CN101287082A - A Collaborative Filtering Recommendation Method Introducing Program Popularity Weight - Google Patents

Abstract

The invention discloses a collaborative filtering recommendation method that introduces the weight of program popularity, which is characterized in that on the interface of IPTV programs, a visual menu of user ratings is provided, and according to the user's viewing time, behavior operations, and program rating data transmitted from the terminal set-top box Make a program recommendation list to the target user, the specific steps include: collect user behavior characteristic information; make a "user-item" scoring matrix A(m, n); calculate popularity weight value; calculate similarity and sort; target user Perform prediction scoring and sorting; make a recommendation list to target users. Compared with the prior art, the present invention is more in line with the objective reality, improves the quality of collaborative filtering, and makes the recommendation more accurate. It actively cuts the programs according to the user's preference and behavior characteristics, and makes personalized recommendations for the programs that the user wants to watch. , to achieve the goal of "watching the TV you want when you want".

Description

A Collaborative Filtering Recommendation Method Introducing Program Popularity Weight

technical field

The invention relates to an IPTV personalized recommendation system, in particular to a collaborative filtering recommendation method that introduces program popularity weights.

Background technique

With the rapid increase of information on the Internet, the phenomenon of so-called "information overload" and "information obsession" appeared, and the recommendation system came into being. It can find resources suitable for users' interests according to the user's operation history and feedback information. It generates personalized recommendations. Today, recommendation technology has been applied in various fields such as e-commerce, digital library, film and television entertainment, etc. Especially in the field of IPTV, with the continuous development of digital TV and communication technology, TV program resources are becoming more and more abundant. On the one hand, users are excited to watch so many programs; It is very distressing to find the show they really like among thousands of shows. Collaborative filtering technology is currently the most successful personalized recommendation technology. Some well-known recommendation systems such as WebWatcher, GroupLens, Firefly, SELECT, LileMinds and Citeseer have adopted collaborative filtering methods. The basic idea is to generate recommendations to target users based on the scoring data of the nearest neighbors with similar scores, that is, to generate a recommendation list for target users based on the views of other users. It is based on the assumption that if users rate some items similarly, they will rate other items similarly. The starting point is to find a group of users with the same interests as you. The term is called "nearest neighbor". The core of nearest neighbor search is to calculate the similarity between two users. For example, user A and user B first need to obtain all the scoring items of user A and user B, and then select an appropriate similarity calculation method to calculate the similarity value of user A and user B based on the scoring item data. The currently used similarity algorithms include Pearson correlation coefficient (PCC), cosine similarity and adjusted cosine similarity. From the above, it can be seen that the key step of collaborative filtering is to find the nearest neighbors of the target user. Whether the accurate nearest neighbors can be found is the key to the accuracy of the recommendation. A more accurate calculation of the similarity between users is the prerequisite for accurate selection of the nearest neighbors. However, in the currently used PCC calculation, it treats the items that are jointly rated by users equally, and does not distinguish the popularity of the items themselves. The accuracy of the personalized recommendation technology is poor, and it is not consistent with the objective reality.

Contents of the invention

The purpose of the present invention is a kind of collaborative filtering recommendation method that introduces program popularity weight designed for the deficiencies in the prior art, it first defines the popularity of TV programs, and then calculates its popularity weight, and in user similarity calculation formula This weight is introduced in , and the calculated similarity is more in line with the actual situation, and the calculated similarity is more accurate, so the nearest neighbor of the target user can be selected more accurately, thereby generating more accurate recommendations.

The purpose of the present invention is achieved in this way: a collaborative filtering recommendation method that introduces program popularity weights is characterized in that on the interface of IPTV programs, a visual menu of user ratings is provided, and according to the user's viewing time and behavior transmitted from the terminal set-top box Operation and program scoring data to make a program recommendation list to target users, the specific steps are as follows:

a. Collect user interest data and make a "user-item" scoring matrix A(m, n);

b. Calculate the popularity weight value of the project in an offline cycle;

c. Find the corresponding popularity weight value for the item rated by the current active user;

d. Make and sort the similarity between the target user a and other users;

e. Select the K users with the highest similarity as their nearest neighbor set;

f. Predict and sort the unrated items of the target user according to the nearest neighbor set;

g. Make a recommendation list of the top N items with the largest predicted ratings to the target user.

The "user-item" rating matrix A(m, n) is arranged in a matrix based on user rating information and user behavior data, rows represent users, columns represent items, and the element values in the matrix represent the rows of users for the column items. degree of liking.

The offline period is calculated once every 30 minutes, and the popularity weight value of the project is based on $w_t=\log \left(\frac{P_{\mathrm{all}}}{P_t}\right)$ For calculation, w _t is the popularity weight; P _all is the sum of popularity of all items; P _t is the popularity of t item;

The similarity between the target user a and other users is calculated by introducing the popularity weight value into the Pearson correlation coefficient.

Compared with the prior art, the present invention is more in line with the objective reality, improves the quality of collaborative filtering, and makes the recommendation more accurate. It actively cuts the programs according to the user's preference and behavior characteristics, and makes personalized recommendations for the programs that the user wants to watch. , to achieve the goal of "watching the TV you want when you want".

Description of drawings

Fig. 1 is a schematic flow chart of the present invention

Fig. 2 is a schematic flow chart of calculating the popularity weight of the project t of the present invention

Detailed ways

Example

Referring to accompanying drawings 1-2, the present invention provides a visual menu of user ratings on the interface of IPTV programs, and makes a program recommendation list to the target user according to the user's viewing time, behavior operation, and program rating data transmitted from the terminal set-top box. Proceed as follows:

1. In the IPTV system, the data collection component acquires information representing user interests by tracking features such as user viewing time and behavioral operations, and stores them in corresponding database tables.

2. The above-mentioned user behavior characteristic information is processed by the system as raw data, and the evaluation is completed on behalf of the user, and then according to the user rating information and user behavior data, the "user-item" scoring matrix A(m, n) is obtained, and the scoring The value of is from 1 to r _max (that is, the scoring range is 1-5). This matrix is stored on the recommendation engine device as a user interest model. The rows represent users, and the columns represent items. The preference level of the item is set to 5 levels, corresponding to: (1) dislike very much, (2) dislike somewhat, (3) general, (4) like somewhat, (5) like very much. If the user has not rated an item, then it is set to 0 in the rating matrix.

3. Calculate the "user-item" scoring matrix A(m, n) in an offline cycle, obtain the popularity weight of each item and store it, and calculate the offline cycle once every 30 minutes, (it can also be determined according to the frequency of user updates ) Among them, the popularity P _t of item t is defined as: the number of times item t is scored, that is, the number of non-zero items in the tth column in the user-item rating matrix, P _t = |U(t)|, it can be seen that The more times it is rated, the more popular the item is, and the less it is rated, the less popular the item is. Its popularity weight w _t is defined as: $w_t=\log \left(\frac{P_{\mathrm{all}}}{P_t}\right),$ Where P _all is the sum of the popularity of all items, that is, the total number of times all items are rated, which can be obtained from the number of all non-zero items in the user-item rating matrix. According to the data in the matrix A(m, n), use the offline cycle to calculate the popularity weight value of each item and record it in W(n), which reflects the role of the corresponding item in calculating the user similarity . For example, user 1 and user 2 both rated the movies "Yangtze River No. 7" and "My Name is Liu Yuejin", but by scanning the rating matrix, we found that "Yangtze River No. "Liu Yuejin" has been rated 40 times, assuming that all ratings on the movie collection are 3000, then the popularity weight w of _{"Yangtze River No. 7"} = log(3000/200) = 2.7, "I My name is Liu Yuejin" popularity weight w _{"My name is Liu Yuejin"} = log(3000/40) = 4.3, so that "My name is Liu Yuejin" plays a greater role in measuring the similarity between user 1 and user 2 than "Yangtze River No. 7".

4. When the target user a arrives, scan the scoring matrix A(m, n) to obtain a set of rated items T _a , for each item t∈T _a , find the corresponding w _t in W(n). The recommendation engine uses the similarity calculation between the target user a and other users u to form the user similarity matrix Sim(m, m) according to the "user-item" scoring matrix A(m, n) and the popularity weight. In the similarity calculation The popularity weight of common scoring items is introduced, and the calculation method is as follows: Based on the Pearson correlation coefficient method, the popularity weight of each item in the public scoring items is added to the numerator, in order to limit the similarity between -1 and 1, The maximum value of the popularity weight of the public scoring item is added to the denominator as a divisor, and the formula is as follows:

$\mathrm{<span\; class="notranslate">\; sim</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ^</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&Center\; Dot;</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}}<span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; l</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&le;</span>\mathrm{<span\; class="notranslate">\; no</span>}}{\mathrm{<span\; class="notranslate">\; Max</span>}}<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; )</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ^</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}<span\; class="notranslate">\; \&Center\; Dot;</span>\sqrt{\underset{\mathrm{<span\; class="notranslate">\; t</span>}<span\; class="notranslate">\; \&Element;</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; ^</span>\mathrm{<span\; class="notranslate">\; T</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; )</span>}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}{<span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; u</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; t</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}}<span\; class="notranslate">\; )</span>}^{\mathrm{<span\; class="notranslate">\; 2</span>}}}}$

It can be seen from the calculation formula that when the popularity P _t of the item t is greater, the popularity weight w _t is smaller, and this weight is added to the similarity calculation formula, and the resulting similarity is smaller. Conversely, when the popularity P _t of t is smaller, the obtained similarity is larger, so it conforms to this common sense. Watching movies with higher popularity together will reflect lower user similarity. Conversely, watching movies with lower popularity together will reflect higher user similarity. Sort the above calculation results from high to low.

5. According to the above-mentioned similarity between the target user a and other users and sort them, find the top k nearest neighbors most similar to the target user a, and form the nearest neighbor set [knn ₁ , knn ₂ , L, knn _k ], such that sim(a, knn ₁ )>sim(a, knn ₂ )>L>sim(a, knn _k ).

6. Scan A(m, n), find the set of unrated items T′ _a of user a, and predict the rating of user a on item j for each unrated item j of active user a, and use the following formula for each t ∈T′ _a calculates the predicted score value;

${\mathrm{<span\; class="notranslate">\; P</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; =</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; a</span>}}}<span\; class="notranslate">\; +</span>\frac{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; sim</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; knn</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; \&times;</span><span\; class="notranslate">\; (</span>{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}{\mathrm{<span\; class="notranslate">\; n</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; ,</span>\mathrm{<span\; class="notranslate">\; j</span>}}<span\; class="notranslate">\; -</span>\stackrel{<span\; class="notranslate">\; \&OverBar;</span>}{{\mathrm{<span\; class="notranslate">\; R</span>}}_{\mathrm{<span\; class="notranslate">\; k</span>}{\mathrm{<span\; class="notranslate">\; n</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}}<span\; class="notranslate">\; )</span>}{\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; k</span>}}{\mathrm{<span\; class="notranslate">\; \&Sigma;</span>}}}\mathrm{<span\; class="notranslate">\; sim</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; a</span>}<span\; class="notranslate">\; ,</span>{\mathrm{<span\; class="notranslate">\; knn</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span>}$

The calculated predicted scores of all unrated items by user a are sorted from large to small, and the top N items with the highest score value are selected to form a recommendation list RecList(N) to the current active user a.

Compared with the existing collaborative filtering recommendation method, the present invention considers the popularity difference of the item itself in the similarity calculation, and reflects it in the calculation formula as a weight, so that the calculation result is more in line with the objective reality, and the recommendation is accurate to a certain extent Performance has improved, improving recommendation quality.

Claims (4)

1. A collaborative filtering recommendation method that introduces program popularity weights, is characterized in that on the interface of IPTV programs, a visual menu of user ratings is provided, and the program is made according to the user viewing time, behavior operations, and program rating data transmitted from the terminal set-top box Recommend the list to the target user, the specific steps are as follows: a. Collect user interest data and make a "user-item" scoring matrix A(m, n); b. Calculate the popularity weight value of the project in an offline cycle; c. Find the corresponding popularity weight value for the item rated by the current active user; d. Make and sort the similarity between the target user a and other users; e. Select the K users with the highest similarity as their nearest neighbor set; f. Predict and sort the unrated items of the target user according to the nearest neighbor set; g. Make a recommendation list of the top N items with the largest predicted ratings to the target user. 2. According to claim 1, the collaborative filtering recommendation method introducing program popularity weights is characterized in that the "user-item" rating matrix A (m, n) is arranged in a matrix with user rating information and user behavior data, Rows represent users, columns represent items, and the value of an element in the matrix represents how much the user in the row likes the item in the column. 3. According to claim 1, the collaborative filtering recommendation method introducing program popularity weight is characterized in that the off-line cycle is calculated once every 30 minutes, and the popularity weight value of the item is $w_t=\log \left(\frac{P_{\mathrm{all}}}{P_t}\right)$ For calculation, w _t is the popularity weight; P _all is the sum of popularity of all items; P _t is the popularity of t item; 4. The collaborative filtering recommendation method that introduces program popularity weights according to claim 1, characterized in that the similarity between the target user a and other users is calculated by introducing the popularity weight value into the Pearson correlation coefficient.

Images