Live broadcast recommendation method based on optimization convolution and self-enhancement sequence denoising

Abstract

The invention provides a live broadcast recommendation method based on optimization convolution and self-enhancement sequence denoising, relates to the technical field of data processing, and adopts a self-enhancement sequence denoising technology to denoise user interaction data. Besides, a convolution layer is improved, a default convolution function of pytorch is replaced, and the recommendation accuracy is improved. The user behavior sequence is processed through an LSTM (Long Short Term Memory) network, and the watching preference is predicted according to the preference change, so that the recommendation method disclosed by the invention obtains relatively good accuracy in the test, can process the user preference change and the user interaction noise, and has relatively good performance under Hit Ratio, NDCG (Non-Discrete Current Generation) and MRR (Multi-Rate Ratio) indexes. And after the convolution improvement, the recommendation accuracy is obviously improved, and each index is superior to that before the improvement.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising. Background Art

[0002] With the development of Internet technology, live streaming platforms have become an important form of media and entertainment. These platforms recommend content to users through recommendation systems to improve user engagement and satisfaction. Recommendation systems mainly rely on technologies such as data mining, machine learning, and statistics to predict users' interest in unseen items by analyzing users' historical behavior and preference data. The core idea is to use existing information to infer unknown information, thereby providing personalized recommendations.

[0003] The Chinese patent "CN107613395B Live Room Recommendation Method, System, Device and Storage Medium" proposes an Internet live room recommendation method and system. The specific steps include user and live room screening, similarity calculation, and live room recommendation. The similarity is calculated based on the valid user set shared by the valid live rooms in the historical time period. Based on the historical live rooms that the user has watched within a preset time, the live rooms with the highest similarity to these historical rooms are recommended to the user.

[0004] Chinese patent "CN105978888B A method and system for recommending virtual gifts in a live broadcast room" proposes a method for recommending virtual gifts in a live broadcast room. Through information processing, the system obtains the live broadcast room and anchor information, and normalizes the content tags. Then, list matching is performed, and the virtual gift list is matched according to the tags and anchor information, and the list is merged and deduplicated.

[0005] The Chinese patent "CN107613395B Live Room Recommendation Method, System, Device and Storage Medium" makes recommendations based on user similarity and live room similarity. On the one hand, the noise in the user interaction data is not processed, and on the other hand, the changes in user preferences over time are not taken into account. Ignoring noise will affect the accuracy of recommendations, and user preferences are highly dynamic. These two aspects will lead to insufficient recommendation accuracy and timeliness. Summary of the invention

[0006] In view of the shortcomings of the prior art, the purpose of the present invention is to propose a live broadcast recommendation method based on optimized convolution and self-enhancement sequence denoising. The noise in the user interaction data is processed and the influence of the noise is reduced by a denoising algorithm. In addition, the long short-term memory network is used to process the user behavior sequence, and recommendations can be made according to changes in user preferences;

[0007] The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising comprises:

[0008] Step 1: Obtain multiple live broadcast room data, multiple user data, and multiple user-live broadcast room interaction data, and construct the live broadcast room embedding vector e for each live broadcast room data v , construct the user's embedding vector e for each user data u ;

[0009] Step 2: According to the interactive relationship between the user and the live broadcast room, the transitional relationship between the live broadcast rooms, the incompatible relationship between the live broadcast rooms, the similarity relationship between users and the dissimilar relationship between users, the embedding vector e of the live broadcast room is v and the user's embedding vector e u updating, wherein the interactive relationship between the user and the live broadcast room, the transitional relationship between the live broadcast rooms, the incompatible relationship between the live broadcast rooms, the similarity relationship between the users, and the dissimilarity relationship between the users are obtained based on multiple live broadcast room data, multiple user data, and multiple interactive data between the user and the live broadcast room;

[0010] Step 2.1: Based on the interaction relationship between the user and the live broadcast room, calculate the user's embedding vector containing the interaction relationship between the user and the live broadcast room and the embedding vector of the live broadcast room containing the interaction relationship between the user and the live broadcast room The interactive relationship between the user and the live broadcast room indicates that there is data interaction between the user and the live broadcast room;

[0011] Step 2.2: According to the transition relationship between the live broadcast rooms, the embedding vector of the live broadcast room containing the transition relationship is calculated The transition relationship indicates that two live broadcast rooms of the same user are viewed in sequence;

[0012] Step 2.3: According to the incompatible relationship between the live broadcast rooms, the embedding vector of the live broadcast room containing the incompatible relationship is calculated The incompatible relationship indicates that for the same user, only one of the two live broadcast rooms is viewed, and both of the two live broadcast rooms have a transition relationship with another live broadcast room other than the two live broadcast rooms;

[0013] Step 2.4: Based on the similarity relationship between users, calculate the embedding vector of users with similar relationship The similarity relationship indicates that two users have the same interaction data for the same live broadcast room;

[0014] Step 2.5: Based on the dissimilar relationships between users, calculate the embedding vector of the users containing the dissimilar relationships The dissimilar relationship indicates that the two users do not have similar interaction data for the same live broadcast room, and the two users have at least one common user with a similar relationship;

[0015] Step 2.6: Embedding vector of the user including the interaction relationship between the user and the live broadcast room Embedding vectors of users with similar relationships and the embedding vectors of users with dissimilar relations Perform the merge process to obtain the updated user embedding vector h u , which is specifically expressed by the following formula:

[0016]

[0017] Among them, f u is the merge function;

[0018] Embedding vector of the live broadcast room containing the interaction relationship between users and the live broadcast room Embedding vector of the live broadcast room including transition relations And the embedding vector of the live broadcast room containing incompatible relations Perform the merge process to obtain the updated embedding vector h of the live broadcast room v , which is specifically expressed by the following formula:

[0019]

[0020] Among them, f v is the merge function;

[0021] Step 3: From the interaction data between multiple users and the live broadcast room, obtain the interaction data between the user and the live broadcast room including the target user, generate a sequence S according to the time sequence of the interaction data including the target user, and update the embedding vector h of the live broadcast room according to the updated embedding vector h of the live broadcast room. v and the updated user embedding vector h u , construct the embedding vector of the interaction data between users and the live broadcast room All embedding vectors in sequence S Construct the vector matrix H S , according to the embedding vector of the interaction data between the user and the live broadcast room in sequence S and vector matrix H S , construct a new sequence;

[0022] According to all the embedding vectors of sequence S and vector matrix H S , construct a new sequence, which specifically includes the following steps:

[0023] Step A1: Based on the embedding vector Through the bidirectional long short-term memory network, the sequential inconsistency score of the interaction data between each user and the live broadcast room containing the target user in the sequence S is calculated.

[0024] Step A2: Based on the embedding vector Calculate the similarity inconsistency score of each interaction data containing the target user in sequence S

[0025] Step A3: For each user's interaction data with the live broadcast room in sequence S, the order inconsistency score is calculated. Inconsistency score with similarity Multiply them together to get the inconsistency score distribution r s , in the inconsistency score distribution r S Get the maximum score of the inconsistent scores, and use the position corresponding to the maximum score as the insertion position;

[0026] Step A4: According to the vector matrix H S , selecting the preceding live broadcast room data and the following live broadcast room data to be inserted in the live broadcast room sequence, the live broadcast room sequence comprising a plurality of live broadcast room data arranged in chronological order, generating the interaction data between the preceding user and the live broadcast room and the interaction data between the following user and the live broadcast room according to the user data of the target user and the preceding live broadcast room data and the following live broadcast room data, inserting the interaction data between the preceding user and the live broadcast room into the preceding order of the insertion position, and at the same time, inserting the interaction data between the following user and the live broadcast room into the following order of the insertion position, thereby generating a new sequence;

[0027] Step 4: Perform denoising on the new sequence to obtain a denoised sequence, and recommend the denoised sequence according to the recommendation algorithm.

[0028] Optionally, step 2.1 specifically includes:

[0029] is the embedding vector of user u at layer 0 and the embedding vector of the live broadcast room v at layer 0 Assign values, which are specifically expressed by the following formula:

[0030]

[0031] The embedding vector of user u is updated by the following formula:

[0032]

[0033] in, is the embedding vector of user u at the k+1th layer, Represents the set of live broadcast rooms that have data interaction with user u. is the embedding vector of the live broadcast room v at the kth layer, where k represents the number of graph convolution layers;

[0034] The embedding vector of the live broadcast room v is updated by the following formula:

[0035]

[0036] in, is the embedding vector of the live broadcast room v at the k+1th layer, Represents the set of users u that have data interaction with the live broadcast room v. is the embedding vector of user u at the kth layer;

[0037] By aggregating multiple layers of graph convolution, the embedding vectors of multiple layers of user u are aggregated, and the embedding vectors of multiple layers of live broadcast rooms v are aggregated at the same time, and the embedding vector of the user containing the interactive relationship between the user and the live broadcast room is obtained. and the embedding vector of the live broadcast room containing the interaction relationship between the user and the live broadcast room It is specifically expressed by the following formula:

[0038]

[0039] Among them, K represents the total number of graph convolution layers.

[0040] Optionally, step 2.2 specifically includes:

[0041] The transition relationship weight is calculated, which is specifically implemented by the following formula:

[0042]

[0043] Among them, v i represents the ith live broadcast room in the live broadcast room sequence, v j Indicates the jth live broadcast room in the live broadcast room sequence, and live broadcast room v i And live broadcast room v j There is a transition relationship, Dis() is used to calculate the distance between two live broadcast rooms in the target sequence, and the target sequence includes multiple users and live broadcast room data arranged in chronological order, u i is the i-th user in the user sequence, wherein the user sequence includes a plurality of user data arranged in chronological order, For live broadcast room v i And live broadcast room v j The weight of the transition relationship;

[0044] Constructing the embedding vector of the live broadcast room containing the transition relationship is specifically achieved through the following formula:

[0045]

[0046] in, is a channel-by-channel over-parameterized convolution, is the embedding vector of the i-th live room data in the live room sequence, is the embedding vector of the jth live room data in the live room sequence; For live broadcast room v i The weight of the transition relationship with the live broadcast room v, For live broadcast room v j The weight of the transition relationship between the live broadcast room v, || is the connection operation.

[0047] Optionally, step 2.3 specifically includes:

[0048] The weight of the incompatible relationship is calculated, which is specifically expressed by the following formula:

[0049]

[0050] Among them, v m represents the mth live room in the live room sequence, v n Indicates the nth live broadcast room in the live broadcast room sequence, and the live broadcast room v m And live broadcast room v n There is an incompatible relationship between k Indicates the kth live broadcast room in the live broadcast room sequence, live broadcast room v k With live broadcast room v m There is a transition relationship, live broadcast room v k With live broadcast room v n There is a transitional relationship; Indicates the live broadcast room v in the target sequence m And live broadcast room v k The order between them is first live room v m Post-live broadcast room k The weight of the transition relationship, Calculated based on the weight formula, the target sequence includes a plurality of users and live broadcast room data arranged in chronological order; Indicates the live broadcast room v in the target sequence k And live broadcast room v m The order between them is first live room v k Post-live broadcast room m The weight of the transition relationship; Indicates the live broadcast room v in the target sequence n And live broadcast room v k The order between them is first live room v n Post-live broadcast room k The weight of the transition relationship; Indicates the live broadcast room v in the target sequence k And live broadcast room v n The order between them is first live room v k Post-live broadcast room n The weight of the transition relationship; Indicates live broadcast room vm And live broadcast room v n The weight of the incompatible relationship, v k Indicates the live broadcast room v m And live broadcast room v n A collection of live broadcast rooms that all have transitional relationships;

[0051] Constructing embedding vectors of incompatible live broadcast rooms The specific formula is as follows:

[0052]

[0053] in, is a channel-by-channel over-parameterized convolution, For live broadcast room v m The weight of the incompatible relationship with the live broadcast room v, For live broadcast room v m The embedding vector of the data, e v is the embedding vector of the data in the live broadcast room v, || represents the connection operation, are trainable parameters of the convolution operation.

[0054] Optionally, step 2.4 specifically includes:

[0055] The similarity weights between users are calculated based on the Jaccard similarity measurement method, which is specifically calculated using the following formula:

[0056]

[0057] Among them, u i represents the i-th user in the user sequence, Jaccard is the metric used to calculate similarity, N(u i ) indicates that the i The set of live broadcast rooms with data interaction, N(u) represents the set of live broadcast rooms with data interaction with user u, To represent user u i The weight of the similarity between and user u; ∩ represents the intersection operation, and ∪ represents the union operation;

[0058] Construct embedding vectors of users with similar relationships This is achieved specifically through the following formula:

[0059]

[0060] Among them, U + is the set of users that have similar relationships with user u, For user u i The embedding vector of the data, e uis the embedding vector of user u’s data, is an aggregate function, and || represents a join operation.

[0061] Optionally, step 2.5 specifically includes:

[0062] The dissimilarity weights between users are calculated using the Jaccard similarity measurement method, which is specifically implemented using the following formula:

[0063]

[0064] Among them, u i represents the ith user in the user sequence, and Jaccard is used to calculate the similarity metric, N(u i ) indicates that the i The set of live broadcast rooms with data interaction, N(u) represents the set of live broadcast rooms with data interaction with user u, To represent user u i The weight of the dissimilarity between and user u, ∩ represents the intersection operation, and ∪ represents the union operation;

[0065] Constructing embedding vectors for users with dissimilar relationships This is achieved specifically through the following formula:

[0066]

[0067] Among them, U - is the set of users that have dissimilar relationships with user u, To represent user u i The weight of the dissimilarity between y and user u, For user u i The embedding vector of the data, e u is the embedding vector of user u’s data, Is an aggregate function.

[0068] Optionally, step A1 specifically calculates the sequential inconsistency score by the following formula:

[0069]

[0070] in, is the forward hidden state of the t-th item in sequence S, is the backward hidden state of the tth item in the sequence S. The forward hidden state and the backward hidden state are obtained based on the bidirectional long short-term memory network. is the embedding vector of the t-th item in the sequence S, ⊙ is the element-wise multiplication operation, ∑ d is the sum over all dimensions, is the sum of the sequence st.

[0071] Optionally, step A2 specifically calculates the similarity inconsistency score by the following formula:

[0072]

[0073] in, s i is the i-th live broadcast room in sequence S, s t is the tth live broadcast room in sequence S, and ni is the length of sequence S.

[0074] Optionally, the preceding live broadcast room data and the following live broadcast room data in step A4 are obtained in the following manner:

[0075] The context representation of the insertion position is obtained through the bidirectional long short-term memory Bi-LSTM, which is implemented by the following formula:

[0076]

[0077] Among them, q L is the above representation of the insertion position, q R The following text indicates the insertion position, s t Indicates the insertion position;

[0078] The similarity between the context representation of the insertion position and all live broadcast rooms is calculated, which is specifically expressed by the following formula:

[0079] Choose the most suitable live broadcast room from multiple live broadcast rooms:

[0080] k L =q L H v , k R =q R H v

[0081] H v is the matrix of embedding vectors of all live broadcast rooms in the live broadcast room sequence, k L Represents the similarity vector of the live broadcast room before the insertion position, k R Represents the similarity vector of the live broadcast room after the insertion position;

[0082] In the obtained multiple k L In the above example, the similarity vector with the largest value is selected, and the live broadcast room data corresponding to the similarity vector with the largest value is used as the previous live broadcast room data. R In the above example, the similarity vector with the largest value is selected, and the live broadcast room data corresponding to the similarity vector with the largest value is used as the subsequent live broadcast room data.

[0083] The beneficial effects of adopting the above technical solution are:

[0084] The recommendation method of the present invention has achieved good accuracy in the test, can handle changes in user preferences and user interaction noise, and has good performance under the indicators of Hit Ratio, NDCG, and MRR. After the improvement of convolution, the recommendation accuracy is significantly improved, and all indicators are better than before the improvement. BRIEF DESCRIPTION OF THE DRAWINGS

[0085] Figure 1 A flowchart of a live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising in an embodiment of the present invention;

[0086] Figure 2 The present invention is a flowchart of another live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising in an embodiment of the present invention. DETAILED DESCRIPTION

[0087] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0088] In view of the problems existing in the prior art, the present invention provides a live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising. Figure 1 , which may include the following steps:

[0089] Step 1: Obtain multiple live broadcast room data, multiple user data, and multiple user-live broadcast room interaction data, and construct the live broadcast room embedding vector e for each live broadcast room data v , construct the user's embedding vector e for each user data u ;

[0090] Specifically, the interactive data generated by user behavior is processed. The data required for model training is extracted from a large amount of data and processed into a format that can be used by the model, including:

[0091] (1) Extract a data set from the database that stores user browsing records and gift-giving records in the live broadcast room. The data set contains information such as the behavior and gift-giving records of each user when entering the live broadcast room. Focus on the behavior of each user every time they enter the live broadcast room and process it as raw data. For each user, calculate the two features of "number of times they enter the live broadcast room" and "total gift value" in each live broadcast room. Then, perform a weighted calculation on the number of entries and the gift value to obtain the user's rating for a certain live broadcast room. This rating indicates the user's liking for the live broadcast room.

[0092] (2) All data are divided into three files, namely ll.item (live broadcast room), ll.user (user), and ll.inter (user-live broadcast room interaction). The field names of these three files are in the format of "original field name: data type".

[0093] The ll.item file contains the item_id:token, duration:float, pv:float, uv:float, person_max_num:float, person_average_duration:float, and gift_income:float fields, which are the live broadcast room id, live broadcast duration, number of times the live broadcast page is viewed, number of unique visitors to the live broadcast page, maximum number of people in the live broadcast room, average viewing time of all users in the live broadcast room, and total gift income.

[0094] The ll.user file contains the user_id:token field, which is the user id.

[0095] The ll.inter file contains the fields user_id:token, item_id:token, rating:float, and timestamp:float, which are user id, live broadcast room id, user's rating of the live broadcast room, and timestamp respectively.

[0096] Among them, the data in the ll.item file is the data of a live broadcast room, the data in the ll.user file is the data of a user, and the data in the ll.inter file is the interaction data between a user and the live broadcast room. All the data together constitute Figure 2 The original data in .

[0097] Similarly, following the above steps, data of multiple live broadcast rooms, data of multiple users, and interaction data of multiple users and live broadcast rooms can be obtained.

[0098] (3) For each live broadcast room data and each user data, each live broadcast room data and each user data can be encoded by Embedding encoding. Specifically, the live broadcast room data is mapped to the embedding vector e v , and maps the user data to the embedding vector e u , which is specifically achieved through the following formula:

[0099] e v =x v W v ;

[0100] e u =x u W u ;

[0101] Among them, W v and W u is the parameter matrix that can be trained, with dimensions of |V|×d and |U|×d, respectively. v is the encoding vector of the live broadcast room data, x u is the encoding vector of user data, x v and x u is the one-hot encoding vector of the item (v) and user (u) ID. This method effectively encodes the item and user information into vectors in a low-dimensional space for subsequent recommendation algorithm processing. v for Figure 1 The initial live broadcast room embedding vector in, e u for Figure 1 The initial user embedding vector in .

[0102] Step 2: According to the interactive relationship between the user and the live broadcast room, the transitional relationship between the live broadcast rooms, the incompatible relationship between the live broadcast rooms, the similarity relationship between users and the dissimilar relationship between users, the embedding vector e of the live broadcast room is v and the user's embedding vector e u updating, wherein the interactive relationship between the user and the live broadcast room, the transitional relationship between the live broadcast rooms, the incompatible relationship between the live broadcast rooms, the similarity relationship between the users, and the dissimilarity relationship between the users are obtained based on multiple live broadcast room data, multiple user data, and multiple interactive data between the user and the live broadcast room;

[0103] Among them, the interactive relationship between users and live broadcast rooms, the transitional relationship between live broadcast rooms, the incompatible relationship between live broadcast rooms, the similar relationship between users and the dissimilar relationship between users are reflected through the live broadcast room relationship subgraph, the user relationship subgraph and the user and live broadcast room relationship subgraph, which specifically include:

[0104] The user relationship subgraph includes a first node, a first connection line, and a second connection line. The first nodes are connected by the first connection line or the second connection line. The first node corresponds to a user. Two first nodes are connected by the first connection line, indicating that there is a similar relationship between the users corresponding to the two first nodes, and the similar relationship indicates that the two users corresponding to the two first nodes have similar interaction data for the live broadcast room; the two first nodes are connected by the second connection line, indicating that there is a dissimilar relationship between the users corresponding to the two first nodes, and the dissimilar relationship indicates that the two users corresponding to the two first nodes do not have similar interaction data for the live broadcast room, and the two users corresponding to the two first nodes have at least one common user with a similar relationship;

[0105] A similar relationship is established based on the users' common interactive behaviors, which means that two users are connected because of similar interaction patterns; a dissimilar relationship means that two users have no record of interacting with the same live broadcast room, but they have at least one common similar user.

[0106] The live broadcast room relationship subgraph includes a second node, a third connection line and an arrow, the second nodes are connected by the third connection line or the arrow, the second node corresponds to a live broadcast room, two second nodes are connected by an arrow, indicating that there is a transition relationship between the two live broadcast rooms corresponding to the two second nodes, and the transition relationship indicates that for the same user, the two live broadcast rooms are viewed in the order from the live broadcast room pointed by the arrow to the live broadcast room pointed by the arrow; the two second nodes are connected by the third connection line, indicating that there is an incompatible relationship between the two live broadcast rooms corresponding to the two second nodes, and the incompatible relationship indicates that for the same household, only one of the two live broadcast rooms corresponding to the two second nodes is viewed, and the live broadcast rooms corresponding to the two second nodes both have a transition relationship with one live broadcast room;

[0107] Transition relationship: This relationship reflects the positive dependency between live broadcast rooms, that is, the situation where two items appear in a certain order in the user behavior sequence.

[0108] Incompatible relationship: indicates the negative dependency between live broadcast rooms, which is used to prevent incompatible live broadcast rooms from being mistakenly combined during the recommendation process.

[0109] The user-live broadcast room relationship subgraph includes the first node, the second node and the fourth connecting line. The first node and the second node are connected by the fourth connecting line, indicating that there is an interactive relationship between the user corresponding to the first node and the live broadcast room corresponding to the second node. The interactive relationship between the user and the live broadcast room indicates that there is data interaction between the user corresponding to the first node and the live broadcast room corresponding to the second node.

[0110] Among them, the interactive relationship between users and live broadcast rooms is constructed by analyzing the interactive data between users and live broadcast rooms, with the aim of quantifying the intensity of interaction between users and live broadcast rooms.

[0111] By establishing these complex relationships, the recommendation system can select and locate items more accurately, thereby more effectively improving the accuracy of recommendations and user satisfaction when enhancing sequential operations.

[0112] Step 2.1: Based on the interaction relationship between the user and the live broadcast room, calculate the user's embedding vector containing the interaction relationship between the user and the live broadcast room and the embedding vector of the live broadcast room containing the interaction relationship between the user and the live broadcast room The interactive relationship between the user and the live broadcast room indicates that there is data interaction between the user and the live broadcast room;

[0113] This step is implemented based on the interaction relationship encoding layer between users and live broadcast rooms. This layer focuses on encoding the directional sequence dependencies between items and distinguishing the live broadcast rooms from and pointed to. The attention mechanism is used to assign different weights to these neighbors, and the convolution operator is used to aggregate information to learn transition-aware representations. LightGCN is used to encode the interaction between users and items. LightGCN is a simplified GCN (graph convolution) model that only retains the core operation of graph convolution and removes the nonlinear activation function and feature transformation matrix. Specifically, it includes:

[0114] is the embedding vector of user u at layer 0 and the embedding vector of the live broadcast room v at layer 0 Assign values, which are specifically expressed by the following formula:

[0115]

[0116] The embedding vector of user u is updated by the following formula:

[0117]

[0118] in, is the embedding vector of user u at the k+1th layer, Represents the set of live broadcast rooms that have data interaction with user u. is the embedding vector of the live broadcast room v at the kth layer, where k represents the number of graph convolution layers;

[0119] The embedding vector of the live broadcast room v is updated by the following formula:

[0120]

[0121] in, is the embedding vector of the live broadcast room v at the k+1th layer, Represents the set of users u that have data interaction with the live broadcast room v. is the embedding vector of user u at the kth layer;

[0122] By aggregating multiple layers of graph convolution, the embedding vectors of multiple layers of user u are aggregated, and the embedding vectors of multiple layers of live broadcast rooms v are aggregated at the same time, and the embedding vector of the user containing the interactive relationship between the user and the live broadcast room is obtained. and the embedding vector of the live broadcast room containing the interaction relationship between the user and the live broadcast room It is specifically expressed by the following formula:

[0123]

[0124] Among them, K represents the total number of graph convolution layers.

[0125] Step 2.2: According to the transition relationship between the live broadcast rooms, the embedding vector of the live broadcast room containing the transition relationship is calculated The transition relationship indicates that two live broadcast rooms of the same user are viewed in sequence;

[0126] This step is implemented based on the live broadcast room transition relationship encoding layer, which focuses on encoding the directional sequence dependencies between live broadcast rooms and distinguishing the live broadcast rooms from and pointed to. The attention mechanism is used to assign different weights to these live broadcast rooms, and the convolution operator is used to aggregate information to learn transition-aware representations, including:

[0127] The transition relationship weight is calculated, which is specifically implemented by the following formula:

[0128]

[0129] Wherein, vi represents the ith live broadcast room in the live broadcast room sequence, vj represents the jth live broadcast room in the live broadcast room sequence, and there is a transition relationship between live broadcast room vi and live broadcast room vj. Dis() is used to calculate the distance between two live broadcast rooms in the target sequence, and the target sequence includes multiple users and live broadcast room data arranged in chronological order. u i is the i-th user in the user sequence, wherein the user sequence includes a plurality of user data arranged in chronological order, For live broadcast room v i And live broadcast room v j The weight of the transition relationship;

[0130] Constructing the embedding vector of the live broadcast room containing the transition relationship is specifically achieved through the following formula:

[0131]

[0132] in, is a channel-by-channel over-parameterized convolution, is the embedding vector of the i-th live room data in the live room sequence, is the embedding vector of the jth live room data in the live room sequence; For live broadcast room v i The weight of the transition relationship with the live broadcast room v, For live broadcast room v j The weight of the transition relationship between the live broadcast room v, || is the connection operation.

[0133] Step 2.3: According to the incompatible relationship between the live broadcast rooms, the embedding vector of the live broadcast room containing the incompatible relationship is calculated The incompatible relationship indicates that for the same user, only one of the two live broadcast rooms is viewed, and both of the two live broadcast rooms have a transition relationship with another live broadcast room other than the two live broadcast rooms;

[0134] This step is based on the incompatibility relationship encoding layer, which encodes the non-directional incompatibility relationships between live broadcast rooms. Since these relationships do not distinguish between different types of live broadcast rooms, a convolution operator similar to the previous layer but with different parameters learns the incompatibility-aware representation, which includes:

[0135] The weight of the incompatible relationship is calculated, which is specifically expressed by the following formula:

[0136]

[0137] Among them, v m represents the mth live room in the live room sequence, v n Indicates the nth live broadcast room in the live broadcast room sequence, and the live broadcast room v m And live broadcast room v n There is an incompatible relationship between k Indicates the kth live broadcast room in the live broadcast room sequence, live broadcast room v k With live broadcast room v m There is a transition relationship, live broadcast room v k With live broadcast room v n There is a transitional relationship; Indicates the live broadcast room v in the target sequence m And live broadcast room v k The order between them is first live room v m Post-live broadcast room k The weight of the transition relationship, Calculated based on the weight formula, the target sequence includes a plurality of users and live broadcast room data arranged in chronological order; Indicates the live broadcast room v in the target sequence k And live broadcast room v m The order between them is first live room v k Post-live broadcast room m The weight of the transition relationship; Indicates the live broadcast room v in the target sequence n And live broadcast room v k The order between them is first live room v n Post-live broadcast room k The weight of the transition relationship; Indicates the live broadcast room v in the target sequence k And live broadcast room v n The order between them is first live room v k Post-live broadcast room n The weight of the transition relationship; Indicates live broadcast room v m And live broadcast room v n The weight of the incompatible relationship, v k Indicates the live broadcast room v m And live broadcast room v n A collection of live broadcast rooms that all have transitional relationships;

[0138] Constructing embedding vectors of incompatible live broadcast rooms The specific formula is as follows:

[0139]

[0140] in, is a channel-by-channel over-parameterized convolution, For live broadcast room v m The weight of the incompatible relationship with the live broadcast room v, For live broadcast room v m The embedding vector of the data, e v is the embedding vector of the data in the live broadcast room v, || represents the connection operation, are trainable parameters of the convolution operation.

[0141] Step 2.4: Based on the similarity relationship between users, calculate the embedding vector of users with similar relationship The similarity relationship indicates that two users have the same interaction data for the same live broadcast room;

[0142] This step is implemented based on the user similarity relationship encoding layer, which identifies and encodes similar users based on interaction patterns. A convolutional function similar to the live broadcast room layer aggregates and learns the user's similar perception representation, including:

[0143] The similarity weights between users are calculated based on the Jaccard similarity measurement method, which is specifically calculated using the following formula:

[0144]

[0145] Among them, u i represents the i-th user in the user sequence, Jaccard is the metric used to calculate similarity, N(u i ) indicates that the i The set of live broadcast rooms with data interaction, N(u) represents the set of live broadcast rooms with data interaction with user u, To represent user u i The weight of the similarity between and user u; ∩ represents the intersection operation, and ∪ represents the union operation;

[0146] Construct embedding vectors of users with similar relationships This is achieved specifically through the following formula:

[0147]

[0148] Among them, U + is the set of users that have similar relationships with user u, For user u i The embedding vector of the data, e u is the embedding vector of user u’s data, is an aggregate function, and || represents a join operation.

[0149] Step 2.5: Based on the dissimilar relationships between users, calculate the embedding vector of the users containing the dissimilar relationships The dissimilar relationship indicates that the two users do not have similar interaction data for the same live broadcast room, and the two users have at least one common user with a similar relationship;

[0150] This step 2.5 is implemented based on the user dissimilarity encoding layer, which focuses on encoding dissimilarity between users to reduce irrelevant recommendations. A convolutional function similar to the incompatible live broadcast room layer but with different parameters creates a dissimilarity-aware representation, including:

[0151] The dissimilarity weights between users are calculated using the Jaccard similarity measurement method, which is specifically implemented using the following formula:

[0152]

[0153] Among them, u i represents the ith user in the user sequence, and Jaccard is used to calculate the similarity metric, N(u i ) indicates that the i The set of live broadcast rooms with data interaction, N(u) represents the set of live broadcast rooms with data interaction with user u, To represent user u i The weight of the dissimilarity between and user u, ∩ represents the intersection operation, and ∪ represents the union operation;

[0154] Constructing embedding vectors for users with dissimilar relationships This is achieved specifically through the following formula:

[0155]

[0156] Among them, U - is the set of users that have dissimilar relationships with user u, To represent user u i The weight of the dissimilarity between y and user u, For user u i The embedding vector of the data, e u is the embedding vector of user u’s data, Is an aggregate function.

[0157] Among them, the user-live broadcast room interaction relationship encoding layer, the live broadcast room transition relationship encoding layer, the live broadcast room incompatible relationship encoding layer, the user similarity relationship encoding layer implementation and the user dissimilar relationship encoding layer constitute the global relationship encoder.

[0158] Among them, the convolution operation mentioned above adds an additional channel-by-channel convolution (depthwise convolution) to the traditional convolution layer, so that each input channel is convolved with a different 2D kernel. This combination of structures constitutes over-parameterization by increasing learnable parameters. It effectively improves the performance of convolutional neural networks in processing tasks while maintaining the computational efficiency of the reasoning stage. In different encoding layers in the global relation encoder, the relationship between users and items is captured through convolution operations. In order to improve the relationship extraction capability of convolution, an additional channel-by-channel convolution (depthwise convolution) is added to enable the subsequent global relation encoder to achieve better results.

[0159] The improvements to convolution are illustrated by the following examples:

[0160] Assume that the input feature map size is H×W×Cin, where H and W are the height and width of the feature map, respectively, and Cin is the number of input channels; Depthwise Convolution, in which each input channel is convolved using an independent convolution kernel, which is different from the standard convolution where all channels share the convolution kernel. Each input channel c uses a convolution kernel Dc of size K×K (the spatial size of the convolution kernel). This convolution is performed independently on each channel to generate an intermediate feature map H×W×Cin. The channel-by-channel convolution operation here does not mix the information between different channels, but only extracts features within each channel. Over-parameterization Component, the core idea of ​​over-parameterization is to increase the number of parameters to improve the feature learning ability of the model. DO-Conv achieves this by adding an additional convolution operation. For the output feature map of each channel, it is further passed through a standard convolution layer. The convolution kernel size of this convolution layer is 1×1, but the number of each convolution kernel is equal to the depth multiplier (that is, the additional number of parameters).

[0161] This operation introduces more parameters in the convolutional layer and enhances the expressiveness of the features.

[0162] Step 2.6: Embedding vector of the user including the interaction relationship between the user and the live broadcast room Embedding vectors of users with similar relationships and the embedding vectors of users with dissimilar relations Perform the merge process to obtain the updated user embedding vector h u , which is specifically expressed by the following formula:

[0163]

[0164] Among them, f u is the merge function;

[0165] Embedding vector of the live broadcast room containing the interaction relationship between users and the live broadcast room Embedding vector of the live broadcast room including transition relations And the embedding vector of the live broadcast room containing incompatible relations Perform the merge process to obtain the updated embedding vector h of the live broadcast room v , which is specifically expressed by the following formula:

[0166]

[0167] Among them, f v is the merge function;

[0168] Step 3: From the interaction data between multiple users and the live broadcast room, obtain the interaction data between the user and the live broadcast room including the target user, generate a sequence S according to the time sequence of the interaction data including the target user, and update the embedding vector h of the live broadcast room according to the updated embedding vector h of the live broadcast room. v and the updated user embedding vector h u , construct the embedding vector of the interaction data between users and the live broadcast room All embedding vectors in sequence S Construct vector matrix H s , according to the embedding vector of the interaction data between the user and the live broadcast room in sequence S and vector matrix H S , construct a new sequence;

[0169] According to all the embedding vectors of sequence S and vector matrix H S , construct a new sequence, the present invention is used to achieve sequence position enhancement, such as Figure 2, which is implemented through sequential discriminator, similarity discriminator, position selector and item selector. The overall goal is to enrich sequence information, alleviate outlier user preferences (OUPs), and provide more accurate and relevant recommendations by strategically selecting items and positions in the sequence. The specific steps include:

[0170] Step A1: Based on the embedding vector Through the bidirectional long short-term memory network, the sequential inconsistency score of the interaction data between each user and the live broadcast room containing the target user in the sequence S is calculated.

[0171] Step A1 is implemented based on the sequential discriminator, using the Bi-LSTM (Bi-directional Long Short-Term Memory) context-aware encoder to analyze the sequential dependencies of items, allowing detection of locations where global transfer relationships and local dependencies may conflict. The sequential discriminator is used to calculate the sequential inconsistency score, which is specifically calculated using the following formula:

[0172]

[0173] in, is the forward hidden state of the t-th item in sequence S, is the backward hidden state of the tth item in the sequence S. The forward hidden state and the backward hidden state are obtained based on the bidirectional long short-term memory network. is the embedding vector of the t-th item in the sequence S, ⊙ is the element-wise multiplication operation, ∑ d is the sum over all dimensions, is the sequence s t Sum.

[0174] Step A2: Based on the embedding vector Calculate the similarity inconsistency score of each interaction data containing the target user in sequence S

[0175] Step A2 is implemented based on a similarity discriminator, which calculates an inconsistency score based on the similarity of the live broadcast room to help further determine the possible inconsistent positions in the sequence. The similarity discriminator is used to calculate the similarity inconsistency score of each item. Specifically, the similarity inconsistency score is calculated by the following formula:

[0176]

[0177] in, s i is the i-th live broadcast room in sequence S, s tis the tth live broadcast room in sequence S, and ni is the length of sequence S.

[0178] Step A3: For each user's interaction data with the live broadcast room in sequence S, the sequential inconsistency score is calculated. Inconsistency score with similarity Multiply them together to get the inconsistency score distribution r s , in the inconsistency score distribution r s Get the maximum score of the inconsistent scores, and use the position corresponding to the maximum score as the insertion position;

[0179] Among them, step A3 is based on the location selector implementation, which identifies the most suitable location in the sequence for inserting the live broadcast room, balancing computational efficiency and effectiveness. It uses global and local information to detect inconsistencies, indicating that the items in these locations may not be well coordinated with the surrounding environment and may benefit from additional information.

[0180] Step A4: According to the vector matrix H S , selecting the preceding live broadcast room data and the following live broadcast room data to be inserted in the live broadcast room sequence, the live broadcast room sequence comprising a plurality of live broadcast room data arranged in chronological order, generating the interaction data between the preceding user and the live broadcast room and the interaction data between the following user and the live broadcast room according to the user data of the target user and the preceding live broadcast room data and the following live broadcast room data, inserting the interaction data between the preceding user and the live broadcast room into the preceding order of the insertion position, and at the same time, inserting the interaction data between the following user and the live broadcast room into the following order of the insertion position, thereby generating a new sequence;

[0181] Step A4 is implemented based on the item selector. After determining the best enhancement position, the most suitable data for insertion is selected. It selects data that can well complement the existing sequence, maintain fluency and minimize interference. The preceding live broadcast room data and the following live broadcast room data are obtained in the following way:

[0182] The context representation of the insertion position is obtained through the bidirectional long short-term memory Bi-LSTM, which is implemented by the following formula:

[0183]

[0184] Among them, q L is the above representation of the insertion position, q R The following text indicates the insertion position, s t Indicates the insertion position;

[0185] The similarity between the context representation of the insertion position and all live broadcast rooms is calculated, which is specifically expressed by the following formula:

[0186] Choose the most suitable live broadcast room from multiple live broadcast rooms:

[0187] k L =q L H v , k R =q R H v

[0188] H v is the matrix of embedding vectors of all live broadcast rooms in the live broadcast room sequence, k L Represents the similarity vector of the live broadcast room before the insertion position, k R Represents the similarity vector of the live broadcast room after the insertion position;

[0189] In the obtained multiple k L In the above example, the similarity vector with the largest value is selected, and the live broadcast room data corresponding to the similarity vector with the largest value is used as the previous live broadcast room data. R In the above example, the similarity vector with the largest value is selected, and the live broadcast room data corresponding to the similarity vector with the largest value is used as the subsequent live broadcast room data.

[0190] Step 4: Perform denoising on the new sequence to obtain a denoised sequence, and recommend the denoised sequence according to the recommendation algorithm.

[0191] The above description is only a preferred embodiment of the present disclosure and an explanation of the technical principles used. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present disclosure is not limited to the technical solutions formed by a specific combination of the above-mentioned technical features, but should also cover other technical solutions formed by any combination of the above-mentioned technical features or their equivalent features without departing from the above-mentioned inventive concept. For example, the above-mentioned features are replaced with the technical features with similar functions disclosed in the embodiments of the present disclosure (but not limited to) to form a technical solution.

Claims

1. A live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising, characterized in that: include: Step 1: Obtain multiple live broadcast room data, multiple user data, and multiple user-live broadcast room interaction data, and construct the live broadcast room embedding vector e for each live broadcast room data v , construct the user's embedding vector e for each user data u ; Step 2: According to the interactive relationship between users and live broadcast rooms, the transitional relationship between live broadcast rooms, the incompatible relationship between live broadcast rooms, the similarity relationship between users and the dissimilarity relationship between users, the embedding vector e of the live broadcast room is v and the user's embedding vector e u updating, wherein the interactive relationship between the user and the live broadcast room, the transitional relationship between the live broadcast rooms, the incompatible relationship between the live broadcast rooms, the similarity relationship between the users, and the dissimilarity relationship between the users are obtained based on multiple live broadcast room data, multiple user data, and multiple interactive data between the user and the live broadcast room; Step 2.1: Based on the interaction relationship between the user and the live broadcast room, calculate the user's embedding vector containing the interaction relationship between the user and the live broadcast room and the embedding vector of the live broadcast room containing the interaction relationship between the user and the live broadcast room The interactive relationship between the user and the live broadcast room indicates that there is data interaction between the user and the live broadcast room; Step 2.2: According to the transition relationship between the live broadcast rooms, the embedding vector of the live broadcast room containing the transition relationship is calculated The transition relationship indicates that two live broadcast rooms of the same user are viewed in sequence; Step 2.3: According to the incompatible relationship between the live broadcast rooms, the embedding vector of the live broadcast room containing the incompatible relationship is calculated The incompatible relationship indicates that for the same user, only one of the two live broadcast rooms is viewed, and both of the two live broadcast rooms have a transition relationship with another live broadcast room other than the two live broadcast rooms; Step 2.4: Based on the similarity relationship between users, calculate the embedding vector of users with similar relationship The similarity relationship indicates that two users have the same interaction data for the same live broadcast room; Step 2.5: Based on the dissimilar relationships between users, calculate the embedding vector of the users containing the dissimilar relationships The dissimilar relationship indicates that the two users do not have similar interaction data for the same live broadcast room, and the two users have at least one common user with a similar relationship; Step 2.6: Embedding vector of the user including the interaction relationship between the user and the live broadcast room Embedding vectors of users with similar relationships and the embedding vectors of users with dissimilar relations Perform the merge process to obtain the updated user embedding vector h u , which is specifically expressed by the following formula: Among them, f u is the merge function; Embedding vector of the live broadcast room containing the interaction relationship between users and the live broadcast room Embedding vector of the live broadcast room including transition relations And the embedding vector of the live broadcast room containing incompatible relations Perform the merge process to obtain the updated embedding vector h of the live broadcast room υ , which is specifically expressed by the following formula: Among them, f υ is the merge function; Step 3: From the interaction data between multiple users and the live broadcast room, obtain the interaction data between the user and the live broadcast room including the target user, generate a sequence S according to the time sequence of the interaction data including the target user, and update the embedding vector h of the live broadcast room according to the updated embedding vector h of the live broadcast room. v and the updated user embedding vector h u , construct the embedding vector of the interaction data between users and the live broadcast room All embedding vectors in sequence S Construct the vector matrix H S , according to the embedding vector of the interaction data between the user and the live broadcast room in sequence S and vector matrix H S , construct a new sequence; According to all the embedding vectors of sequence S and vector matrix H S , construct a new sequence, which specifically includes the following steps: Step A1: Based on the embedding vector Through the bidirectional long short-term memory network, the sequential inconsistency score of the interaction data between each user and the live broadcast room containing the target user in the sequence S is calculated. Step A2: Based on the embedding vector Calculate the similarity inconsistency score of each interaction data containing the target user in sequence S Step A3: For each user's interaction data with the live broadcast room in sequence S, the sequential inconsistency score is calculated. Inconsistency score with similarity Multiply them together to get the inconsistency score distribution r s , in the inconsistency score distribution r s Get the maximum score of the inconsistent scores, and use the position corresponding to the maximum score as the insertion position; Step A4: According to the vector matrix H S , selecting the preceding live broadcast room data and the following live broadcast room data to be inserted in the live broadcast room sequence, the live broadcast room sequence comprising a plurality of live broadcast room data arranged in chronological order, generating the interaction data between the preceding user and the live broadcast room and the interaction data between the following user and the live broadcast room according to the user data of the target user and the preceding live broadcast room data and the following live broadcast room data, inserting the interaction data between the preceding user and the live broadcast room into the preceding order of the insertion position, and at the same time, inserting the interaction data between the following user and the live broadcast room into the following order of the insertion position, thereby generating a new sequence; Step 4: Perform denoising on the new sequence to obtain a denoised sequence, and recommend the denoised sequence according to the recommendation algorithm.

2. The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising according to claim 1 is characterized in that: Step 2.1 specifically includes: is the embedding vector of user u at layer 0 and the embedding vector of the live broadcast room v at layer 0 Assign values, which are specifically expressed by the following formula: The embedding vector of user u is updated by the following formula: in, is the embedding vector of user u at the k+1th layer, Represents the set of live broadcast rooms that have data interaction with user u. is the embedding vector of the live broadcast room v at the kth layer, where k represents the number of graph convolution layers; The embedding vector of the live broadcast room v is updated by the following formula: in, is the embedding vector of the live broadcast room v at the k+1th layer, Represents the set of users u that have data interaction with the live broadcast room v. is the embedding vector of user u at the kth layer; By aggregating multiple layers of graph convolution, the embedding vectors of multiple layers of user u are aggregated, and the embedding vectors of multiple layers of live broadcast rooms v are aggregated at the same time, and the embedding vector of the user containing the interactive relationship between the user and the live broadcast room is obtained. and the embedding vector of the live broadcast room containing the interaction relationship between the user and the live broadcast room It is specifically expressed by the following formula: Among them, K represents the total number of graph convolution layers.

3. The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising according to claim 1 is characterized in that: Step 2.2 specifically includes: The transition relationship weight is calculated, which is specifically implemented by the following formula: Among them, v i represents the ith live broadcast room in the live broadcast room sequence, v j Indicates the jth live broadcast room in the live broadcast room sequence, and live broadcast room v i And live broadcast room v j There is a transition relationship, Dis() is used to calculate the distance between two live broadcast rooms in the target sequence, and the target sequence includes multiple users and live broadcast room data arranged in chronological order, u i is the i-th user in the user sequence, wherein the user sequence includes a plurality of user data arranged in chronological order, For live broadcast room v i And live broadcast room v j The weight of the transition relationship; Constructing the embedding vector of the live broadcast room containing the transition relationship is specifically achieved through the following formula: in, is a channel-by-channel over-parameterized convolution, is the embedding vector of the i-th live room data in the live room sequence, is the embedding vector of the jth live room data in the live room sequence; For live broadcast room v i The weight of the transition relationship with the live broadcast room v, For live broadcast room v j The weight of the transition relationship between the live broadcast room v, || is the connection operation.

4. The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising according to claim 1 is characterized in that: Step 2.3 specifically includes: The weight of the incompatible relationship is calculated, which is specifically expressed by the following formula: Among them, v m represents the mth live room in the live room sequence, v n Indicates the nth live broadcast room in the live broadcast room sequence, and the live broadcast room v m And live broadcast room v n There is an incompatible relationship between k Indicates the kth live broadcast room in the live broadcast room sequence, live broadcast room v k With live broadcast room v m There is a transition relationship, live broadcast room v k With live broadcast room v n There is a transitional relationship; Indicates the live broadcast room v in the target sequence m And live broadcast room v k The order between them is first live room v m Post-live broadcast room k The weight of the transition relationship, Calculated based on the weight formula, the target sequence includes a plurality of users and live broadcast room data arranged in chronological order; Indicates the live broadcast room v in the target sequence k And live broadcast room v m The order between them is first live room v k Post-live broadcast room m The weight of the transition relationship; Indicates the live broadcast room v in the target sequence n And live broadcast room v k The order between them is first live room v n Post-live broadcast room k The weight of the transition relationship; Indicates the live broadcast room v in the target sequence k And live broadcast room v n The order between them is first live room v k Post-live broadcast room n The weight of the transition relationship; Indicates live broadcast room v m And live broadcast room v n The weight of the incompatible relationship, v k Indicates the live broadcast room v m And live broadcast room v n A collection of live broadcast rooms that all have transitional relationships; Constructing embedding vectors of incompatible live broadcast rooms The specific formula is as follows: in, is a channel-by-channel over-parameterized convolution, For live broadcast room v m The weight of the incompatible relationship with the live broadcast room v, For live broadcast room v m The embedding vector of the data, e υ is the embedding vector of the data in the live broadcast room v, || represents the connection operation, are trainable parameters of the convolution operation.

5. The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising according to claim 1 is characterized in that: Step 2.4 specifically includes: The similarity weights between users are calculated based on the Jaccard similarity measurement method, which is specifically calculated using the following formula: Among them, u i represents the i-th user in the user sequence, Jaccard is the metric used to calculate similarity, N(u i ) indicates that the i The set of live broadcast rooms with data interaction, N(u) represents the set of live broadcast rooms with data interaction with user u, To represent user u i The weight of the similarity between and user u; ∩ represents the intersection operation, and ∪ represents the union operation; Construct embedding vectors of users with similar relationships This is achieved specifically through the following formula: Among them, U + is the set of users that have similar relationships with user u, For user u i The embedding vector of the data, e u is the embedding vector of user u’s data, is an aggregate function, and || represents a join operation.

6. The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising according to claim 1 is characterized in that: Step 2.5 specifically includes: The dissimilarity weights between users are calculated using the Jaccard similarity measurement method, which is specifically implemented using the following formula: Among them, u i represents the ith user in the user sequence, and Jaccard is used to calculate the similarity metric, N(u i ) indicates that the i The set of live broadcast rooms with data interaction, N(u) represents the set of live broadcast rooms with data interaction with user u, To represent user u i The weight of the dissimilarity between and user u, ∩ represents the intersection operation, and ∪ represents the union operation; Constructing embedding vectors for users with dissimilar relationships This is achieved specifically through the following formula: Among them, U - is the set of users that have dissimilar relationships with user u, To represent user u i The weight of the dissimilarity between y and user u, For user u i The embedding vector of the data, e u is the embedding vector of user u’s data, Is an aggregate function.

7. The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising according to claim 1 is characterized in that: Step A1 specifically calculates the sequential inconsistency score by the following formula: in, is the forward hidden state of the t-th item in sequence S, is the backward hidden state of the tth item in the sequence S. The forward hidden state and the backward hidden state are obtained based on the bidirectional long short-term memory network. is the embedding vector of the t-th item in the sequence S, ⊙ is the element-wise multiplication operation, ∑ d is the sum over all dimensions, is the sequence s t Sum.

8. The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising according to claim 1 is characterized in that: Step A2 specifically calculates the similarity inconsistency score by the following formula: in, s i is the i-th live broadcast room in sequence S, s t is the tth live broadcast room in sequence S, and ni is the length of sequence S.

9. The live broadcast recommendation method based on optimized convolution and self-enhanced sequence denoising according to claim 1, characterized in that: The preceding live broadcast room data and the subsequent live broadcast room data in step A4 are obtained in the following manner: The context representation of the insertion position is obtained through the bidirectional long short-term memory Bi-LSTM, which is implemented by the following formula: Among them, q L is the above representation of the insertion position, q R The following text indicates the insertion position, s t Indicates the insertion position; The similarity between the context representation of the insertion position and all live broadcast rooms is calculated, which is specifically expressed by the following formula: Choose the most suitable live broadcast room from multiple live broadcast rooms: k L =q L H υ ,k R =q R H υ H υ is the matrix of embedding vectors of all live broadcast rooms in the live broadcast room sequence, k L Represents the similarity vector of the live broadcast room before the insertion position, k R Represents the similarity vector of the live broadcast room after the insertion position; In the obtained multiple k L In the above example, the similarity vector with the largest value is selected, and the live broadcast room data corresponding to the similarity vector with the largest value is used as the previous live broadcast room data. R In the above example, the similarity vector with the largest value is selected, and the live broadcast room data corresponding to the similarity vector with the largest value is used as the subsequent live broadcast room data.

Images