors of the corresponding layers of the cross network to extract high-level feature information. The feature vector obtained by crossing the target feature cross-representation vector of the cross network with the dynamic weight network is determined as the high-order cross feature vector.

[0175] The vectors for each layer of the multilayer perceptron include: scene-sensitive feature weight vector and bias vector.

[0176] Assuming that both the cross-network and the dynamic weight network of the multilayer perceptron have 3 layers, the scene-sensitive feature weight vector and bias vector of each layer of the 3-layer multilayer perceptron will be connected to the cross-network. That is, the scene-sensitive feature weight vector and bias vector of each layer will be multiplied and added with the feature cross-representation vector of each layer of the cross-network to complete the extraction of high-level feature information. The feature after the cross-representation vector of the 3rd layer of the cross-network is crossed with the dynamic weight network is taken as the high-order cross feature vector.

[0177] In summary, the present invention employs a dynamic weight network and a cross network for feature crossing, which can fit the Cartesian product feature crossing and solve the sparsity problem. It is also similar to the DNN model in achieving automatic feature crossing and preventing overgeneralization.

[0178] To further optimize the above embodiments, the construction apparatus may further include:

[0179] The mapping unit is used to map the constructed multi-scenario recommendation model from multi-dimensional to one-dimensional to obtain a one-dimensional output vector;

[0180] The fourth processing unit is used to process the one-dimensional output vector using the Sigmoid function to obtain the recommendation model score.

[0181] In summary, this embodiment can determine the ranking of each scenario in multiple scenarios based on the recommendation model score, thereby making it easier to recommend content that users are more interested in.

[0182] It should be noted that for the specific working principles of each component in the device embodiment, please refer to the corresponding section of the method embodiment, which will not be repeated here.

[0183] Corresponding to the above embodiments, the present invention also discloses a scene recommendation device. This scene recommendation method employs the multi-scene recommendation model obtained by the construction method of the multi-scene recommendation model described in the above embodiments. The recommendation device includes:

[0184] The data acquisition unit is used to acquire user data;

[0185] A feature extraction unit is used to extract user features from the user data;

[0186] The recommendation unit is used to input the user features into the multi-scenario recommendation model to obtain scenario recommendation results.

[0187] User data includes, but is not limited to, user age, user gender, geographical location, and data related to the videos watched (such as video channels).

[0188] User characteristics may include: user age, user gender, geographical location, tags and types of videos watched, and other discrete or continuous features, as well as the video channels watched.

[0189] In summary, when making scenario recommendations to users, this invention adopts a multi-scenario recommendation model that better adjusts the scenario recommendation results and scenario ranking weights. Therefore, it can recommend better scenarios to users based on their characteristics, thereby achieving personalized recommendations.

[0190] Corresponding to the above embodiments, such as Figure 5 As shown, the present invention also provides an electronic device, which may include: a processor 1 and a memory 2;

[0191] The processor 1 and memory 2 communicate with each other via communication bus 3.

[0192] Processor 1, for executing at least one instruction;

[0193] Memory 2 is used to store at least one instruction;

[0194] Processor 1 may be a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits configured to implement embodiments of the present invention.

[0195] Memory 2 may include high-speed RAM memory, and may also include non-volatile memory, such as at least one disk storage device.

[0196] The processor executes at least one instruction to implement the embodiment of the above-described method for constructing a multi-scenario recommendation model.

[0197] Corresponding to the above embodiments, the present invention also discloses a computer-readable storage medium that stores at least one instruction, which, when executed by a processor, implements the embodiments shown in the above-described method for constructing a multi-scenario recommendation model.

[0198] In summary, this invention discloses an electronic device and a computer-readable storage medium. It acquires basic feature vectors and scene-sensitive feature vectors, processes the basic feature vectors using a multilayer perceptron to obtain deep feature vectors, performs feature cross processing on the basic feature vectors to obtain target feature cross representation vectors, constructs a dynamic weight network based on the scene-sensitive feature vectors, and then performs feature cross processing on the scene-sensitive feature vectors and target feature cross representation vectors based on the dynamic weight network to obtain higher-order cross feature vectors. By concatenating the higher-order cross feature vectors and deep feature vectors, and constructing a multi-scene recommendation model through a neural network, this invention first processes the basic feature vectors separately to obtain the target feature cross representation vectors. Then, it uses a dynamic weight network to perform cross processing on the target feature cross representation vectors and scene-sensitive feature vectors to amplify the importance of the basic features of the recommendation model, thereby better adjusting the subsequent scene recommendation results and scene ranking weights. This effectively avoids the problem that the multi-scene recommendation model is dominated by data-rich scenes, making it difficult to recommend scenes with limited data.

[0199] Finally, it should be noted that in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0200] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0201] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for constructing a multi-scenario recommendation model, characterized in that, include: Obtain basic feature vectors and scene-sensitive feature vectors. The basic feature vectors contain the basic features required to build a multi-scene recommendation model, and the scene-sensitive feature vectors contain distinguishing features for different scenes. The distinguishing features include video channels. The basic feature vectors are processed by a multilayer perceptron to obtain deep feature vectors; The basic feature vectors are subjected to feature cross processing to obtain the target feature cross representation vector; A dynamic weighted network is constructed based on the scene-sensitive feature vectors. Based on the dynamic weight network, feature cross processing is performed on the scene-sensitive feature vector and the target feature cross representation vector to obtain a higher-order cross feature vector; The higher-order cross feature vector and the deep feature vector are concatenated together, and a multi-scenario recommendation model is obtained by constructing a neural network.

2. The construction method according to claim 1, characterized in that, The acquisition of the basic feature vector and the scene-sensitive feature vector includes: Obtain the basic features of the recommendation model, which include the basic features required to build a multi-scenario model; The basic features of the recommendation model are transformed into vector form to obtain the basic feature vector; Obtain scene-sensitive features, which include distinguishing features for different scenes; The scene-sensitive features are converted into vector form to obtain the scene-sensitive feature vector.

3. The construction method according to claim 1, characterized in that, The step of performing feature cross processing on the basic feature vector to obtain the target feature cross representation vector includes: The current layer of the cross-network performs a vector transpose operation on the feature cross representation vector of the previous layer to obtain the feature cross transpose vector of the previous layer; The target feature cross representation vector is obtained by performing feature cross between the transpose vector of the previous layer and the basic feature vector; Wherein, when the current layer is the first layer of the cross network, the feature cross representation vector of the previous layer is the basic feature vector.

4. The construction method according to any one of claims 1 to 3, characterized in that, The construction of the dynamic weight network based on the scene-sensitive feature vector includes: The scene-sensitive feature vector is passed through a fully connected neural network to obtain a fixed-dimensional scene-sensitive feature vector; The fixed-dimensional scene-sensitive feature vector is reshaped to obtain the dynamic weight network; The dynamic weight network includes a multilayer perceptron, the number of layers of which is the same as the number of layers of the cross network, and each layer of the multilayer perceptron has a corresponding scene-sensitive feature weight vector and bias vector.

5. The construction method according to claim 4, characterized in that, The step of performing feature cross processing on the scene-sensitive feature vector and the target feature cross representation vector based on the dynamic weight network to obtain a higher-order cross feature vector includes: The vectors of each layer of the multilayer perceptron are cross-processed with the feature cross-representation vectors of the corresponding layers of the cross network to extract high-level feature information, and the feature vector obtained by crossing the target feature cross-representation vector of the cross network with the dynamic weight network is determined as the high-order cross feature vector. The vectors of each layer of the multilayer perceptron include: the scene-sensitive feature weight vector and the bias vector.

6. The construction method according to claim 1, characterized in that, Also includes: The constructed multi-scenario recommendation model is mapped from multi-dimensional to one-dimensional to obtain a one-dimensional output vector; The one-dimensional output vector is processed using the Sigmoid function to obtain the recommendation model score.

7. A scenario recommendation method, characterized in that, The multi-scenario recommendation model obtained by constructing the multi-scenario recommendation model according to any one of claims 1 to 6, wherein the method includes: Obtain user data; Extract user features from the user data; The user features are input into the multi-scenario recommendation model to obtain scenario recommendation results.

8. A device for constructing a multi-scenario recommendation model, characterized in that, include: The vector acquisition unit is used to acquire basic feature vectors and scene-sensitive feature vectors. The basic feature vectors contain the basic features required to build a multi-scene recommendation model, and the scene-sensitive feature vectors contain distinguishing features for different scenes. The distinguishing features include video channels. The first processing unit is used to process the basic feature vector through a multilayer perceptron to obtain a deep feature vector; The second processing unit is used to perform feature cross processing on the basic feature vector to obtain the target feature cross representation vector. Network construction unit, used to construct a dynamic weight network based on the scene-sensitive feature vector; The third processing unit is used to perform feature cross processing on the scene-sensitive feature vector and the target feature cross representation vector based on the dynamic weight network to obtain a higher-order cross feature vector. The model building unit is used to concatenate the high-order cross feature vector and the deep feature vector together, and construct a multi-scenario recommendation model through a neural network.

9. A scene recommendation device, characterized in that, The multi-scenario recommendation model obtained by constructing the multi-scenario recommendation model according to any one of claims 1 to 6, wherein the apparatus comprises: The data acquisition unit is used to acquire user data; A feature extraction unit is used to extract user features from the user data; The recommendation unit is used to input the user features into the multi-scenario recommendation model to obtain scenario recommendation results.

10. An electronic device, characterized in that, The electronic device includes: a memory and a processor; The memory is used to store at least one instruction; The processor is used to execute the at least one instruction to implement the method for constructing a multi-scenario recommendation model as described in any one of claims 1 to 6.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores at least one instruction, which, when executed by a processor, implements the method for constructing a multi-scenario recommendation model as described in any one of claims 1 to 6.

Images