Short video recommendation method and device based on heterogeneous multi-interest modeling

Abstract

The application discloses a short video recommendation method and device based on heterogeneous multi-interest modeling. The method comprises the following steps: obtaining a historical interactive short video sequence of a user; constructing short video category embedding vectors which are orthogonal to each other; calculating classification scores of each short video in the historical interactive short video sequence on each short video category, so as to obtain classification probabilities of each short video on each short video category; sequentially weighting and converging short video embedding in each short video category according to the classification probabilities, so as to obtain soft interest representations of the user on each short video category; after dividing the short video into the category with the maximum probability, cutting the historical interactive short video sequence according to the category of the short video, and calculating vector representations of sub-sequences through a recurrent neural network, hard interest representations of the user on each short video category are obtained; and based on the soft interest representations and the hard interest representations, a short video recommendation result of the user is obtained. The application can realize diversified and accurate short video recommendation.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of recommendation system, and particularly relates to a short video recommendation method and device based on heterogeneous multi-interest modeling. BACKGROUND

[0002] In recent years, short video platforms such as TikTok, Kuaishou, MX TakaTak, etc. have become increasingly popular, and the number of short videos has exploded. Overall, short video applications display a single video in full-screen mode each time and play automatically and repeatedly. Usually, when a user sees the cover of a short video, the user can determine whether the user is interested in the video. With the increase in the number of short videos, if the short videos recommended by the platform do not meet the user's interests, the user is likely to leave the application. Therefore, efficient short video recommendation has become an important task.

[0003] In order to quickly generate personalized recommendation results from millions of candidate short videos, consistent with other large-scale recommendation systems, short video recommendation usually includes two stages, namely, a recall stage and a ranking stage. The present application mainly focuses on the short video recall stage, which requires that the recommendation method can quickly recall several hundred or several thousand short videos from millions or hundreds of millions of candidate short videos.

[0004] In order to achieve efficient personalized recommendation for different users, the following challenges need to be addressed. First, users have multiple interests, and the number of interests is also personalized. For example, a user may be interested in travel and pet-related short videos. These interest points play an implicit but very important role in determining which short videos the user decides to watch. Therefore, it is necessary to disentangle multiple user interest vectors from the user interaction sequence, where each interest vector represents a user's interest point; then, recommendations are generated according to the learned multiple interest points. However, most existing short video recommendation methods do not disentangle the user's multiple interests, and some methods, although they learn the user's multiple interest points, ultimately aggregate them into a single interest vector when serving. Next, the breadth and depth of each user interest may change over time. For example, a user who likes travel has watched a lot of short videos related to European scenery, but the interest may change to Asian scenery later. Further, although platforms encourage short video bloggers to add text labels when publishing videos, the problem of missing and inaccurate labels in short videos is still very serious. In addition, so far, unlike the e-commerce scenario where goods usually have a very complete classification system, short videos still lack a complete classification system.

[0005] Currently, some short video recommendation methods that utilize multi-modal information including visual, auditory and text features are proposed (Y. Wei, X. Wang, et. al. MMGCN: Multi-modal Graph Convolution Network for Personalized Recommendation of Micro-video. In Proceedings of the 27th ACM International Conference on Multimedia, 1437-1445). However, these methods are more suitable for the ranking stage rather than the recall stage. This is because the acquisition of multi-modal information usually requires a large amount of computing resources, and it is very expensive to process hundreds of millions of short videos in the recall stage.

[0006] So far, people have carried out some research to solve the problem of short video recommendation, but none of the methods can model user multi-interest while guaranteeing recommendation accuracy, diversity and algorithm running efficiency, so as to be applied to the short video recall stage. SUMMARY

[0007] The technical problem solved by the present application is to provide a short video recommendation method and device based on heterogeneous multi-interest modeling, which can mine multiple user interests implied in user historical interaction sequences, and realize diversified and accurate short video recommendation.

[0008] The technical content of the present application includes:

[0009] A short video recommendation method based on heterogeneous multi-interest modeling, the method comprising:

[0010] obtaining a user's historical interaction short video sequence;

[0011] constructing a set of mutually orthogonal short video category embedding vectors; wherein the number of short video category embedding vectors in the set of mutually orthogonal short video category embedding vectors is the number of short video categories;

[0012] combining the set of mutually orthogonal short video category embedding vectors, calculating the classification score of each short video in the historical interaction short video sequence in each short video category, to obtain the classification probability of each short video in each short video category;

[0013] based on the classification probability, sequentially weighting and converging the short video embedding in each short video category to obtain the user's soft interest representation in each short video category;

[0014] The short video is divided into a category to which the short video belongs with the highest probability, and then the historical interaction short video sequence is segmented according to the category to which the short video belongs, and a vector representation of a subsequence is calculated through a recurrent neural network to obtain a hard interest representation of the user in each short video category;

[0015] Based on the soft interest representation and the hard interest representation, a short video recommendation result of the user is obtained.

[0016] Further, in combination with the set of mutually orthogonal short video category embedding vectors, a classification score of each short video in the historical interaction short video sequence in each short video category is calculated to obtain a classification probability of each short video in each short video category, including:

[0017] For each short video in the historical interaction short video sequence, an initial vector v il is randomized, where i represents the serial number of the user, and l represents the serial number of the short video.

[0018] Based on a set of mutually orthogonal short video category embedding vectors {g1, g2, …, g k}, a classification score of each short video in each short video category is calculated

[0019] The short video classification score is amplified , where 1≤s≤k.

[0020] The amplified short video classification score is converted through a softmax function to obtain a classification probability of each short video in each short video category.

[0021] Further, the short video is divided into a category to which the short video belongs with the highest probability, and then the historical interaction short video sequence is segmented according to the category to which the short video belongs, and a vector representation of a subsequence is calculated through a recurrent neural network to obtain a hard interest representation of the user in each short video category, including:

[0022] The short video is divided into a category to which the short video belongs with the highest probability, and then the historical interaction short video sequence is segmented according to the category to which the short video belongs to obtain k non-intersecting subsequences;

[0023] Each non-empty subsequence is input into a GRU neural network, and the last hidden vector of the GRU neural network is taken as the vector representation of the non-empty subsequence;

[0024] The vector representation of the non-empty subsequence is taken as the hard interest representation of the user in each short video category.

[0025] Further, the short video recommendation result of the user is obtained based on the soft interest representation and the hard interest representation.

[0026] The interest vector of the user in each short video category is calculated wherein i represents the serial number of the user, represents the soft interest representation of the user in short video category j, represents the hard interest representation of the user in short video category j;

[0027] The short video recommendation result of the interest vector is obtained by calculating the similarity between the interest vector and the candidate short video embedding.

[0028] The short video recommendation results of the plurality of interest vectors are integrated to obtain the short video recommendation result of the user.

[0029] A short video recommendation device based on heterogeneous multi-interest modeling, the device comprising:

[0030] A short video acquisition module for acquiring a historical interactive short video sequence of a user;

[0031] An embedding vector construction module for constructing a set of mutually orthogonal short video category embedding vectors {g1, g2, …, g k} and obtaining the embedding {v i1 ,v i2 ,…,v il} of each short video in the user interaction sequence;

[0032] A classification probability calculation module for combining the short video category embedding vectors {g1, g2, …, g k} to calculate the short video embedding vector of each short video in the historical interactive short video sequence in each short video category, so as to obtain the classification probability of each short video in each short video category;

[0033] A soft interest calculation module for aggregating the short video embedding in each short video category based on the classification probability to obtain the soft interest representation of the user in each short video category;

[0034] A hard interest calculation module for dividing the historical interactive short video sequence according to the classification result of the short video and obtaining the hard interest representation of the user in each short video category by calculating the vector representation of the subsequence;

[0035] A short video recommendation module for obtaining the short video recommendation result of the user based on the soft interest representation and the hard interest representation. ​

[0036] Furthermore, the orthogonal loss function during the training of the embedding vector construction module

[0037] Furthermore, the uniform loss function during the training of the classification probability calculation module Where, σ w Let μ represent the standard deviation of the distribution w of the number of short videos. w This represents the average value of the distribution w of the number of short videos, where w = [w1, ..., w2]. j ,…,w k The estimated number of short videos classified into short video category j is [value missing]. Indicates user u i The classification probability of the l-th short video in short video category j, This represents the set of training users.

[0038] Furthermore, the uniqueness loss function of the hard interest calculation module is trained.

[0039] Furthermore, the loss of training the device Among them, cross-entropy loss This represents a short video randomly sampled from a sequence of future interactions of user i. Short video vectors, This represents the interest vector of user i for short video category j, where ∈ represents the temperature coefficient. λ represents the negative sample set formed by randomly sampling short videos that the user has never interacted with from the full set of candidate short videos. o λ represents the first weighting coefficient. f λ represents the second weighting coefficient. q This represents the third weighting coefficient.

[0040] An electronic device includes a memory and a processor, the memory storing a computer program and the processor being configured to run the computer program to perform any of the methods described above.

[0041] Compared with the prior art, the present invention has the following advantages and features:

[0042] (1) In the face of short videos without systematic and standardized category classification, this invention proposes to learn the mutually orthogonal implicit category information of short videos. Based on this, using the correspondence between user interests and short video categories as a bridge, this invention can identify multiple heterogeneous soft and hard interests of users by designing soft and hard allocation strategies for video categories.

[0043] (2) The application proposes a two-stage training strategy. In the first stage (pre-training stage), the soft interest of the user is identified, and the confidence of interest disintegration is improved. In the second stage (fine-tuning stage), the hard interest of the user is identified, the complete disintegration of the user interest is realized, and the evolution of a single interest over time is modeled.

[0044] (3) The application realizes effective and efficient short video recommendation without relying on multi-modal information of short videos, surpasses the recall rate of the existing multi-interest method, and guarantees the diversity of the recommendation results. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 The flowchart corresponding to the method of the application.

[0046] Figure 2 The influence of multi-interest on the diversity of recommendation. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical scheme and advantages of the application more clear, the application is further described in detail below in combination with the drawings and examples.

[0048] The short video recommendation method of the application is to input the user historical interactive short video sequence into the user interest extraction module, output the learned user multiple interest vectors, and generate a group of short video recommendation results according to each interest vector of the user, and select the K short videos with the highest relevance score as the final recommendation results. In the pre-training stage, the user historical interactive short video sequence is input into the soft interest extraction module to obtain multiple soft interest vectors of the user, and in the fine-tuning stage, the user historical interactive short video sequence is input into the soft interest extraction module and the hard interest extraction module respectively to obtain multiple soft interest vectors and multiple hard interest vectors of the user.

[0049] The short video recommendation method based on heterogeneous multi-interest modeling of the application comprises the following steps: Figure 1 The short video recommendation method based on heterogeneous multi-interest modeling of the application comprises the following steps:

[0050] Step 1: Obtain the user's historical interactive short video sequence.

[0051] Step 2: Construct a group of mutually orthogonal short video category embedding vectors.

[0052] The number of short video category embedding vectors in the group of mutually orthogonal short video category embedding vectors is the number of short video categories. Specifically, k parameterized, to-be-learned short video category embeddings {g1, g2, …, gk} are set, wherein, k} and and ||g *||2=1. To keep the heterogeneity between different short video categories and reduce the information overlap between multiple user interests, the short video category embeddings are pairwise orthogonal. Therefore, the orthogonal loss is constructed as follows:

[0053]

[0054] The embedding of a short video category can be regarded as a set of orthonormal bases. Thus, for each short video v The cosine similarity between the short video vector and the short video category vector can be calculated to obtain its coordinates in the set of orthonormal bases, denoted as where This coordinate vector measures the association between the short video v i and each short video category. The larger the coordinate value, the stronger the association.

[0055] Step 3: Combine the set of mutually orthogonal short video category embedding vectors to calculate the classification score of each short video in the historical interaction short video sequence in each short video category, to obtain the classification probability of each short video in each short video category.

[0056] The coordinate value of any short video v il in each short video category is calculated in turn, i.e. where Then, the coordinate value is amplified by dividing by a number less than 1, and then converted into the probability of category allocation by the softmax function. The probability calculation formula is as follows:

[0057]

[0058] Step 4: Based on the classification probability, the short video embedding in each short video category is aggregated in turn to obtain the soft interest representation of the user in each short video category.

[0059] The invention aggregates the embedding of the short video divided into each category in turn to form multiple interests of the user, as follows:

[0060]

[0061] To avoid the aggregation of short videos into a certain category, the method constrains the number of short videos divided into each category to be as equal as possible. The estimated value of the number of short videos divided into category j by the user's historical interaction is Thus, the distribution of the number of short videos of the user in different categories is w=[w1,…,w k ]. Then, the smaller the discrete coefficient of the distribution vector is, the better, and the following uniform loss is formed as follows:

[0062]

[0063] In the above formula, σ w is the standard deviation of the vector w, μ w is the average value of the vector w.

[0064] At this stage, the jth interest vector of the user u i is

[0065] Step 5: According to the number of short video categories, the historical interaction short video sequence of the user is divided, and the vector representation of the subsequence is calculated to obtain the hard interest representation of the user in each short video category.

[0066] The hard classification divides each historical interaction short video v i of the user u il only to the category with the maximum classification probability, that is, In order to further improve the confidence of the category division, the present application establishes a uniqueness loss based on cross entropy as follows, forcing the short video to be divided into only one category and far away from other categories.

[0067]

[0068] Based on the hard category of the short video, the historical interaction sequence s i of the user u i will be divided into k disjoint subsequences wherein each subsequence with a length exceeding δ h generates a corresponding user hard interest, so the number of user hard interests is dynamic and personalized. If the user has very extensive interests, he may have multiple hard interests; if the user's interests are concentrated, he may have only one or two hard interests. Further, in order to capture the evolution of each interest of the user, for each non-empty subsequence a sequence modeling method is used to learn the corresponding hard interest vector Here, in order to control the complexity of the method, GRU is used to model the subsequence to generate the user hard interest. Specifically, the short video subsequence corresponding to each interest is input into the GRU, and the last hidden vector is passed through a fully connected layer to obtain the user hard interest

[0069] Step 6: Based on the soft interest representation and the hard interest representation, a short video recommendation result of the user is obtained.

[0070] The present application combines the above-mentioned user soft interest and user hard interest to obtain the jth interest vector of the user u i is Then, a set of short video recommendation results can be generated based on each user's interest vector, and the K short videos with the highest relevance scores can be selected as the final recommendation results.

[0071] Specifically, for a user's interest vector, the cosine similarity between the interest vector and the candidate short video embeddings is calculated sequentially, and the K short videos with the highest similarity are selected in descending order. Then, the similarity of the multiple recommendation results during recall is sorted from high to low, and the K elements with the highest similarity are taken as the final recommendation results.

[0072] Based on steps 2-6 above, this invention constructs a short video recommendation method based on heterogeneous multi-interest modeling. The final training loss is as follows, where λ o , λ f , λ q It is the coefficient of the regularization term.

[0073]

[0074]

[0075] Among them, the cross-entropy loss is based on the data from user u. i Future Interaction Sequence A short video is randomly sampled as a positive sample. (where f) i The negative sample set is formed by randomly sampling short videos that the user will interact with in the future (the number of short videos the user will interact with in the future) and short videos that the user has never interacted with, from the full set of candidate short videos. structure.

[0076] To avoid the additional overhead of negative sampling, a common practice is to use intra-batch softmax. However, this method limits the diversity of negative samples. Since the goal of recall is to eliminate a large number of obviously irrelevant short videos, random sampling is needed from all candidate short video pools to obtain a richer pool of negative sample short videos. To balance sampling overhead and negative sample diversity, this invention randomly samples one negative sample for each training sample, and shares the negative sample within each batch of training samples. Let the batch size be B, then there are B negative samples. Therefore, the multi-class probability is predicted by calculating the value of the sampling softmax function in B+1 samples, where one is a positive sample and the other B are negative samples.

[0077] In the embodiment, the Python programming language is used, which can run on mainstream computer system platforms. The operating system used in the embodiment is Ubuntu 16.04, the CPU is required to be Intel Core i7, the memory requirement is 8G or more, the hard disk space requirement is 16G or more, the GPU is NVIDIA GeForce 3090, and the video memory is 24G.

[0078] The present application is implemented based on the PyTorch 1.7.0 framework, and the Adam optimization algorithm (Kingma DP, Ba J. 2014. Adam: A method for stochastic optimization [J]. arXiv preprint arXiv:1412.6980.) is used as the gradient descent algorithm to solve all the parameters to be learned in the neural network.

[0079] The short video and interest embedding matrix in the method are implemented using the torch.nn.Embedding class, the linear transformation is implemented using the torch.nn.Linear class, the activation function is implemented using the torch.nn.ReLU class, the GRU recurrent neural network is implemented using the torch.nn.GRU class, and the cross-entropy loss is implemented using the torch.nn.CrossEntropyLoss class. When performing forward calculation, first, the short video and interest embedding vectors are normalized in the two-norm, so that the correlation between the short video and the interest uses the cosine similarity instead of the inner product. In the pre-training stage, the method only needs to calculate the soft interest, at this time, the method will calculate the cosine similarity between the user historical interaction short video and each interest embedding and divide by a temperature coefficient less than 1, then calculate the probability of the short video assigned to each interest through the torch.softmax() function, and then based on the probability, the short video assigned to the same interest is weighted and summed to obtain each soft interest of the user. In the fine-tuning stage, in addition to calculating the soft interest, the method also needs to calculate the user hard interest. When calculating the hard interest, the torch.argmax() function is used to obtain the interest number with the maximum probability of each short video and count the number of short videos belonging to each interest, then the index and torch.split() function are used to obtain the short video subsequence related to each interest, then the pad_sequence() function and the pack_padded_sequence() function are used to pad the sequence, and input into the GRU layer to obtain the corresponding user hard interest. The predicted score of the candidate short video by the user is calculated as the inner product of the user embedding and the short video embedding. The predicted scores of the positive and negative samples are input into the cross-entropy loss to obtain the main loss of the recommendation task, and the sum of the main loss, the orthogonal loss, the uniform loss and the uniqueness loss is the final training loss.

[0080] The configuration parameters used in the embodiment are shown in Table 1.

[0081] Table 1 Configuration parameter settings

[0082] Name Value Description o ]]> ​ 10 Coefficient of loss of orthogonality f ]]> ​ 1 Coefficient of loss of uniformity q ]]> ​ 1 Coefficient of loss of uniqueness ∈ 0.1 Temperature coefficient

[0083] The main parameters to be learned are shown in Table 2.

[0084] Table 2 Main parameters to be learned configuration

[0085] Name Parameter dimensions Short video embedding matrix 157691,64 Interest embedding matrix 8,64 GRU parameter matrix 64,64 Fully connected layer 64,64

[0086] The data in the embodiment of the application is derived from MX-TakaTak, one of the largest short video platforms in India. The interaction logs on the MX-TakaTak platform from September 18, 2021 to September 28, 2021 are collected, and the interaction data of 50,000 users randomly sampled are used to construct a data set, which includes 157,691 short videos, 33,863,980 interactions, and an interaction matrix density of 0.45%. The interaction behavior that meets the following conditions is regarded as a positive interaction: the interaction behavior is a satisfied interaction (such as likes, collections), the number of repeated plays is >1.8 times, or the play duration is >15 seconds.

[0087] In order to prevent the data penetration problem in the training and evaluation stage, the training set, the validation set and the test set are strictly constructed in time sequence in this embodiment. Specifically, the last day of data is used to construct the test set, the second last day of data is used to construct the validation set, and then the other data is used to construct the training set.

[0088] In order to evaluate the performance of the method, this embodiment uses two indicators, recall@K (Recall@K) and hit rate@K (HitRate@K), to evaluate the recommendation results. The value of K is set to 50, 100 and 200.

[0089] This embodiment uses uniform distribution All parameters in the initialization method are optimized using the Adam optimizer. The learning rate is searched from [0.1, 0.01, 0.001, 0.0001], and finally set to 0.001, λ q , λ o , λ f is searched in [0.5, 1, 5, 10, 20], and finally set to λ q =1, λ o =10, λ f= 1, and the temperature coefficient τ is set to 0.1. This embodiment searches the number of interests in [1, 2, 4, 8, 16] and selects the number that performs best on the recall@200 indicator of the validation set, and finally determines the number of interests to be 8. This embodiment sets the embedding dimension of all experiments to be 64, and the batch size to be 1024. Training is stopped when the indicator recall@200 does not improve on the validation set for 5 consecutive rounds, and the test set is tested.

[0090] The final performance of this embodiment on the test set is: the recall@50 reaches 14.24%, the recall@100 reaches 13.73%, the recall@200 reaches 20.21%, the hit rate@50 reaches 57.44%, the hit rate@100 reaches 67.36%, and the hit rate@200 reaches 75.98%, which exceeds other existing multi-interest methods, indicating that the recommended results of the method have high accuracy.

[0091] The SPPMI value between short videos interacted by the same user (Levy O, Goldberg Y. Neural word embedding as implicit matrix factorization[J]. 2014: 2177-2185.) can measure the similarity between short videos. This embodiment uses the SPPMI value between recommended short videos when the number of interests is set to 1 and the number of interests is set to 4 to measure the diversity of the recommended results, and the evaluation results are shown in Table 2. Figure 2 As can be seen, the SPPMI value between short videos recalled by the same interest vector is relatively high in the recommended results of the multi-interest method, while the SPPMI value between short videos recalled by different interest vectors is relatively low. At the same time, it can be seen that the SPPMI value between recommended short videos of the single-interest method is generally high, indicating that the recommended short videos have high similarity and the recommended results are relatively single. This shows that the multi-interest method can improve the diversity of the recommended results.

[0092] The above embodiments are provided only for the purpose of describing the present application, and are not intended to limit the scope of the present application. The scope of the present application is defined by the appended claims. Various equivalent substitutions and modifications made without departing from the spirit and principles of the present application shall be encompassed within the scope of the present application.

Claims

1. A short video recommendation method based on heterogeneous multi-interest modeling, characterized in that, The method comprises: acquiring a historical short video sequence of user interaction; constructing a set of mutually orthogonal short video category embedding vectors; wherein the number of short video category embedding vectors in the set of mutually orthogonal short video category embedding vectors is the number of short video categories; combining the set of mutually orthogonal short video category embedding vectors, calculating the classification score of each short video in the historical short video sequence of user interaction on each short video category to obtain the classification probability of each short video on each short video category; based on the classification probability, sequentially weighting and converging short video embedding on each short video category to obtain the soft interest representation of the user on each short video category; dividing short videos into the category with the highest probability, cutting the historical short video sequence of user interaction according to the category of the short video, and calculating the vector representation of the subsequence through a recurrent neural network to obtain the hard interest representation of the user on each short video category; based on the soft interest representation and the hard interest representation, acquiring the short video recommendation result of the user; wherein the acquisition of the short video recommendation result of the user based on the soft interest representation and the hard interest representation comprises: calculating an interest vector of the user on each short video category wherein i represents the serial number of the user, denotes a soft interest representation of the user on short video category j, denotes a hard interest representation of the user on short video category j; by calculating the interest vector a similarity between the candidate short video embedding and the interest vector a short video recommendation result; combining the plurality of interest vectors the short video recommendation result of the user.

2. The method of claim 1, wherein, the combination of the set of mutually orthogonal short video category embedding vectors, the calculation of the classification score of each short video in the historical short video sequence of user interaction on each short video category to obtain the classification probability of each short video on each short video category comprises: for each short video in the sequence of historical interactive short videos, randomize an initial vector v il ; where i denotes the sequence number of the user, and I denotes the sequence number of the short video. Based on a set of mutually orthogonal short video category embedding vectors {g1, g2,..., g k}, calculate the classification score of each short video on each short video category Amplifying the above classification score where 1≤s≤k; Converting the amplified short video embedding vector by a softmax function Obtain the classification probability of each short video on each short video category.

3. The method of claim 1, wherein, the cutting of the historical short video sequence of user interaction according to the number of short video categories and the calculation of the vector representation of the subsequence to obtain the hard interest representation of the user on each short video category comprises: dividing short videos into the category with the highest probability, and then cutting the historical short video sequence of user interaction according to the category of the short video to obtain k non-intersecting subsequences; inputting each non-empty subsequence into a GRU neural network and taking the last hidden vector of the GRU neural network as the vector representation of the non-empty subsequence; taking the vector representation of the non-empty subsequence as the hard interest representation of the user on each short video category.

4. A short video recommendation device based on heterogeneous multi-interest modeling, characterized in that, The device comprises: a short video acquisition module for acquiring a historical short video sequence of user interaction; The embedding vector construction module is configured to construct a set of mutually orthogonal short video category embedding vectors {g1, g2, …, g k} and obtain the embedding {v i1 ,v i2 ,…,v il} of each short video in the user interaction sequence. a classification probability calculation module, configured to combine the short video category embedding vectors {g1, g2, …, g k} to calculate short video embedding vectors of each short video in the historical interaction short video sequence on each short video category, so as to obtain classification probabilities of each short video on each short video category; a soft interest calculation module for sequentially converging short video embedding on each short video category based on the classification probability to obtain the soft interest representation of the user on each short video category; a hard interest calculation module for cutting the historical short video sequence of user interaction according to the classification result of the short video and calculating the vector representation of the subsequence to obtain the hard interest representation of the user on each short video category; a short video recommendation module for acquiring the short video recommendation result of the user based on the soft interest representation and the hard interest representation; wherein the acquisition of the short video recommendation result of the user based on the soft interest representation and the hard interest representation comprises: calculating an interest vector of the user on each short video category wherein i represents the serial number of the user, represents a soft interest representation of the user on short video category j, represents a hard interest representation of the user on short video category j; by calculating the interest vector a similarity between the candidate short video embedding and the interest vector a short video recommendation result; combining the plurality of interest vectors the short video recommendation result of the user.

5. The apparatus of claim 4, wherein, loss when training the embedding vector construction module 6. The apparatus of claim 5, wherein, Loss function when training the classification probability calculation module wherein σ w denotes the standard deviation of the short video quantity distribution w, μ w denotes the mean value of the short video quantity distribution w, the short video quantity distribution w = [w1, …, w j , …, w k ], the estimated value of the number of short videos divided into the short video category j is denotes the classification probability of the lth short video of the user u i in the short video category j, u denotes the training user set.

7. The apparatus of claim 6, wherein, loss to train the hard interest computing module 8. The apparatus of claim 7, wherein, Loss of training the device wherein, cross-entropy loss represents a short video vector of a short video randomly sampled from a future interaction short video sequence of a user i, represents an interest vector of the user i to a short video category j, represents a temperature coefficient, represents a negative sample set formed by randomly sampling short videos that the user has never interacted from a full candidate short video set, o represents a first weight coefficient, f represents a second weight coefficient, q represents a third weight coefficient.

9. An electronic device, comprising: a device comprising a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the method according to any one of claims 1-3.

Images