A short video multi-behavior click prediction method based on multi-task learning

Abstract

The application discloses a short video multi-behavior click prediction method based on multi-task learning, and steps are as follows: extracting user basic features and short video basic features; introducing a word embedding model and a graph embedding model to fully learn time sequence information in a user historical behavior sequence and extract sequence features; establishing an MMOE multi-task model to learn the connection and difference of different behaviors; and introducing an FM model to learn low-order feature interaction on the basis; assigning different weights to user behaviors such as viewing, commenting, liking, clicking an avatar and forwarding, performing weighted average, and predicting a probability value of a short video click behavior of the user. The application introduces the word embedding model and the graph embedding model to extract the sequence features, introduces the FM model on the basis of the MMOE model, can not only learn the connection and difference of different behaviors, but also learn low-order feature interaction, so that the model has low-order feature memory performance and high-order feature generalization performance, and the click prediction accuracy is improved.

Description

Technical Field

[0001] This invention belongs to the field of information recommendation, and specifically relates to a method for predicting multi-behavior clicks in short videos based on multi-task learning. Background Technology

[0002] In recent years, with the advent of the big data era, a massive amount of short video information has been generated, making it difficult for people to find videos they like. This has created an urgent need for a system capable of personalized video recommendations. Therefore, personalized short video recommendation systems have emerged, with click-through rate (CTR) prediction being a key issue. Personalized short video recommendation systems predict the probability of a user clicking on a recommended short video, then sort the recommendation list and recommend the short video with the highest predicted probability to the user. This achieves personalized recommendations, improves the user viewing experience, increases user engagement, and ultimately boosts platform revenue while increasing traffic.

[0003] Currently, the mainstream method for predicting short video click behavior is based on deep learning. For multiple unrelated learning tasks, it is generally chosen to build a model for each task separately. However, when there is some potential relationship between multiple tasks, such as only being able to post a comment after viewing one, if the approach of modeling separately is still adopted, the correlation between multiple tasks is ignored. Deep learning models map large-scale sparse input features to low-dimensional embedding vectors and then convert them into fixed-length vectors, making the user interest representation invariant, but ignoring the fact that user behavior may change dynamically over time in real life. Summary of the Invention

[0004] The purpose of this invention is to provide a method for predicting multi-behavior clicks in short videos based on multi-task learning, so as to solve the problems of the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for predicting multi-behavior clicks in short videos based on multi-task learning, comprising the following steps:

[0006] S1. Extract basic features from the user behavior dataset and the short video dataset respectively to obtain the user basic feature set and the short video basic feature set; then select and merge features from the user basic feature set and the short video basic feature set to obtain the user basic features and the short video basic features respectively.

[0007] S2. The user behavior features extracted from the user basic feature set are divided into user viewing sequences according to the chronological order. Then, the viewing sequences of all users are obtained. The viewing sequences of all users are used as input and their corresponding short video word vector features are used as output to build and train a sequence feature extraction model.

[0008] S3. Based on the user viewing sequence, generate a user viewing path map. Using the user viewing sequence as input and the user viewing sequence vector features as output, construct and train a user sequence feature extraction model; using the user viewing path map as input and the user potential viewing sequence vector features as output, construct and train a user potential viewing sequence feature extraction model.

[0009] S4. Using user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features as inputs, and their corresponding user behavior prediction probabilities as outputs, construct and train a user multi-behavior click prediction model, and assign weights to user behaviors during the training process.

[0010] Furthermore, the aforementioned step S1 includes the following sub-steps:

[0011] S101. Extract user ID features and user behavior features from the user behavior dataset to obtain the basic user feature set; extract short video features and short video multimodal information from the short video dataset to obtain the basic short video feature set.

[0012] S102. Select and merge features from user ID features and user behavior features to obtain basic user features; merge user ID features and short video features to obtain basic short video features.

[0013] Furthermore, the aforementioned step S2 includes the following sub-steps:

[0014] S201. The user behavior features extracted from the user basic feature set are divided according to the chronological order to obtain the user viewing sequence as follows:

[0015] W1 = [w1, w2, ..., w T ],

[0016] Where w i This represents the video watched at the i-th timestamp, where i = 1, 2, ..., T;

[0017] S202, Obtain the viewing sequence of all m users as follows:

[0018] W m =[W1,W2,...,W m ],

[0019] S203. Using the viewing sequences of m users as input and their corresponding short video word vector features as output, a sequence feature extraction model is constructed and trained using the Word2Vec word embedding method.

[0020] Furthermore, in the aforementioned step S3, generating the user viewing path graph specifically involves: based on the user's viewing sequence W1 = [w1, w2, ..., w...], ... T Generate the corresponding user viewing path graph G = (V, E), where V represents the set of short video nodes in the graph and E represents the set of time viewing order edges in the graph.

[0021] Furthermore, in step S3 above, when constructing the user's potential viewing sequence model, the Random Walk algorithm is used to calculate the probability of the user's potential viewing sequence from the user viewing path graph as follows:

[0022]

[0023] Where M is the adjacency matrix of G, M ij N represents the weighted edge from node i to node j. + (v i ) represents all nodes v i The set of next-hop nodes.

[0024] Furthermore, in step S3 above, when constructing the user sequence feature extraction model, the Skip-gram algorithm is used, and the following optimization method is employed:

[0025] S301. Using the user viewing sequence as input and the user viewing sequence vector features as output, construct the optimization objective as follows:

[0026]

[0027] Where w is the window size selected by the Skip-gram algorithm, and v is the node size. i The probabilities of reaching nodes within each window are independent, and Φ is a mapping function: Φ:V→R d d represents the dimension, and Pr() represents the probability function;

[0028] S302. The probability transition formula in step S301 is further optimized as follows:

[0029]

[0030] S303. Add negative sampling to optimize the result of step S302 as follows:

[0031]

[0032] Wherein, N(v) i )' is node v i The negative sampling is σ, where σ is the activation function.

[0033] Furthermore, the aforementioned step S4 includes the following sub-steps:

[0034] S401. Using the MMOE model, with user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features as inputs, the model learns the relationships and differences between these features to obtain their corresponding higher-order features. The output f of the k-th subtask of the MMOE model is... k As shown in the following formula:

[0035]

[0036] Where n is the number of expert subnets, i is the expert subnet number, k is the task number, x is the input feature of the model, and g k (x) i f represents the weight of the i-th expert subnet in the k-th task. i (x) is the output of the i-th expert subnetwork. is the linear transformation matrix of the k-th gated unit corresponding to the i-th expert subnet, and softmax is the activation function;

[0037] S402. Using the FM model, with user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features as input, the interaction of each feature is learned to obtain its corresponding low-order features as follows:

[0038]

[0039]

[0040] Where w0∈R, w∈R n v i v j ∈R k n is the number of features, w0 is the global bias, k is the dimension of the latent vector for each feature, and the parameters w and latent vector v are also present. i ,v j Iterative updates are performed during each training session.

[0041] S403. In step S401, a fully connected layer is added after the output of each subtask of the MMOE model. The output of each subtask is concatenated with the output of the FM model in step S402 and input into the tower network of the MMOE model.

[0042] Furthermore, in step S4 above, uAUC and GAUC are used as evaluation metrics during the training of the user multi-behavior click prediction model, and user behaviors are assigned weights, specifically including:

[0043] Calculate uAUC using the following formula:

[0044]

[0045] Where n is the number of valid users in the test set, and AUC i Let AUC be the prediction result for the i-th effective user;

[0046] Calculate GAUC using the following formula:

[0047]

[0048] Among them, uAUC i For the uAUC value of the i-th action, W i Let be the weight value of the i-th action.

[0049] Furthermore, the aforementioned user behaviors include user viewing behavior, user liking behavior, user double-tapping avatar behavior, user forwarding behavior, user favorite behavior, user posting behavior, and user following behavior.

[0050] Furthermore, the aforementioned method for predicting clicks on multiple behaviors in short videos based on multi-task learning sets the weights for user viewing behavior (4), user liking behavior (3), user double-tapping avatar behavior (2), user forwarding behavior (1), user favorite behavior (1), user posting behavior (1), and user following behavior (1).

[0051] Compared with existing technologies, the beneficial effects of this invention are as follows: Compared with current mainstream click-through rate prediction models, this invention fills the gap in the MMOE model's utilization of the temporal feature information contained in user behavior. It introduces the Word2Vec word embedding model and the graph embedding model to fully learn the temporal information contained in the user's historical behavior sequence, extract sequence features, and inputs the sequence features, user basic features, short video basic features, short video keywords and short video classification tag features, as well as multimodal features that integrate video text, images, and sound information into the MMOE model to learn the connections and differences between different behaviors. Secondly, it introduces the FM model on the basis of the MMOE model to learn low-order feature interactions, so that the model has the memory performance of low-order features, thereby improving the accuracy of click-through rate prediction. Attached Figure Description

[0052] Figure 1 This is a flowchart of the present invention. Detailed Implementation

[0053] To better understand the technical content of the present invention, specific embodiments are described below in conjunction with the accompanying drawings.

[0054] In this invention, various aspects of the invention are described with reference to the accompanying drawings, in which numerous illustrative embodiments are shown. Embodiments of the invention are not limited to those depicted in the drawings. It should be understood that the invention is implemented through any of the various concepts and embodiments described above, as well as the concepts and embodiments described in detail below, because the concepts and embodiments disclosed herein are not limited to any particular implementation. Furthermore, some aspects of the invention disclosed may be used alone or in any suitable combination with other aspects of the invention disclosed.

[0055] like Figure 1 As shown, a short video multi-behavior click prediction method based on multi-task learning includes the following steps:

[0056] S1. Extract basic features from user behavior datasets and short video datasets through feature engineering to obtain user basic feature sets and short video basic feature sets.

[0057] The user basic feature set includes ID features: user ID, video ID, video author ID, background music artist ID, and background music ID; and user behavior features: including the duration of video playback, the duration of time spent watching videos, and seven behaviors such as viewing comments, liking, clicking on avatars, forwarding, collecting, posting comments, and following.

[0058] The basic feature set of short videos includes short video features: video duration, video keywords, and video tags;

[0059] Short video multimodal information: includes multimodal feature information of video text, images, and sound.

[0060] Next, features were selected and merged from the basic feature set. Features were selected from ID features and user behavior features and merged to form basic user features. Features were selected from ID features and short video features and merged to form basic short video features.

[0061] S2. First, the user behavior features extracted from the user basic feature set are sorted in chronological order. For a specific user, their behavior data is arranged into a sequence in chronological order, denoted as W1 = [w1, w2, ..., w...]. T ] indicates that w i This represents the video watched at the i-th timestamp; then, the viewing sequence of all m users is obtained as follows:

[0062] W m =[W1,W2,...,W m ],

[0063] Finally, using the viewing sequences of m users as input and their corresponding short video word vector features as output, a sequence feature extraction model is constructed and trained using the Word2Vec word embedding method.

[0064] S3. Using the user's viewing sequence as input and the user's viewing sequence vector features as output, construct and train a user sequence feature extraction model:

[0065] First, based on the user's viewing sequence W1 = [w1, w2, ..., w T For example, if a user watches 100 short videos, then these 100 short videos can be treated as nodes and sorted according to the viewing time to form a one-way path with 100 nodes. If the above operation is performed on n users, n paths will be generated. The intersection of these n paths forms a graph G = (V, E), where V represents the set of short video nodes in the graph and E represents the set of edges in the graph based on the viewing time.

[0066] Using the user's viewing path map as input and the user's potential viewing sequence feature vector as output, a user potential viewing sequence feature extraction model is constructed and trained.

[0067] The Random Walk algorithm is used to obtain potential viewing sequences for users: A Random Walk is performed on the generated graph. First, the number of paths to be generated, q, and the number of nodes in each path, p, are set. Then, a node is randomly selected, and a random walk is performed until a path with p nodes is generated. This operation is performed q times to generate q paths. The transition probability formula for the random walk is defined as:

[0068]

[0069] Where M is the adjacency matrix of G, M ij N represents the weighted edge from node i to node j. + (v i ) represents all nodes v i The next hop node set is obtained by random walk, which is based on the probability value assigned according to the edge weight. A large number of sequences are obtained from the graph through this probability value.

[0070] Secondly, the user's viewing sequence is input into the Skip-gram model for training: q paths are treated as q sentences and input into the Skip-gram model for training, finally obtaining the embedding vector of the short video. The optimization objective is:

[0071]

[0072] In the formula: w is the window size selected by the Skip-gram model, assuming node v iThe probabilities of reaching nodes within each window are independent, and Φ is a mapping function: Φ:V→R d Let d represent the dimension, then the transition probability formula can be further optimized as follows:

[0073]

[0074] To reduce computational complexity, negative sampling was introduced to optimize the objective formula.

[0075]

[0076] In the formula: N(v) i )' is node v i The negative sampling is σ, where σ is the activation function.

[0077] S4. Using user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features as inputs, and their corresponding user behavior prediction probabilities as outputs, construct and train a user multi-behavior click prediction model, and assign weights to user behaviors during the training process. Specifically:

[0078] First, using the MMOE model, with user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features as input, the model learns the relationships and differences between these features to obtain their corresponding higher-order features. The MMOE model divides the parameter-sharing layer into several expert subnets, each of which is a multilayer perceptron responsible for independently learning the coupling relationships between multiple tasks. Different expert subnets do not share parameters. Simultaneously, the MMOE model sets up a gating unit for each subtask, responsible for calculating the weights of each expert subnet in the subtask, making the selection of expert subnets more flexible for different tasks and avoiding mutual interference between weakly related tasks. The output f of the k-th subtask of the MMOE model is... k As shown in the following formula:

[0079]

[0080] Where n is the number of expert subnets, i is the expert subnet number, k is the task number, x is the input feature of the model, and g k (x) i f represents the weight of the i-th expert subnet in the k-th task. i (x) is the output of the i-th expert subnetwork. is the linear transformation matrix of the k-th gated unit corresponding to the i-th expert subnet, and softmax is the activation function;

[0081] Secondly, an FM model is introduced at the output of the MMOE model. Using the FM model, user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features are taken as inputs, and their corresponding low-order features are taken as outputs. The interaction of each feature is learned to obtain its corresponding low-order features as follows:

[0082]

[0083] Where w0∈R, w∈R n v i v j ∈R k n is the number of features, w0 is the global bias, k is the dimension of the latent vector for each feature, and the parameters w and latent vector v are also present. i ,v j Iterative updates are performed during each training session.

[0084] A fully connected layer is added after the output of each subtask in the MMOE model, concatenating the output of each subtask with the output of the FM model, and then inputting it into the tower network of the MMOE model.

[0085] This paper assigns different weights to seven user behaviors: viewing comments, liking, clicking on profile pictures, forwarding, saving, posting comments, and following. uAUC and GAUC are used as evaluation metrics, with uAUC serving as the evaluation metric for each user behavior. Since the behavior of different users in short video click prediction scenarios is quite complex, this paper uses GAUC as the evaluation metric. Different weights are assigned to the uAUC of the four user behaviors, and a weighted average is performed to reduce the impact of ranking distortion among different users. The weights are set as follows: viewing behavior (4), liking behavior (3), double-clicking profile picture (2), forwarding behavior (1), saving behavior (1), posting behavior (1), and following behavior (1).

[0086] In training the user multi-behavior click prediction model, uAUC and GAUC are used as evaluation metrics. User behaviors are assigned weights, and uAUC is calculated using the following formula:

[0087]

[0088] Where n is the number of valid users in the test set, and AUC i Let AUC be the prediction result for the i-th effective user;

[0089] Calculate GAUC using the following formula:

[0090]

[0091] Among them, uAUC i For the uAUC value of the i-th action, W i This is the weight value for the i-th action. Although the present invention has been described above with reference to preferred embodiments, it is not intended to limit the invention. Those skilled in the art can make various modifications and refinements without departing from the spirit and scope of the invention. Therefore, the scope of protection of the present invention shall be determined by the claims.

Claims

1. A method for predicting multi-behavior clicks in short videos based on multi-task learning, characterized in that, Includes the following steps: S1. Extract basic features from the user behavior dataset and the short video dataset respectively to obtain the user basic feature set and the short video basic feature set; then select and merge features from the user basic feature set and the short video basic feature set to obtain the user basic features and the short video basic features respectively. S2. The user behavior features extracted from the user basic feature set are divided into user viewing sequences according to the chronological order. Then, the viewing sequences of all users are obtained. The viewing sequences of all users are used as input and their corresponding short video word vector features are used as output to build and train a sequence feature extraction model. S3. Generate a user viewing path map based on the user viewing sequence. Using the user viewing sequence as input and the user viewing sequence vector features as output, construct and train a user sequence feature extraction model. Using the user viewing path map as input and the user's potential viewing sequence vector features as output, a user potential viewing sequence feature extraction model is constructed and trained. Specifically, when constructing the user sequence feature extraction model, the Skip-gram algorithm is used, and the following optimization method is employed: S301. Using the user viewing sequence as input and the user viewing sequence vector features as output, construct the optimization objective as follows: Where w is the window size selected by the Skip-gram algorithm, and v is the node size. i The probabilities of reaching nodes within each window are independent, and Φ is a mapping function: Φ:V→R d d represents the dimension, and Pr() represents the probability function; S302. The probability transition formula in step S301 is further optimized as follows: S303. Add negative sampling to optimize the result of step S302 as follows: Wherein, N(v) i )′ is node v i Negative sampling, where σ is the activation function; S4. Using user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features as inputs, and their corresponding user behavior prediction probabilities as outputs, construct and train a user multi-behavior click prediction model, and assign weights to user behaviors during the training process; specifically including the following sub-steps: S401. Using the MMOE model, with user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features as inputs, learn the relationship and differences between each feature and obtain its corresponding higher-order features. S402. Using the FM model, with user basic features, short video basic features, short video word vector features, user viewing sequence vector features, and user potential viewing sequence vector features as input, learn the interaction of each feature to obtain its corresponding low-order features. S403. In step S401, a fully connected layer is added after the output of each subtask of the MMOE model. The output of each subtask is concatenated with the output of the FM model in step S402 and input into the tower network of the MMOE model.

2. The method for predicting multi-behavior clicks in short videos based on multi-task learning according to claim 1, characterized in that... Step S1 includes the following sub-steps: S101. Extract user ID features and user behavior features from the user behavior dataset to obtain the basic user feature set; extract short video features and short video multimodal information from the short video dataset to obtain the basic short video feature set. S102. Select and merge features from user ID features and user behavior features to obtain basic user features; merge user ID features and short video features to obtain basic short video features.

3. The short video multi-behavior click prediction method based on multi-task learning according to claim 2, characterized in that, Step S2 includes the following sub-steps: S201. The user behavior features extracted from the user basic feature set are divided according to the chronological order to obtain the user viewing sequence as follows: W1=[w1,w2,...,w T ], Where w i This represents the video watched at the i-th timestamp, where i = 1, 2, ..., T; S202, Obtain the viewing sequence of all m users as follows: IN m =[W1,W2,...,W m ], S203. Using the viewing sequences of m users as input and their corresponding short video word vector features as output, a sequence feature extraction model is constructed and trained using the Word2Vec word embedding method.

4. The short video multi-behavior click prediction method based on multi-task learning according to claim 3, characterized in that, In step S3, generating the user viewing path graph specifically involves: based on the user's viewing sequence W1 = [w1, w2, ..., w...] T Generate the corresponding user viewing path graph G = (V, E), where V represents the set of short video nodes in the graph and E represents the set of time viewing order edges in the graph.

5. The short video multi-behavior click prediction method based on multi-task learning according to claim 4, characterized in that, In step S3, when constructing the user's potential viewing sequence model, the Random Walk algorithm is used to calculate the probability of the user's potential viewing sequence from the user viewing path graph as follows: Where M is the adjacency matrix of G, M ij N represents the weighted edge from node i to node j. + (v i ) represents all nodes v i The set of next-hop nodes.

6. The short video multi-behavior click prediction method based on multi-task learning according to claim 5, characterized in that, In step S401, the output f of the k-th subtask of the MMOE model k As shown in the following formula: Where n is the number of expert subnets, i is the expert subnet number, k is the task number, x is the input feature of the model, and g k (x) i f represents the weight of the i-th expert subnet in the k-th task. i (x) is the output of the i-th expert subnetwork. is the linear transformation matrix of the k-th gated unit corresponding to the i-th expert subnet, and softmax is the activation function; In step S402, the corresponding low-order features are obtained as follows: Where w0∈R, w∈R n v i v j ∈R k n is the number of features, w0 is the global bias, k is the dimension of the latent vector for each feature, and the parameters w and latent vector v are also present. i ,v j Iterative updates are performed during each training session.

7. The short video multi-behavior click prediction method based on multi-task learning according to claim 6, characterized in that, In step S4, uAUC and GAUC are used as evaluation metrics during the training of the user multi-behavior click prediction model, and user behaviors are assigned weights, specifically including: Calculate uAUC using the following formula: Where n is the number of valid users in the test set, and AUC i Let AUC be the prediction result for the i-th effective user; Calculate GAUC using the following formula: Among them, uAUC i For the uAUC value of the i-th action, W i Let be the weight value of the i-th action.

8. The short video multi-behavior click prediction method based on multi-task learning according to claim 7, characterized in that, User behavior includes user viewing behavior, user liking behavior, user double-tapping avatar behavior, user forwarding behavior, user saving behavior, user posting behavior, and user following behavior.

9. The short video multi-behavior click prediction method based on multi-task learning according to claim 8, characterized in that, The weights for user viewing behavior are set to 4, user liking behavior to 3, user double-tapping avatar behavior to 2, user forwarding behavior to 1, user favorite behavior to 1, user posting behavior to 1, and user following behavior to 1.

Images