A personalized video recommendation method and system based on automatic feature combination

The embedded representation and combination of user and video features is performed through the deep neural network module, combining popular and manually selected videos, and solving the problem of failing to effectively combine user and video features in the prior art, realizing personalized video recommendations Accuracy and generalization ability.

CN114168790BActive Publication Date: 2024-11-26CENT SOUTH UNIV +1
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202111138288.5
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2021-09-27
Publication Date
2024-11-26
Estimated Expiration
2041-09-27

AI Technical Summary

Technical Problem

The existing video recommendation methods fail to effectively consider the cross-combination of user and video features, making it difficult for users to obtain a personalized video recommendation list.

Method used

A similar recall submodule and precise sorting module based on deep neural network are used to combine the embedded representation vectors of users and videos to generate a personalized video recommendation list, and combine and sort together popular and manually selected videos.

Benefits of technology

It realizes the rapid generation of personalized video recommendation candidate sets, explores user interests and hobbies, provides users with personalized video recommendations, and improves the accuracy and generalization ability of recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114168790B_ABST
    Figure CN114168790B_ABST
Patent Text Reader

Abstract

The present invention discloses a personalized video recommendation method and system based on automatic feature combination, the method comprising: according to the user unique identifier, through the similarity recall submodule based on deep neural network, obtaining the embedded representation vector of the user and the video; inputting the embedded representation vector of the user and the video into the vector similarity retrieval tool, searching and obtaining multiple initial recommended videos of interest to the user; combining the multiple initial recommended videos with multiple popular videos obtained through the popular recall submodule and multiple manually selected videos obtained through the manual recall submodule to form a video recommendation candidate list; through the user unique identifier and the corresponding video recommendation candidate list, inputting the accurate sorting module to sort and obtain an ordered video recommendation list. The present invention can automatically combine user features and video features to provide personalized video recommendations for users.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of video recommendation, and in particular to a personalized video recommendation method and system based on automatic feature combination. Background Art

[0002] With the rapid development of information technology and the Internet, especially the development of short video technology, users on Internet video platforms can not only be video consumers, but also video producers. With the improvement of mobile device performance and the widespread application of fourth-generation and fifth-generation mobile communication technologies, video production has broken through the limitations of time and space. On short video platforms, a large number of videos are produced every second. The explosive growth in the amount of videos and users has also brought about the problem of video overload. Video overload refers to the fact that it is very difficult for video users to find videos of interest from a large number of videos because the amount of videos is too large and complicated.

[0003] Video recommendation system is an important tool to solve this contradiction and has great research value. Due to the importance of video recommendation system, in the early stage of its development, the common recommendation method was to simply sort by the sales volume of items, the number of clicks on topics, or the number of news readings. However, this method is too one-sided and not personalized enough. Subsequently, personalized recommendation algorithms based on matrix decomposition appeared, but this low-dimensional and linear method lacks sufficient generalization ability, that is, the effect on untrained data samples is not good enough.

[0004] Although the currently more common deep learning-based algorithms meet the requirements of personalized recommendations and have good generalization capabilities, they fail to take into account the cross-combination of user and video features. However, in fact, feature combination is of practical significance. For example, a user is from Changsha, is a male, and likes to eat betel nuts. If the two features "Changsha" and "male" are not combined and trained together, the model may conclude that "Changsha people" like to eat betel nuts, or that "males" like to eat betel nuts, which is inappropriate. At present, some practitioners also use manual methods to combine features and conduct training, but this method is not general and flexible enough. Summary of the invention

[0005] The present invention provides a personalized video recommendation method and system based on automatic feature combination, which aims to solve the technical problem that the current video recommendation method fails to take into account the cross-combination of user and video features, resulting in video users having difficulty in accurately obtaining a personalized video recommendation list.

[0006] In order to solve the above technical problems, the technical solution proposed by the present invention is:

[0007] A personalized video recommendation method based on automatic feature combination includes the following steps:

[0008] According to the user's unique identifier, the embedding representation vector of the user and video is obtained through the similarity recall submodule based on the deep neural network;

[0009] Input the embedding representation vectors of the user and the video into a vector similarity retrieval tool to search for multiple initial recommended videos that the user is interested in;

[0010] Combining the multiple initial recommended videos, the multiple popular videos obtained through the popular recall submodule, and the multiple manually selected videos obtained through the manual recall submodule to form a video recommendation candidate list;

[0011] The user's unique identifier and the corresponding video recommendation candidate list are input into the precise sorting module for sorting to obtain an ordered video recommendation list.

[0012] Preferably, the similarity recall submodule based on the deep neural network is trained by the following steps:

[0013] Collect user-side information, video-side information, and historical interaction information between users and videos; calculate the user's preference weight for the video based on the historical interaction information between users and videos; construct a training database where each sample is: user-video-preference weight;

[0014] The user's unique identifier and the corresponding user-side information, as well as the video's unique identifier and the corresponding video-side information are taken as input, the user's preference weight for the video is used as the supervision label, the video that the user has not interacted with is used as the negative sampling data, and the user and video embedding representation vectors are used as output. Supervised learning is performed to train a similarity recall submodule based on a deep neural network.

[0015] Preferably, the similar recall submodule, the popular recall submodule, the manual recall submodule and the precise sorting module are all trained offline and updated periodically or irregularly.

[0016] Preferably, the calculation model of the user's preference weight for the video is:

[0017] weight=(click_time×click_weight+like_time×like_weight+favorite_time×favorite_weight)×watch_time_ratio

[0018] Among them, weight is the user's preference weight for the video; click_time, like_time, and favorite_time respectively represent the number of times the user clicks on the video, whether the user likes the video, and whether the user collects the video; click_weight, like_weight, and favorite_weight respectively represent the weights of the corresponding behaviors; watch_time_ratio represents the proportion of the user's viewing time of the video to the total video length.

[0019] Preferably, the precise sorting module is trained by the following steps:

[0020] The user's unique identifier, user-side information, and the video unique identifier and video-side information in the video recommendation candidate list are taken as input and respectively input into the parallel feature automatic combination layer and deep learning layer; the output vectors of the feature automatic combination layer and the deep learning layer are concatenated and input into a neural network hidden layer, and the sorted video recommendation list corresponding to the user's unique identifier is output; the absolute value loss function is used to measure the error between the user's actual preference weight for the video and the training result, and the gradient descent algorithm is used to train the model to converge to a stable state with a small error as an accurate sorting module.

[0021] Preferably, the feature automatic combination layer includes the following multiple hidden layers, and the input and output of each hidden layer are as follows:

[0022] x l+1 =x0x l w l+1 +b l+1 +x l

[0023] Among them, x l represents the output of the lth hidden layer, w l represents the neuron weight of the lth hidden layer, b l represents the bias of the lth hidden layer.

[0024] Preferably, the feature automatic combination layer uses the following function as the neural network activation function:

[0025]

[0026] y i =a i ×(1-p i )×y i +p i ×y i ;

[0027] where y i is the output of neuron i, E[y i] and Var[y i ] is the mean and variance obtained during training, ∈ is a small positive number; a i is a very small hyperparameter, usually set to 0.01, p i is a trainable variable used to activate the propagation of neurons.

[0028] The present invention also provides a personalized video recommendation system based on automatic feature combination, comprising:

[0029] A side information module is used to collect user side information and video side information, where the user side information includes user portraits and the video side information includes video attribute information;

[0030] The interaction record module is used to record the historical interaction information between users and videos, including user click behavior, like behavior, collection behavior and viewing time;

[0031] Rapid Recall Module, and

[0032] The precise sorting module is used to sort the user's unique identifier and the corresponding video recommendation candidate list to obtain an ordered video recommendation list;

[0033] Among them, the rapid recall module includes:

[0034] Similarity recall submodule, used to obtain the embedding representation vector of users and videos based on the user's unique identifier;

[0035] Vector similarity retrieval tool, used to search for multiple initial recommended videos that the user is interested in based on the embedding representation vectors of users and videos;

[0036] Popular recall submodule, used to obtain multiple popular videos;

[0037] A manual recall submodule, used to obtain multiple manually selected videos;

[0038] And, a video combination module is used to combine multiple initial recommended videos, multiple popular videos, and multiple manually selected videos to form a video recommendation candidate list.

[0039] The present invention also provides a computer system, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above methods when executing the computer program.

[0040] The present invention also provides a computer storage medium on which a computer program is stored, and when the program is executed by a processor, the steps in any of the above methods are implemented.

[0041] The present invention has the following beneficial effects:

[0042] The personalized video recommendation method and system based on automatic feature combination of the present invention can quickly generate a recommendation candidate set and explore user interests and hobbies through user and video interaction records; and combine and sort the initial recommended videos with popular videos and manually selected videos to obtain a final video recommendation list, which can provide personalized video recommendations for users.

[0043] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0045] Figure 1 A flowchart of a personalized video recommendation method based on automatic feature combination according to an embodiment of the present invention;

[0046] Figure 2 This is an architecture diagram of a personalized video recommendation system based on automatic feature combination according to an embodiment of the present invention;

[0047] Figure 3 A flowchart of calculating a user's video preference weight according to an embodiment of the present invention;

[0048] Figure 4 is a schematic diagram of a similarity recall submodule according to an embodiment of the present invention;

[0049] Figure 5 Schematic diagram of a precise sorting module according to an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0051] See also Figure 1 The personalized video recommendation method based on automatic feature combination of the present invention, the prediction stage, includes the following steps:

[0052] According to the user's unique identifier, the embedding representation vector of the user and video is obtained through the similarity recall submodule based on the deep neural network;

[0053] Input the embedding representation vectors of the user and the video into a vector similarity retrieval tool to search for multiple initial recommended videos that the user is interested in;

[0054] Combining the multiple initial recommended videos, the multiple popular videos obtained through the popular recall submodule, and the multiple manually selected videos obtained through the manual recall submodule to form a video recommendation candidate list;

[0055] The user's unique identifier and the corresponding video recommendation candidate list are input into the precise sorting module for sorting to obtain an ordered video recommendation list.

[0056] In the above steps, see Figure 5 ,The similar recall submodule, popular recall submodule, manual recall submodule and precise sorting module are all trained offline and ,updated (released) periodically or irregularly.

[0057] The similarity recall submodule based on the deep neural network of this embodiment is preferably trained by the following steps:

[0058] Collect user side information, video side information, and historical interaction information between users and videos. When collecting data, record the user side information and video side information in the database, collect the interaction information of user videos through the embedding technology, and store it in the database. In this embodiment, the Kafka message queue and Hive big data storage engine are used.

[0059] The user's preference weight for the video is calculated based on the historical interaction information between the user and the video. A training database is constructed for each sample: user-video-preference weight.

[0060] The user's unique identifier and the corresponding user-side information, as well as the video's unique identifier and the corresponding video-side information are used as input, the user's preference weight for the video is used as the supervision label, the video that the user has not interacted with is used as the negative sampling data, and the user and video embedding representation vectors are used as outputs. Supervised learning is performed to train a similar recall submodule based on a deep neural network. During implementation, the similar recall submodule in the fast recall module is regularly trained and released through offline training.

[0061] The precise sorting module of this embodiment is preferably trained by the following steps:

[0062] The user's unique identifier, user-side information, and the video's unique identifier and video-side information in the video recommendation candidate list are taken as input and respectively input into the parallel feature automatic combination layer and deep learning layer; after the output vectors of the feature automatic combination layer and the deep learning layer are spliced, the spliced ​​vector is input into a neural network hidden layer with an output dimension of 1, and a sorted video recommendation list corresponding to the user's unique identifier is output; the absolute value loss function is used to measure the error between the user's actual preference weight for the video and the training result, and the gradient descent algorithm is used to train the model to converge to a stable state with a small error as an accurate sorting module.

[0063] Through the above steps, the personalized video recommendation method and system based on automatic feature combination of the present invention can quickly generate a recommendation candidate set through the user and video interaction records, and explore the user's interests and hobbies; and combine and sort the initial recommended videos with popular videos and manually selected videos to obtain a final video recommendation list, which can provide personalized video recommendations for users.

[0064] Corresponding to the above method, an embodiment of the present invention further provides a personalized video recommendation system based on automatic feature combination, including:

[0065] The side information module is used to collect user side information and video side information. User side information includes user portraits, and video side information includes video attribute information. Video side information includes the category, tag, language, and length of the video filled in by the creator when uploading the video. User side information includes the user's gender, hobbies, region, and other information filled in when the user registers. After collecting these side information, they can be used as training data for the fast recall module and the precise sorting module.

[0066] The interaction record module is used to record the historical interaction information between users and videos, including user click behavior, like behavior, favorite behavior and viewing time. And calculate the preference weight of the interactive video and store it in the database. The user interaction behavior is confirmed by three identifiers, namely user identifier (uId), video identifier (vId) and behavior identifier (like_time, favorite_time, watch_time_ratio). It is saved by serialization using a message queue, where the queue adopts a first-in-first-out consumption logic. The data is written to the big data storage engine Hive, and the user's preference weight for the video is calculated by writing a Hive structured query statement.

[0067] After using the interaction recording module, we can obtain a large amount of interaction data (user, video, preference weight). We use the user unique identifier (uId), user-side information, video unique identifier (vId), and video-side information as input data, and the preference weight as a supervision label to carry out subsequent training and prediction of the fast recall module and precise sorting module.

[0068] The fast recall module is mainly used to generate a candidate video list. It includes:

[0069] The similarity recall submodule is used to obtain the embedded representation vectors of users and videos based on the user's unique identifier; the user's unique identifier (uId), user-side information, video unique identifier (vId), and video-side information are used as input data, and the preference weights are used as labels. Supervised learning is performed through a deep neural network to obtain the embedded representation vectors of users and videos.

[0070] The vector similarity retrieval tool is used to search for multiple initial recommended videos that the user is interested in based on the embedded representation vectors of the user and the video. A high-performance vector retrieval tool is used to obtain the N1 (for example, N1 is 30) video vIds that the user is most likely to be interested in based on the user's embedded representation vector.

[0071] The popular recall submodule is used to obtain multiple popular videos; select the most popular N2 (for example, N2 is 10) video vIds.

[0072] The manual recall submodule is used to obtain multiple manually selected videos; manually select N3 (for example, N3 is 10) video vIds.

[0073] And, a video combination module is used to combine multiple initial recommended videos, multiple popular videos, and multiple manually selected videos to form a video recommendation candidate list. The video candidate lists obtained by the above three methods are merged to obtain the final candidate list candidate_vIds.

[0074] The personalized video recommendation system of this embodiment also includes: a precise sorting module, which is used to sort the user's unique identifier and the corresponding video recommendation candidate list to obtain an ordered video recommendation list. The main purpose of this module is to further accurately sort the candidate_vIds obtained by the fast recall module. In the precise sorting module, a deep neural network is used to construct a feature automatic combination layer and a deep learning layer respectively, and the strong robustness of the deep neural network is used to automatically obtain a suitable feature combination method. And the deep neural network layer is used to improve the generalization ability of the model.

[0075] Figure 3 This is a flowchart for calculating the user's video preference weight. It mainly includes four steps:

[0076] Step 1: Tracking. On the client side, tracking technology can be used to generate corresponding record data when users trigger certain specific behaviors. Object serialization technology is used to convert each record data into a string for transmission.

[0077] The record specification is: {"uId":uId, "vId":vId, "behave_type":behave_type, "num":num};

[0078] uId is the unique user ID, and vId is the unique video ID. behave_type has three values: like, favorite, and end. num represents the value of the corresponding behavior. For example, when behave_type is like, num can be 1 and 0, indicating that the user has liked and unliked the video, respectively. Similarly, when behave_type is favorite, num's 0 and 1 represent that the user has favorited and unfavorited the video. When behave_type is end, num's value range is between 0 and 1, indicating the proportion of the user's viewing time to the total video time.

[0079] Step 2: Store the records in the database. After obtaining the standardized record data, you can receive the record data through the Kafka message queue, and after the number of data records accumulates to a certain amount (for example, 1000), store the record data in the Hive distributed database. This can avoid writing to the Hive database every time a record data is generated, which can effectively improve database performance.

[0080] Design a table in the Hive database (for example, the table name is records) to store record data. The table includes at least five fields: uId, vId, like_time, favorite_time, and watch_time_ratio.

[0081] Step 3: Customize query statements to calculate preference weights. After obtaining the recorded data, you need to customize query statements to calculate the user's preference weights for the video and retrieve the data for subsequent model training.

[0082] Preference weight calculation formula:

[0083] weight=(click_time×click_weight+like_time×like_weight+favorite_time×favorite_weight)×watch_time_ratio

[0084] Among them, weight is the user's preference weight for the video; click_time, like_time, and favorite_time respectively represent the number of times the user clicks on the video, whether the user likes the video, and whether the user collects the video; click_weight, like_weight, and favorite_weight respectively represent the weights of the corresponding behaviors; watch_time_ratio represents the proportion of the user's viewing time of the video to the total video length.

[0085] For example, when clikc_weight is 1, like_weight is 2, and favorite_weight is 3, you can customize the structured query statement as follows:

[0086] SELECT uId,vId,(click_time*1+like_time*2+favorite_time*3)*watch_time_ratio ASweight FROM records;

[0087] Step 4: Get trainable data. After customizing the structured query statement, you can use the statement:

[0088] hive-e "SELECT uId,vId,(click_time*1+like_time*2+favorite_time*3)*watch_time_ratio as weight FROM records;">>data.csv;

[0089] Save it as a csv file for training.

[0090] Figure 4 Schematic diagram of the similarity recall submodule, which mainly includes three layers: input layer, representation layer and matching layer.

[0091] Input layer: The main function of the input layer is to convert the user or video input (such as uId, vId, and corresponding side information) into a vector in Euclidean space. Methods that can be used in this step include one-hot encoding, etc.

[0092] Representation layer: The representation layer is the key layer for learning user and video representation vectors. It mainly utilizes the strong representation ability of deep neural networks. Through several hidden layers and using the gradient descent algorithm, the weight value of each layer is continuously optimized. The final output of the representation layer can be used as the representation vector of the user and video.

[0093] Matching layer: The matching layer is the key layer that determines the representation effect. In this layer, by using the vector inner product method, we can get a training sample and then get the user's model score for the video, which can also be understood as the model prediction preference weight. The calculation method is as follows:

[0094]

[0095] Among them, a and b represent the representation vectors of the user and video in the representation layer in a training sample respectively. By comparing this score with the actual preference weight of the user for the video calculated in the interaction record module, supervised learning can be established, and the model can be continuously optimized through back propagation.

[0096] After obtaining the representation vectors of users and videos, high-performance vector similarity retrieval tools (such as faiss and Elasticsearch) can be used to quickly obtain the N1 videos that the specified user is most interested in.

[0097] Figure 5 The figure is a schematic diagram of the precise sorting module. This module is mainly composed of the feature automatic combination layer and the deep learning layer. The deep learning layer is similar to the representation layer of the similarity recall submodule.

[0098] The characteristic of the feature automatic combination layer is that it can explicitly learn feature intersection. The input and output of each layer are:

[0099] x l+1 =x0x l w l+1 +b l+1 +x l

[0100] Among them, x l represents the output of the lth hidden layer, w l represents the neuron weight of the lth hidden layer, b l represents the bias of the lth hidden layer.

[0101] The feature automatic combination layer forms synthetic features by multiplying features, which helps to represent nonlinear relationships. The special structure of the feature automatic combination layer makes the order of cross features increase as the depth of the network increases. For the L-layer feature automatic combination layer, the order of cross features can reach up to L+1. The time complexity and space complexity of the feature automatic combination layer are both linear functions of the input dimension. Therefore, the cost of the feature automatic combination layer is almost negligible compared to the deep learning layer on the right.

[0102] In the actual training and prediction process, the embedded representations of the user and video uId, vId, and side information are concatenated to obtain a new vector. The calculation method is as follows:

[0103] o * =o 1 +o 2

[0104] Among them 1 Represents the final output vector of the feature automatic combination layer, o 2 Represents the final output of the deep learning layer, + here represents the splicing operation, that is, o 2 The elements of are added to o in sequence 1 rear.

[0105] Get o *After that, it is sent as input to a neural network hidden layer with an output dimension of 1. In this way, the outputs of the feature automatic combination layer and the deep learning layer are combined, and the absolute value loss function is used to measure the difference between the user's actual preference weight for the video and the training result. The loss function is calculated as follows:

[0106] L(Y,f(x))=|Yf(x)|

[0107] Where Y represents the user's actual preference weight for the video, and f(x) represents the output value of the model, which means the user's preference weight for the video predicted by the model. In fact, there is an error between the model output value and the user's actual preference weight. Therefore, an algorithm is needed to train the model to a state with a smaller error.

[0108] The present invention uses a stochastic gradient descent algorithm. Its calculation method is:

[0109]

[0110] where f i (x) represents the output of randomly sampled sample i, Indicates partial derivative, η represents learning rate, which can adjust the amplitude of gradient descent and the rate of model convergence. In fact, in practical applications, η can be initialized to a relatively large value, and then gradually reduced as the number of training iterations increases. The advantage of this is that it can ensure the convergence rate while maximizing the convergence effect and model accuracy. If a large η value is maintained, the model may not converge.

[0111] In actual training, in order to improve the robustness of the model, an activation function needs to be used, that is, the output of each neuron needs to be mapped by an activation function before being output. The activation function used in this embodiment is as follows:

[0112]

[0113] y i =a i ×(1-p i )×y i +p i ×y i

[0114] where y i is the output of neuron i, E[y i ] and Var[y i ] is the mean and variance obtained during training, and ∈ is a small positive number in order to smooth the results. i is a very small hyperparameter, usually set to 0.01, p iis a trainable variable used to activate the propagation of neurons.

[0115] The present invention also provides a computer system, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of any of the above embodiments when executing the computer program.

[0116] The present invention also provides a computer storage medium on which a computer program is stored. When the program is executed by a processor, the steps in any of the above embodiments are implemented.

[0117] In summary, the present invention aims at the deficiencies of the existing recommendation system, discloses a personalized video recommendation method and system based on automatic feature combination, automatically records user video interaction records, quickly generates initial recommended videos, combines and sorts the initial recommended videos with popular videos and manually selected videos to obtain a final video recommendation list. The present invention can automatically combine user features and video features to provide personalized video recommendations for users.

[0118] The above are only preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A personalized video recommendation method based on automatic feature combination, It is characterized in that The following steps are involved: According to the user's unique identifier, the embedding representation vector of the user and video is obtained through the similarity recall submodule based on the deep neural network; Inputting the embedding representation vectors of the user and the video into a vector similarity retrieval tool to search for a plurality of initial recommended videos that the user is interested in; Combining the multiple initial recommended videos, the multiple popular videos obtained by the popular recall submodule, and the multiple manually selected videos obtained by the manual recall submodule to form a video recommendation candidate list; The user's unique identifier and the corresponding video recommendation candidate list are input into the precise sorting module for sorting to obtain an ordered video recommendation list; The similarity recall submodule based on deep neural network is trained by the following steps: Collecting user-side information, video-side information, and historical interaction information between the user and the video; calculating the user's preference weight for the video based on the historical interaction information between the user and the video; Construct each sample as: a training database of user-video-preference weights; The user's unique identifier and the corresponding user-side information, as well as the video's unique identifier and the corresponding video-side information are used as input, the user's preference weight for the video is used as the supervision label, the video that the user has not interacted with is used as the negative sampling data, and the user and video embedding representation vectors are used as outputs. Supervised learning is performed to train a similarity recall submodule based on a deep neural network. The precise sorting module is trained by the following steps: The user's unique identifier, user-side information, and the video's unique identifier and video-side information in the video recommendation candidate list are used as inputs, and are respectively input into a mutually parallel feature automatic combination layer and a deep learning layer; the output vectors of the feature automatic combination layer and the deep learning layer are spliced ​​and input into a neural network hidden layer, and a sorted video recommendation list corresponding to the user's unique identifier is output; the absolute value loss function is used to measure the error between the user's actual preference weight for the video and the training result, and the gradient descent algorithm is used to train the model to converge to a stable state with a small error as an accurate sorting module; The feature automatic combination layer includes the following multiple hidden layers, and the input and output of each hidden layer are as follows: x l+1 =x 0 x l w l+1 +b l+1 +x l Among them, x l represents the output of the lth hidden layer, w l represents the neuron weight of the lth hidden layer, b l represents the bias of the lth hidden layer; The feature automatic combination layer uses the following function as the neural network activation function: and i =a i ×(1-p i )×y i +p i ×y i ; where y i is the output of neuron i, E[y i ] and Var[y i ] is the mean and variance obtained during training, ∈ is a small positive number; a i is a very small hyperparameter, set to 0.01, p i is a trainable variable used to activate the propagation of neurons.

2. The personalized video recommendation method based on automatic feature combination according to claim 1, It is characterized in that The similar recall submodule, popular recall submodule, manual recall submodule and precise sorting module are all trained offline and updated periodically or irregularly.

3. The personalized video recommendation method based on automatic feature combination according to claim 1, It is characterized in that The calculation model of the user's preference weight for the video is: weight=(click_time×click_weight+like_time×like_weight+favorite_time×favorite_weight)×watch_time_ratio Among them, weight is the user's preference weight for the video; click_time, like_time, and favorite_time respectively represent the number of times the user clicks on the video, whether the user likes the video, and whether the user collects the video; click_weight, like_weight, and favorite_weight respectively represent the weights of the corresponding behaviors; watch_time_ratio represents the proportion of the user's viewing time of the video to the total video length.

4. A personalized video recommendation system based on automatic feature combination as claimed in any one of claims 1 to 3, It is characterized in that include: A side information module, used to collect user side information and video side information, wherein the user side information includes user portraits, and the video side information includes attribute information of the video; The interaction record module is used to record the historical interaction information between users and videos, including user click behavior, like behavior, collection behavior and viewing time; Rapid Recall Module, and The precise sorting module is used to sort the user's unique identifier and the corresponding video recommendation candidate list to obtain an ordered video recommendation list; Among them, the rapid recall module includes: Similarity recall submodule, used to obtain the embedding representation vector of users and videos based on the user's unique identifier; A vector similarity retrieval tool, used to search for a plurality of initial recommended videos of interest to the user based on the embedding representation vectors of the user and the video; Popular recall submodule, used to obtain multiple popular videos; A manual recall submodule, used to obtain multiple manually selected videos; And, a video combination module is used to combine the multiple initial recommended videos, the multiple popular videos, and the multiple manually selected videos to form a video recommendation candidate list.

5. A computer system comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, It is characterized in that When the processor executes the computer program, the steps of the method according to any one of claims 1 to 3 are implemented.

6. A computer storage medium having a computer program stored thereon, It is characterized in that When the program is executed by a processor, the steps of the method described in any one of claims 1 to 3 are implemented.

Citation Information

Patent Citations

  • Personalized video recommending system and method

    CN106326413A

  • Information recommendation method and apparatus, electronic device, and storage medium

    WO2021135701A1