A method for realizing recommendation related to programs on IPTV

Abstract

The application relates to a method for realizing relevant recommendations of programs on IPTV, which improves program-based collaborative filtering similarity, and then fuses the improved collaborative filtering similarity with content attribute similarity. The improved program-based collaborative filtering similarity algorithm is adopted, and the improved collaborative filtering similarity algorithm considers the influence of program watching time difference, watching time period and working day or not. Meanwhile, the improved collaborative filtering similarity is linearly fused with the content attribute similarity. The cold start problem of newly online programs is solved, and the diversification problem of IPTV family members and the problem that user interest changes with time are considered.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of IPTV, and particularly relates to a method for realizing related recommendation of programs on IPTV. BACKGROUND

[0002] Intelligent recommendation can help users find the most likely interesting data in a large amount of data and recommend the data to the users, and has been widely applied in the fields of e-commerce product recommendation, news recommendation, music video recommendation and the like.

[0003] As a large-screen video watching device, IPTV is different from a mobile terminal APP, and the APP intelligent recommendation is committed to providing video recommendation services for a single user, while the IPTV intelligent recommendation is committed to providing comprehensive and personalized large-screen video recommendation services for all family members as a unit. The members of a family are complex and diversified, and the user behaviors generated simultaneously are also diversified.

[0004] Related recommendation is a common recommendation method in intelligent recommendation, which recommends similar programs based on the programs on which the user behaviors occur. At present, there are many calculation methods of similarity, such as collaborative filtering, content attribute similarity and knowledge graph. As one of the most successful technologies applied in the recommendation system, how to more efficiently apply the item-based collaborative filtering algorithm to the intelligent recommendation of IPTV will have a very important influence on the final recommendation effect.

[0005] At present, the item-based collaborative filtering algorithm and the content attribute similarity have been widely applied in intelligent recommendation. In the early stage, these algorithms are respectively applied in the related recommendation of IPTV, and through A / B test comparison, the item-based collaborative filtering algorithm has obtained better effect than the content attribute similarity algorithm.

[0006] However, the item-based collaborative filtering algorithm still ignores the following problems: first, the members of an IPTV family are diversified; second, the user interest is constantly changing over time; and third, the cold start problem of new online programs. This makes the recommendation result in IPTV inaccurate, and the recommended content more and more monotonous over time, and the recommendation effect poor. SUMMARY

[0007] In order to overcome the above-mentioned defects, the purpose of the present application is to provide a method for realizing related recommendation of programs on IPTV, solve the existing problems, increase the accuracy, diversity and novelty of the recommendation result, and make the related recommendation column obtain better recommendation effect.

[0008] The technical scheme adopted by the present application to solve its technical problems is: a method for realizing relevant recommendation of programs on IPTV, which improves the program-based collaborative filtering similarity, then fuses the improved collaborative filtering similarity with the content attribute similarity, and the specific scheme is as follows:

[0009] S1, obtaining user behavior records;

[0010] S2, determining a time period when the user behavior occurs;

[0011] S3, determining whether it is a weekday when the user behavior occurs;

[0012] S4, constructing a historical viewing sequence of the user;

[0013] S5, calculating the collaborative filtering similarity between programs, including:

[0014] The program-based collaborative filtering similarity algorithm is improved, which mainly solves the complexity of the IPTV family members and the problem that the user interest changes over time, and the similarity calculation method before improvement is as follows:

[0015]

[0016] Wherein, |N(i)| and |N(j)| respectively represent the number of users who like program i and program j, and |N(i)∩N(j)| represents the number of users who like both program i and program j;

[0017] The improved similarity calculation method is as follows:

[0018]

[0019] The algorithm introduces date-related items, time period-related items, and weekday-related items, and requires that alpha, beta, and gamma are all values greater than 0;

[0020] |d ui -d uj |*alpha is a date-related item related to the date of watching the program, wherein d ui is the date when user u generates behavior on program i, d uj is the date when user u generates behavior on program j, alpha represents a date-related parameter, the farther the time when user u generates behavior on program i and program j, the greater the time attenuation item, and the lower the similarity; on the contrary, the higher the similarity.

[0021] g(ts ui , ts uj )*beta is a time period-related item, wherein ts ui represents the time period when user u generates behavior on program i, ts ujdenotes the time period in which user u generates behavior on program j, and β denotes the time period related parameter, and the specific calculation formula is as follows:

[0022]

[0023] Here, we divide the twenty-four hours of each day into four time intervals (23-05), (06-11), (12-16), (17-22) according to the hour number, and if the time in which user u generates behavior on program i and program j is in the same time period, it is considered that the family members watching program i and the family members watching program j are more likely to belong to the same person, the time period related item is 0, and the similarity of program i and program j is higher; otherwise, if the time is in different time periods, it is considered that the two behaviors are likely to come from two different family members, and the time relatedness is 1*β, and the similarity of program i and program j is lower.

[0024] h(w ui , w uj )*γ is a weekday related item, wherein w ui is whether the date in which user u generates behavior on program i is a weekday, w uj is whether the date in which user u generates behavior on program j is a weekday, and γ denotes the weekday related parameter, and the specific calculation formula is as follows:

[0025]

[0026] Here, the national statutory workday calendar is pre-imported into the database, and if the dates in which user u generates behavior on program i and program j are both weekdays (or non-weekdays), it is considered that the family members watching program j and the family members watching program i are more likely to belong to the same person, and the weekday related item is considered to be 0, and the similarity of program i and program j is higher; otherwise, if one of the behaviors is on a weekday and the other is on a non-weekday, it is considered that the two behaviors are likely to come from two different family members, and the weekday relatedness is 1*γ, and the similarity of program i and program j is lower.

[0027] S6, obtaining the content similarity between programs;

[0028] S7, linearly fusing the improved collaborative filtering similarity and the content attribute similarity, and the addition of the content similarity solves the cold start problem and the data sparsity problem of the collaborative filtering; and the addition of the collaborative filtering solves the problem of poor diversity and novelty of the content similarity.

[0029] The similarity calculation formula after fusion is as follows:

[0030] Sim ij =ItemCF ij+ω*Content ij

[0031] wherein ItemCF ij represents improved collaborative filtering similarity, Content ij represents content attribute similarity, and ω is a linear parameter.

[0032] S8, selecting the most similar TOPN programs for each program as the result of relevant recommendation, and recommending to the user.

[0033] The present application has the following beneficial effects: improving the program-based collaborative filtering similarity algorithm, the improved collaborative filtering similarity algorithm taking into account the influence of program viewing time difference, viewing time period, and viewing whether it is a weekday; at the same time, linearly fusing the improved collaborative filtering similarity and the content attribute similarity; solving the cold start problem of newly online programs, and taking into account the diversification problem of IPTV family members and the problem that user interest changes over time. BRIEF DESCRIPTION OF DRAWINGS

[0034] Figure 1 is a flowchart of the present application. DETAILED DESCRIPTION

[0035] The present application will now be further described in detail in conjunction with the accompanying drawings.

[0036] According to Figure 1 The specific steps are as follows:

[0037] S1, obtaining user behavior records, obtaining user viewing behavior records from a data warehouse using spark, obtaining user viewing date and time;

[0038] S2, determining the time period in which the user behavior occurs, according to the time period division standard: (23-05) is ts1, (06-11) is ts2, (12-16) is ts3, and (17-22) is TS4, determining the time period in which the user behavior occurs;

[0039] S3, determining whether it is a weekday when the user behavior occurs, according to a work calendar and the viewing date, determining whether it is a weekday when the user behavior occurs;

[0040] S4, constructing the user's historical viewing sequence;

[0041] S5, calculating the collaborative filtering similarity between programs, including:

[0042] S51, constructing a co-occurrence matrix based on the user's historical viewing sequence, and calculating the co-occurrence score using the formula for each user:

[0043]

[0044] The algorithm introduces date-related terms, time period-related terms, and workday-related terms, and requires that α, β, and γ are all greater than 0;

[0045] |d ui -d uj |*α is a date-related term related to the viewing date of a program, where d ui is the date on which user u performed an action on program i, d uj is the date on which user u performed an action on program j, and α represents a date-related parameter. The farther apart the time at which user u performed an action on program i and program j, the greater the time decay term and the lower the similarity; conversely, the higher the similarity.

[0046] g(ts ui , ts uj )*β is a time period-related term, where ts ui represents the time period in which user u performed an action on program i, ts uj represents the time period in which user u performed an action on program j, and β represents a time period-related parameter, with the specific calculation formula being:

[0047]

[0048] Here, we divide the twenty-four hours of each day into four time intervals (23-05), (06-11), (12-16), and (17-22) according to the number of hours. If the time at which user u performed an action on program i and program j is within the same time period, then we consider that the family members who watched program i and program j are more likely to be the same person, the time period-related term is 0, and the similarity between program i and program j is higher; conversely, if the times are within different time periods, then we consider that the two actions are likely to come from two different family members, in which case the time-related term is 1*β, and the similarity between program i and program j is lower.

[0049] h(w ui , w uj )*γ is a workday-related term, where w ui is whether the date on which user u performed an action on program i is a workday, w uj is whether the date on which user u performed an action on program j is a workday, and γ represents a workday-related parameter, with the specific calculation formula being:

[0050]

[0051] Here, we will import the national statutory working day calendar in advance into the database, if the date of the behavior of user u to program i and program j is in working day (or non-working day), it is considered that the family members watching program j and the family members watching program i are more likely to belong to the same person, and the working day correlation term is 0, the higher the similarity of program i and program j; on the contrary, if one of the behaviors is working day and the other is non-working day, it is considered that the two behaviors may come from two different family members, at this time, the working day correlation is 1*γ, and the similarity of program i and program j is lower.

[0052] S52, calculate the total co-occurrence score of program i and program j, the calculation formula is:

[0053]

[0054] S53, count the number of watching users of each program |N(i)|, |N(j)|;

[0055] S54, summarize the above formula to calculate the collaborative filtering similarity between programs, using the formula:

[0056] The calculation is carried out.

[0057] S6, obtain the content attribute similarity between programs;

[0058] S7, linearly fuse the content attribute similarity and the collaborative filtering similarity, the calculation formula is:

[0059] Sim ij =ItemCF sim→ij +ω*Content sim→ij ;

[0060] S8, select the most similar TOPN program for each program as the result of related recommendation, and recommend to the user.

[0061] The application is not limited to the described embodiments, and anyone should know that structural changes made under the inspiration of the application, any technical solution with the same or similar to the application, falls within the protection scope of the application.

[0062] The technical, shape and structure parts not described in detail in the application are well-known technologies.

Claims

1. A method for IPTV program implementation related recommendation, characterized in that: S1, obtaining user behavior records; S2, determining the time period when the user behavior occurs; S3, determining whether it is a weekday when the user behavior occurs; S4, constructing the user's historical viewing sequence; S5, calculating the collaborative filtering similarity between programs, including S51, the calculation method of the calculation of the collaborative filtering similarity between programs is: Wherein: The algorithm introduces date related items, time period related items, and weekday related items, and requires that α, β, γ are all values greater than 0; |d ui -d uj |*α is a date dependent term related to the date of program viewing, where d ui is the date on which user u produced an action on program i, d uj is the date on which user u produced an action on program j, and a denotes a date dependent parameter; g(t sui , t suj )*β represents the time-related term, where ts ui ts represents the time period during which user u interacts with program i. uj This represents the time period during which user u interacts with program j, and β represents parameters related to this time period. The specific calculation formula is as follows: h(w ui , w uj )*γ is the weekday-related term, where w ui is whether the date when user u generates behavior on program i is a weekday, w uj is whether the date when user u generates behavior on program j is a weekday, and γ represents the weekday-related parameter, and the specific calculation formula is: S52, calculating the total co-occurrence score of program i and program j, the calculation formula is: S53, counting the number of viewing users of each program ∣N(i)∣, ∣N(j)∣; S54, summarizing the above formula, calculating the collaborative filtering similarity between programs, using the formula: performing the calculation; S6, obtaining the content similarity between programs, the similarity calculation formula after fusion is as follows: Sim ij = ItemCF ij + ω * Content ij where ItemCF ij denotes the improved collaborative filtering similarity, Content ij denotes the content attribute similarity, and ω is a linear parameter. S7, linearly fusing the improved collaborative filtering similarity and the content attribute similarity; S8, selecting the most similar TOPN programs for each program as the result of related recommendation and recommending to the user.

Images