A method and system for recommending programs by time period for Internet television

Abstract

The present invention discloses a method and system for recommending programs by time period for Internet TV, which relates to the technical field of intelligent TV program recommendation, including obtaining the historical records of users watching TV programs, dividing them by time period, and constructing a user behavior sequence; extracting the user's personalized interest feature vector, the historical program time-related feature vector, the candidate program time-related feature vector, and the candidate program general feature vector; calculating the short-term interest feature vector based on the above feature vector, and further calculating the user's long-term preference score and short-term preference score for the candidate program, and obtaining the final candidate program prediction probability by fusion calculation; constructing and optimizing the recommendation loss function, obtaining an optimized time-division program recommendation model, outputting a time-division TV program recommendation list, and realizing the time-division recommendation of TV programs. The present invention takes into account the situation that different customers have different interests in different time periods, and realizes more accurate and personalized program recommendations for customers.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent television program recommendation, and more specifically, to a method and system for recommending programs by time period for Internet television. Background Art

[0002] In recent years, the information resources on the Internet have been continuously enriched, and people are becoming more and more dependent on online information. In this era, it is becoming increasingly difficult for users to obtain the parts that they are really interested in when faced with a large amount of information. How to locate the information that users are interested in and display it to users has also become a challenge that information systems need to face. Therefore, personalized recommendation systems that actively recommend preferred content to users by modeling the user's historical behavior data have important and practical application significance. In recent years, recommendation systems have been successfully applied to many fields, and the most common application scenarios are online video, social networks, online music, e-commerce, etc.

[0003] Traditional collaborative filtering methods are often used as personalized recommendation methods for TV programs. However, such methods make personalized recommendations for different users based on the user's previous historical records, but they do not take into account the impact of time on user interests, that is, users' interest preferences will be different in different time periods of the day (such as early morning, morning, noon and evening). As a result, existing TV program recommendation technologies have problems with recommendation accuracy and personalization.

[0004] The prior art discloses a program recommendation method and device, wherein the program recommendation method includes: obtaining feedback information from users based on a program recommendation system, wherein the program recommendation system is used to calculate a program recommendation list, and the program recommendation system is constructed based on two or more program recommendation models, wherein the two or more program recommendation models are respectively constructed by different program recommendation algorithms; updating the weights of the two or more program recommendation models when calculating the program recommendation list based on the feedback information; updating the program recommendation list according to the updated weights of the two or more program recommendation models when calculating the program recommendation list; although the application utilizes user feedback information to dynamically adjust the weights of each program recommendation model in the program recommendation system, it does not consider the relationship between user interest preferences and time, and the accuracy and personalization of program recommendations are insufficient. Summary of the invention

[0005] In order to overcome the defect of the above-mentioned prior art that the different interests and preferences of users in different time periods are not taken into account, resulting in inaccurate program recommendations, the present invention provides a method and system for recommending programs by time period for Internet TV, which can provide users with more accurate and personalized TV program recommendations in different time periods.

[0006] In order to solve the above technical problems, the technical solution of the present invention is as follows:

[0007] The present invention provides a method for recommending programs by time period for Internet television, comprising:

[0008] S1: Obtain the user's history of watching TV programs, divide it into time periods, and construct a user behavior sequence;

[0009] S2: Extract features from user behavior sequences to obtain user personalized interest feature vectors, historical program time-related feature vectors, candidate program time-related feature vectors, and candidate program general feature vectors in different time periods;

[0010] S3: Calculate the short-term interest feature vector based on the time-related feature vector of the historical program and the time-related feature vector of the candidate program;

[0011] S4: Calculate the user's long-term preference score and short-term preference score for the candidate program based on the user's personalized interest feature vector, the candidate program's time-related feature vector, the candidate program's general feature vector, and the short-term interest feature vector;

[0012] S5: Calculate the prediction probability of the final candidate program based on the user's long-term preference score and short-term preference score for the candidate program;

[0013] S6: construct a time-segment program recommendation model, construct a recommendation loss function based on the predicted probability of the final candidate program, optimize the recommendation loss function, and obtain an optimized time-segment program recommendation model;

[0014] S7: Obtain the user behavior sequence of the user to be recommended, input it into the optimized time-based program recommendation model, obtain a time-based TV program recommendation list, and implement time-based TV program recommendation.

[0015] Preferably, the specific method of constructing the user behavior sequence is:

[0016] Get the user's TV program viewing history, including user set and TV program set; set a time period, assuming that there are T periods in total, and within a time period, arrange the TV program records watched by each user from small to large time, and divide them into M time periods; record the user set as U = {u1,…, p ,…, |U|},u p represents the pth user, |U| represents the number of users; the TV program set is recorded as I = {item1,…,tem q ,…,tem |I|}, item q represents the qth TV program, |I| represents the number of TV programs; the user behavior sequence is constructed as represents all historical TV programs of the pth user in the tth period, Indicates the TV program watched by the p-th user in the m-th time period in the t-th cycle.

[0017] Preferably, the specific method of step S2 is:

[0018] Map the user behavior sequence from the high-dimensional space to the low-dimensional feature space, and use the following mapping function to extract features:

[0019]

[0020]

[0021]

[0022]

[0023] In the formula, represents the personalized interest feature vector of the user in the mth period of the current cycle, represents the time-related feature vector of the candidate program in the mth period of the current cycle, represents the time-related feature vector of the historical program in the mth period of the previous cycle, Φ I (·) represents the general feature vector of candidate programs in the current cycle; represents a set of real numbers, and d represents the dimension of the low-dimensional feature space.

[0024] Preferably, the specific method for calculating the short-term interest feature vector is:

[0025] Calculate the time period feature vector of the current cycle through the time-related feature vector of the historical program in the previous cycle:

[0026]

[0027] In the formula, x n represents the nth period feature vector of the previous cycle, f agg (*) indicates average pooling operation, represents the time-related feature vector of the jth historical program in the nth period of the previous cycle, represents the TV program of the p-th user in the n-th time period in the previous cycle, n=1,…,M;

[0028] The programs watched by the user in the current period of the current cycle are related to the programs watched in different periods of the previous cycle to varying degrees; the attention weight of each period of the previous cycle relative to any period of the current cycle is calculated:

[0029]

[0030] In the formula, represents the attention weight of the nth time period in the current cycle to the ith candidate program in the mth time period in the previous cycle, exp(*) represents the exponential function, e(*) represents the similarity calculation function, and x k represents the k-th period feature vector, It represents the time-related feature vector representation of the i-th candidate program in the m-th time period of the current cycle, in:

[0031]

[0032] In the formula, Represents x n and The similarity value, W m represents the attention layer weight of the mth period,

[0033] The feature vectors of each time period are combined with the corresponding attention weights to obtain the short-term interest feature vector:

[0034]

[0035] In the formula, Represents the short-term interest feature vector of the p-th user in the m-th time period in the current cycle relative to the ith candidate program.

[0036] Preferably, the specific method for calculating the user's long-term preference score and short-term preference score for the candidate programs is:

[0037] Based on the user's personalized interest feature vector and the general feature vector of the candidate program, the user's long-term preference score for the candidate program is calculated:

[0038]

[0039] In the formula, represents the long-term preference score of the p-th user for the ith candidate program in the m-th time period of the current cycle, represents the personalized interest feature vector of the p-th user in the m-th period of the current cycle, Φ I (i) represents the general feature vector of the i-th candidate program;

[0040] Based on the time-related feature vector and short-term interest feature vector of the candidate programs, the user's short-term preference score for the candidate programs is calculated:

[0041]

[0042] In the formula, It represents the short-term preference score of the p-th user for the ith candidate program in the m-th time period of the current cycle.

[0043] Preferably, the specific method for calculating the predicted probability of the final candidate program is:

[0044]

[0045] In the formula, represents the predicted probability of the final candidate program for the i-th candidate program by the p-th user in the m-th time period of the current cycle, z m It represents the importance score of long-term and short-term preferences in the mth period. Represents all historical TV programs of the p-th user in the t-1th period.

[0046] Preferably, the recommendation loss function is specifically:

[0047]

[0048] In the formula, l represents the recommendation loss function, T represents the total number of time periods; s represents negative samples, that is, samples that do not exist in TV programs in; σ(*) represents the logistic function, λ represents the regularization coefficient, Θ represents the set of trainable parameters, represents the trainable parameters of the user's personalized interest feature vector in the mth period, Represents the trainable parameters of the time-dependent feature vector of the candidate programs in the mth time period, represents the trainable parameters of the time-related feature vector of the historical program in the mth period, Trainable parameters representing the general feature vector of candidate programs.

[0049] Preferably, the specific method of optimizing the recommendation loss function to obtain an optimized time-segment program recommendation model is:

[0050] The recommendation loss function l is optimized iteratively until for each user, each positive candidate program has Obtain the optimized time-segment program recommendation model and obtain the total recommendation function f=<f 1 ,…,f m ,…,f M >, f m Represents the program recommendation probability function for the mth time period.

[0051] Preferably, the specific method of establishing the time-slot TV program recommendation list is:

[0052] The user behavior sequence of the user to be recommended is input into the optimized time-slot program recommendation model, and the recommendation probability of each candidate program in each time slot is calculated using the total recommendation function; for each time slot, all candidate program recommendation probabilities are sorted from large to small, and the top N candidate programs are used as recommended programs for that time slot; the recommended programs in each time slot are summarized to establish a time-slot TV program recommendation list.

[0053] The present invention also provides a system for recommending programs by time period for Internet television, comprising:

[0054] The data acquisition processing module is used to obtain the user's historical records of watching TV programs, divide them into time periods, and construct a user behavior sequence;

[0055] A feature extraction module is used to extract features from user behavior sequences to obtain user personalized interest feature vectors, historical program time-related feature vectors, candidate program time-related feature vectors, and candidate program general feature vectors in different time periods;

[0056] A short-term interest feature calculation module, used to calculate a short-term interest feature vector based on a historical program time-related feature vector and a candidate program time-related vector;

[0057] The long-term and short-term interest program prediction module is used to calculate the user's long-term preference score and short-term preference score for the candidate program based on the user's personalized interest feature vector, the candidate program time-related feature vector, the candidate program general feature vector and the short-term interest feature vector;

[0058] A final program prediction module, used to calculate the prediction probability of the final candidate program based on the user's long-term preference score and short-term preference score for the candidate program;

[0059] The recommendation model training module is used to build a time-segment program recommendation model, build a recommendation loss function based on the predicted probability of the final candidate program, optimize the recommendation loss function, and obtain an optimized time-segment program recommendation model;

[0060] The program time-segment recommendation module is used to obtain the user behavior sequence of the user to be recommended, input it into the optimized time-segment program recommendation model, obtain the time-segment TV program recommendation list, and realize the time-segment recommendation of TV programs.

[0061] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0062] The present invention obtains the historical records of users watching TV programs and divides them into time periods to construct a user behavior sequence; extracts the user's personalized interest feature vector, the historical program time-related feature vector, the candidate program time-related feature vector and the candidate program general feature vector from the user behavior sequence; then calculates the user's short-term interest feature vector in different time periods based on the above feature vectors, and further calculates the user's long-term preference score and short-term preference score for the candidate program; due to the different influences of the user's long-term interest and short-term interest in different time periods, the final candidate program prediction probability is calculated by fusion; finally, a recommendation loss function is constructed based on the final candidate program prediction probability, and the recommendation loss function is optimized to optimize the time-divided program recommendation model. The optimized time-divided program recommendation model is used to obtain a time-divided TV program recommendation list, taking into account the situation that different customers have different interests in different time periods, and realizing more accurate and personalized program recommendations for customers. BRIEF DESCRIPTION OF THE DRAWINGS

[0063] Figure 1 This is a flow chart of a method for recommending programs by time period for Internet television as described in Example 1. Figure 2 This is a schematic diagram of the structure of a system for recommending programs by time period for Internet television as described in Example 3. DETAILED DESCRIPTION

[0064] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0065] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0066] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0067] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0068] Example 1

[0069] This embodiment provides a method for recommending programs by time period for Internet television. Figure 1 As shown, including:

[0070] S1: Obtain the user's history of watching TV programs, divide it into time periods, and construct a user behavior sequence;

[0071] S2: Extract features from user behavior sequences to obtain user personalized interest feature vectors, historical program time-related feature vectors, candidate program time-related feature vectors, and candidate program general feature vectors in different time periods;

[0072] S3: Calculate the short-term interest feature vector based on the time-related feature vector of the historical program and the time-related feature vector of the candidate program;

[0073] S4: Calculate the user's long-term preference score and short-term preference score for the candidate program based on the user's personalized interest feature vector, the candidate program's time-related feature vector, the candidate program's general feature vector, and the short-term interest feature vector;

[0074] S5: Calculate the prediction probability of the final candidate program based on the user's long-term preference score and short-term preference score for the candidate program;

[0075] S6: construct a time-segment program recommendation model, construct a recommendation loss function based on the predicted probability of the final candidate program, optimize the recommendation loss function, and obtain an optimized time-segment program recommendation model;

[0076] S7: Obtain the user behavior sequence of the user to be recommended, input it into the optimized time-based program recommendation model, obtain a time-based TV program recommendation list, and implement time-based TV program recommendation.

[0077] In the specific implementation process, this embodiment obtains the user's historical records of watching TV programs and divides them by time period to construct a user behavior sequence; extracts the user's personalized interest feature vector, the historical program time-related feature vector, the candidate program time-related feature vector and the candidate program general feature vector from the user behavior sequence; the user's personalized interest feature vector is related to the user's own long-term interests, the historical program time-related feature vector and the candidate program time-related feature vector respectively reflect the relationship between the historical program, the candidate program and the time period, and the candidate program general feature vector is the feature of the candidate program itself, which is not related to the time period; then calculate the user's short-term interest feature vector in different time periods, and further calculate the user's long-term preference score and short-term preference score for the candidate program; due to the different degrees of influence of the user's long-term interest and short-term interest in different time periods, the final candidate program prediction probability is calculated by fusion; finally, based on the final candidate program prediction probability, a recommendation loss function is constructed, and the recommendation loss function is optimized to achieve the optimization of the time-divided program recommendation model. The optimized time-divided program recommendation model is used to obtain a time-divided TV program recommendation list, taking into account the situation that different customers have different interests in different time periods, and realizing more accurate and personalized program recommendations for customers.

[0078] Example 2

[0079] This embodiment provides a method for recommending programs by time period for Internet television, including:

[0080] S1: Obtain the user's history of watching TV programs, divide it by time period, and construct a user behavior sequence; the specific method is:

[0081] Get the user's TV program viewing history, including user set and TV program set; set a time period, assuming that there are T periods in total, and within a time period, arrange the TV program records watched by each user from small to large time, and divide them into M time periods; record the user set as U = {u1,…, p ,…, |U|},u p represents the pth user, |U| represents the number of users; the TV program set is recorded as I = {item1,…,tem q ,…,tem |I|}, item q represents the qth TV program, |I| represents the number of TV programs; the user behavior sequence is constructed as represents all historical TV programs of the pth user in the tth period, Indicates the TV program watched by the p-th user in the m-th time period in the t-th cycle.

[0082] In this embodiment, a day is taken as a time period and divided into four time periods, wherein the early morning is 00:00-06:59, the morning is 07:00-11:59, the noon is 12:00-18:59, and the night is 19:00-23:59.

[0083] S2: Extract features from the user behavior sequence to obtain the user's personalized interest feature vectors, historical program time-related feature vectors, candidate program time-related feature vectors, and candidate program general feature vectors in different time periods; the specific method is:

[0084] Map the user behavior sequence from the high-dimensional space to the low-dimensional feature space, and use the following mapping function to extract features:

[0085]

[0086]

[0087]

[0088]

[0089] In the formula, represents the personalized interest feature vector of the user in the mth period of the current cycle, represents the time-related feature vector of the candidate program in the mth period of the current cycle, represents the time-related feature vector of the historical program in the mth period of the previous cycle, Φ I (·) represents the general feature vector of candidate programs in the current period; represents a set of real numbers, and d represents the dimension of the low-dimensional feature space.

[0090] S3: Calculate a short-term interest feature vector based on the time-related feature vectors of the historical programs and the time-related vectors of the candidate programs; the specific method of calculating the short-term interest feature vector is:

[0091] Calculate the time period feature vector through the historical program time-related feature vector:

[0092]

[0093] In the formula, x n represents the feature vector of the nth period of the previous cycle, f agg (*) indicates average pooling operation, represents the time-related feature vector of the jth historical program in the nth period of the previous cycle, represents the TV program of the p-th user in the n-th time period in the t-1-th cycle, n=1,…,M;

[0094] The programs watched by the user in the current period of the current cycle are related to the programs watched in different periods of the previous cycle to varying degrees; the attention weight of each period of the previous cycle relative to any period of the current cycle is calculated:

[0095]

[0096] In the formula, represents the attention weight of the nth time period in the previous cycle to the ith candidate program in the mth time period in the current cycle, exp(*) represents the exponential function, e(*) represents the similarity calculation function, and x k represents the k-th period feature vector, The time-related feature vector representation of the i-th candidate program in the m-th time period of the current cycle is represented by, in:

[0097]

[0098] In the formula, Represents x n and The similarity value, W m represents the attention layer weight of the mth period,

[0099] The feature vectors of each time period are combined with the corresponding attention weights to obtain the short-term interest feature vector:

[0100]

[0101] In the formula, Represents the short-term interest feature vector of the p-th user in the m-th time period in the current cycle relative to the ith candidate program.

[0102] S4: Calculate the user's long-term preference score and short-term preference score for the candidate program based on the user's personalized interest feature vector, the candidate program's time-related feature vector, the candidate program's general feature vector, and the short-term interest feature vector; specifically:

[0103] Based on the user's personalized interest feature vector and the general feature vector of the candidate program, the user's long-term preference score for the candidate program is calculated:

[0104]

[0105] In the formula, represents the long-term preference score of the p-th user for the ith candidate program in the m-th time period of the current cycle, represents the personalized interest feature vector of the p-th user in the m-th period of the current cycle, Φ I (i) represents the general feature vector of the i-th candidate program;

[0106] Based on the time-related feature vector and short-term interest feature vector of the candidate programs, the user's short-term preference score for the candidate programs is calculated:

[0107]

[0108] In the formula, It represents the short-term preference score of the p-th user for the ith candidate program in the m-th time period of the current cycle.

[0109] S5: Calculate the prediction probability of the final candidate program based on the user's long-term preference score and short-term preference score for the candidate program; specifically:

[0110]

[0111] In the formula, represents the predicted probability of the final candidate program for the i-th candidate program by the p-th user in the m-th time period of the current cycle, z m It represents the importance score of long-term and short-term preferences in the mth period. Represents all historical TV programs of the p-th user in the t-1th period.

[0112] S6: construct a time-segment program recommendation model, construct a recommendation loss function based on the predicted probability of the final candidate program, optimize the recommendation loss function, and obtain an optimized time-segment program recommendation model;

[0113] The recommended loss function is specifically:

[0114]

[0115] In the formula, l represents the recommendation loss function, T represents the total number of time periods; s represents negative samples, that is, samples that do not exist in TV programs in; σ(*) represents the logistic function, λ represents the regularization coefficient, Θ represents the set of trainable parameters, represents the trainable parameters of the user's personalized interest feature vector in the mth period, represents the trainable parameters of the time-related feature vector of the candidate program in the mth time period, represents the trainable parameters of the time-related feature vector of the historical program in the mth period, Trainable parameters representing a general feature vector of candidate programs;

[0116] The recommendation loss function l is optimized iteratively until for each user, each positive candidate program has Obtain the optimized time-segment program recommendation model and obtain the total recommendation function f=〈f 1 ,…,f m ,…,f M >, f m Represents the program recommendation probability function for the mth time period.

[0117] S7: Obtain the user behavior sequence of the user to be recommended, input the optimized time-based program recommendation model, obtain a time-based TV program recommendation list, and implement time-based TV program recommendation.

[0118] The specific method of establishing the time-divided TV program recommendation list is as follows:

[0119] The user behavior sequence of the user to be recommended is input into the optimized time-slot program recommendation model, and the recommendation probability of each candidate program in each time slot is calculated using the total recommendation function; for each time slot, all candidate program recommendation probabilities are sorted from large to small, and the top N candidate programs are used as recommended programs for that time slot; the recommended programs for each time slot are summarized to establish a time-slot TV program recommendation list.

[0120] Example 3

[0121] The present invention also provides a system for recommending programs by time period for Internet television. Figure 2 As shown, including:

[0122] The data acquisition processing module is used to obtain the user's historical records of watching TV programs, divide them into time periods, and construct a user behavior sequence;

[0123] A feature extraction module is used to extract features from user behavior sequences to obtain user personalized interest feature vectors, historical program time-related feature vectors, candidate program time-related feature vectors, and candidate program general feature vectors in different time periods;

[0124] A short-term interest feature calculation module, used to calculate a short-term interest feature vector based on a historical program time-related feature vector and a candidate program time-related vector;

[0125] The long-term and short-term interest program prediction module is used to calculate the user's long-term preference score and short-term preference score for the candidate program based on the user's personalized interest feature vector, the candidate program time-related feature vector, the candidate program general feature vector and the short-term interest feature vector;

[0126] A final program prediction module, used to calculate the prediction probability of the final candidate program based on the user's long-term preference score and short-term preference score for the candidate program;

[0127] The recommendation model training module is used to build a time-segment program recommendation model, build a recommendation loss function based on the predicted probability of the final candidate program, optimize the recommendation loss function, and obtain an optimized time-segment program recommendation model;

[0128] The program time-segment recommendation module is used to obtain the user behavior sequence of the user to be recommended, input the optimized time-segment program recommendation model, obtain the time-segment TV program recommendation list, and realize the time-segment recommendation of TV programs.

[0129] The same or similar reference numerals correspond to the same or similar components;

[0130] The terms used in the drawings to describe positional relationships are only used for illustrative purposes and should not be construed as limiting this patent;

[0131] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A method for recommending programs by time period for Internet television, characterized in that: include: S1: Obtain the user's history of watching TV programs, divide it into time periods, and construct a user behavior sequence; The specific method of constructing the user behavior sequence is: Get the user's TV program viewing history, including user set and TV program set; set a time period, assuming that there are T periods in total, and within a time period, arrange the TV program records watched by each user from small to large in time, and divide them into M time periods; record the user set as U = {u1,…,u p ,…,u |U| },u p represents the pth user, |U| represents the number of users; the TV program set is recorded as I = {item1,…,item q ,…,item |I| }, item q represents the qth TV program, |I| represents the number of TV programs; the user behavior sequence is constructed as represents all historical TV programs of the pth user in the tth period, represents the TV program watched by the p-th user in the m-th time period in the t-th cycle; S2: Extract features from the user behavior sequence to obtain the user personalized interest feature vectors, historical program time-related feature vectors, candidate program time-related feature vectors, and candidate program general feature vectors in different time periods. The specific method is as follows: Map the user behavior sequence from the high-dimensional space to the low-dimensional feature space, and use the following mapping function to extract features: In the formula, represents the personalized interest feature vector of the user in the mth period of the current cycle, represents the time-related feature vector of the candidate program in the mth period of the current cycle, represents the time-related feature vector of the historical program in the mth period of the previous cycle, Φ I A general feature vector representing candidate programs of the current cycle; represents a set of real numbers, and d represents the dimension of the low-dimensional feature space; S3: Based on the time-related feature vectors of historical programs and the time-related feature vectors of candidate programs, a short-term interest feature vector is calculated. The specific method is as follows: Calculate the time period feature vector of the current cycle through the time-related feature vector of the historical program in the previous cycle: In the formula, x n represents the nth period feature vector of the previous cycle, f agg (*) indicates average pooling operation, represents the time-related feature vector of the jth historical program in the nth period of the previous cycle, represents the TV program of the p-th user in the n-th time period in the previous cycle, n=1,…,M; The programs watched by the user in the current period of the current cycle are related to the programs watched in different periods of the previous cycle to varying degrees; Calculate the attention weight of each period in the previous cycle relative to any period in the current cycle: In the formula, represents the attention weight of the nth time period in the current cycle to the ith candidate program in the mth time period in the previous cycle, exp(*) represents the exponential function, e(*) represents the similarity calculation function, and x k represents the k-th period feature vector, It represents the time-related feature vector representation of the i-th candidate program in the m-th time period of the current cycle, in: In the formula, Represents x n and The similarity value, W m represents the attention layer weight of the mth period, The feature vectors of each time period are combined with the corresponding attention weights to obtain the short-term interest feature vector: In the formula, represents the short-term interest feature vector of the p-th user in the m-th time period in the current cycle relative to the i-th candidate program; S4: Calculate the user's long-term preference score and short-term preference score for the candidate program based on the user's personalized interest feature vector, the candidate program's time-related feature vector, the candidate program's general feature vector, and the short-term interest feature vector. The specific method is as follows: Based on the user's personalized interest feature vector and the general feature vector of the candidate program, the user's long-term preference score for the candidate program is calculated: In the formula, represents the long-term preference score of the p-th user for the ith candidate program in the m-th time period of the current cycle, represents the personalized interest feature vector of the p-th user in the m-th period of the current cycle, Φ I (i) represents the general feature vector of the i-th candidate program; Based on the time-related feature vector and short-term interest feature vector of the candidate programs, the user's short-term preference score for the candidate programs is calculated: In the formula, represents the short-term preference score of the p-th user for the ith candidate program in the m-th time period of the current cycle; S5: Based on the user's long-term preference score and short-term preference score for the candidate programs, the prediction probability of the final candidate programs is calculated. The specific method is: In the formula, represents the predicted probability of the final candidate program for the i-th candidate program by the p-th user in the m-th time period of the current cycle, z m It represents the importance score of long-term and short-term preference in the mth period, represents all historical TV programs of the p-th user in the t-1th period; S6: construct a time-segment program recommendation model, construct a recommendation loss function based on the predicted probability of the final candidate programs, optimize the recommendation loss function, and obtain an optimized time-segment program recommendation model; S7: Obtain the user behavior sequence of the user to be recommended, input it into the optimized time-based program recommendation model, obtain a time-based TV program recommendation list, and implement time-based TV program recommendation.

2. The method for recommending programs by time period for Internet TV according to claim 1, characterized in that: The recommended loss function is specifically: In the formula, l represents the recommendation loss function, T represents the total number of time periods; s represents negative samples, that is, samples that do not exist in TV programs in; σ(*) represents the logistic function, λ represents the regularization coefficient, Θ represents the set of trainable parameters, represents the trainable parameters of the user's personalized interest feature vector in the mth period, Represents the trainable parameters of the time-dependent feature vector of the candidate programs in the mth time period, represents the trainable parameters of the time-related feature vector of the historical program in the mth period, Trainable parameters representing the general feature vector of candidate programs.

3. The method for recommending programs by time period for Internet TV according to claim 2, characterized in that: The specific method of optimizing the recommendation loss function to obtain an optimized time segment program recommendation model is as follows: The recommendation loss function l is optimized iteratively until for each user, each positive candidate program has Obtain the optimized time-segment program recommendation model and obtain the total recommendation function f= <f 1 ,…,f m ,…,f M >, f m Represents the program recommendation probability function for the mth time period.

4. The method for recommending programs by time period for Internet TV according to claim 3, characterized in that: The specific method of establishing a recommended list of TV programs by time period is as follows: Input the user behavior sequence of the user to be recommended into the optimized time-segment program recommendation model, and use the total recommendation function to calculate the recommendation probability of each candidate program in each time period; For each time period, all candidate programs are ranked from high to low in terms of recommendation probability, and the top N candidate programs are used as recommended programs for that time period; Summarize the recommended programs for each time period and create a recommended list of TV programs for each time period.

5. A system for recommending programs by time period for Internet television, characterized in that: include: The data acquisition processing module is used to obtain the user's historical records of watching TV programs, divide them into time periods, and construct a user behavior sequence; The specific method of constructing the user behavior sequence is: Get the user's TV program viewing history, including user set and TV program set; set a time period, assuming that there are T periods in total, and within a time period, arrange the TV program records watched by each user from small to large in time, and divide them into M time periods; record the user set as U = {u1,…,u p ,…,u |U| },u p represents the pth user, |U| represents the number of users; the TV program set is recorded as I = {item1,…,item q ,…,item |I| }, item q represents the qth TV program, |I| represents the number of TV programs; the user behavior sequence is constructed as represents all historical TV programs of the pth user in the tth period, represents the TV program watched by the p-th user in the m-th time period in the t-th cycle; The feature extraction module is used to extract features from the user behavior sequence to obtain the user's personalized interest feature vectors, historical program time-related feature vectors, candidate program time-related feature vectors and candidate program general feature vectors in different time periods. The specific method is as follows: Map the user behavior sequence from the high-dimensional space to the low-dimensional feature space, and use the following mapping function to extract features: In the formula, represents the personalized interest feature vector of the user in the mth period of the current cycle, represents the time-related feature vector of the candidate program in the mth period of the current cycle, represents the time-related feature vector of the historical program in the mth period of the previous cycle, Φ I A general feature vector representing candidate programs of the current cycle; represents a set of real numbers, and d represents the dimension of the low-dimensional feature space; The short-term interest feature calculation module is used to calculate the short-term interest feature vector based on the time-related feature vectors of the historical programs and the time-related feature vectors of the candidate programs. The specific method is as follows: Calculate the time period feature vector of the current cycle through the time-related feature vector of the historical program in the previous cycle: In the formula, x n represents the nth period feature vector of the previous cycle, f agg (*) indicates average pooling operation, represents the time-related feature vector of the jth historical program in the nth period of the previous cycle, represents the TV program of the p-th user in the n-th time period in the previous cycle, n=1,…,M; The programs watched by the user in the current period of the current cycle are related to the programs watched in different periods of the previous cycle to varying degrees; Calculate the attention weight of each period in the previous cycle relative to any period in the current cycle: In the formula, represents the attention weight of the nth time period in the current cycle to the ith candidate program in the mth time period in the previous cycle, exp(*) represents the exponential function, e(*) represents the similarity calculation function, and x k represents the k-th period feature vector, It represents the time-related feature vector representation of the i-th candidate program in the m-th time period of the current cycle, in: In the formula, Represents x n and The similarity value, W m represents the attention layer weight of the mth period, The feature vectors of each time period are combined with the corresponding attention weights to obtain the short-term interest feature vector: In the formula, represents the short-term interest feature vector of the p-th user in the m-th time period in the current cycle relative to the i-th candidate program; The long-term and short-term interest program prediction module is used to calculate the user's long-term preference score and short-term preference score for the candidate program based on the user's personalized interest feature vector, the candidate program time-related feature vector, the candidate program general feature vector and the short-term interest feature vector. The specific method is as follows: Based on the user's personalized interest feature vector and the general feature vector of the candidate program, the user's long-term preference score for the candidate program is calculated: In the formula, represents the long-term preference score of the p-th user for the ith candidate program in the m-th time period of the current cycle, represents the personalized interest feature vector of the p-th user in the m-th period of the current cycle, Φ I (i) represents the general feature vector of the i-th candidate program; Based on the time-related feature vector and short-term interest feature vector of the candidate programs, the user's short-term preference score for the candidate programs is calculated: In the formula, represents the short-term preference score of the p-th user for the ith candidate program in the m-th time period of the current cycle; The final program prediction module is used to calculate the prediction probability of the final candidate program based on the user's long-term preference score and short-term preference score for the candidate program. The specific method is as follows: In the formula, represents the predicted probability of the final candidate program for the i-th candidate program by the p-th user in the m-th time period of the current cycle, z m It represents the importance score of long-term and short-term preference in the mth period, represents all historical TV programs of the p-th user in the t-1th period; The recommendation model training module is used to build a time-segment program recommendation model, build a recommendation loss function based on the predicted probability of the final candidate program, optimize the recommendation loss function, and obtain an optimized time-segment program recommendation model; The program time-segment recommendation module is used to obtain the user behavior sequence of the user to be recommended, input the optimized time-segment program recommendation model, obtain the time-segment TV program recommendation list, and realize the time-segment recommendation of TV programs.

Images