CN101495941A - Neighborhood optimization for content recommendation - Google Patents

Abstract

A cost function is stochastically optimized using, e.g., simulated annealing to render a neighborhood of entities based on which content recommendations can be provided to a user of a home entertainment system (14). The cost function represents a normalized sum of rating similarity scores from entities (15) of the neighborhood that are related to content items viewed by the user.

Description

Be used for the neighborhood optimization of commending contents

The application requires the right of priority of the U.S. Provisional Patent Application sequence number 60/835,020 of submission on August 1st, 2006, and this patented claim is incorporated herein by reference.

Technical field

The present invention relates generally to be used for the system and method for commending contents.

Background technology

Developed based on (rating) the system and method for grading from neighbours' (neighbor) (for example, other domestic system in the cable television system or other entertainment network) content to user's content recommendation of home entertainment system.Understand easily, for to specific user's content recommendation, it is very important selecting neighborhood (neighborhood) member, and this is because if used wrong neighborhood, recommend so may concerning its institute at the specific people effect not quite.

As understanding at this, the current method that neighborhood is selected not is the best approach that they may have.For example, in collaborative filtering (CF), the consumers' opinions of project grading form is collected, and is requested to make when recommending in system, and system for example comes recognition category like the project of user to advise that these users like in the past based on demographic similarity.This method is only based on the judgement to user's neighborhood.

Can use the similarity between two users to come the assesses user similarity, but, rely on this method separately, must limit neighbours' number in advance usually, and not have good method to learn that how many neighbours are best as recognizing at this based on cosine.Therefore, under the situation of the neighbours' that do not know to make the personal user to be benefited optimal number, general predetermined fixed neighbours number.The fact that makes problem become complicated is: be used to form more than the granule amount in 1,000 other users' the known universe of sum of method of neighborhood of fair-sized, therefore, as recognizing, preferably can limit pseudo-best (pseudo-optimal) neighborhood at this feasiblely.At this problem the present invention has been proposed.

Summary of the invention

A kind of being used for to the domestic system content recommendation so that described content is presented at method on the display device of described domestic system disclosed.This method comprises: use the cost function describe below to define other user's neighborhood, and based on described neighborhood, provide at least one commending contents to the user of described domestic system.

In certain embodiments, described cost function expression is from the entity of described neighborhood and the relevant grading similarity scoring sum of content item that watch with described user.Described cost function can be by normalization.If desired, described grading similarity scoring can be based on the similarity scoring of cosine, and can use the random device such as simulated annealing to optimize described cost function.In certain embodiments, can use the operational character (operator) that describes below and be called as " moving ", " exchange ", " exchange-exchange " and " exchange-move " to realize described simulated annealing.

On the other hand, a kind of system receives the grading from entity, and described entity always comprises the potential neighbours of object user in the neighborhood.This system returns to the object user based on described grading with commending contents.This system comprises at least one server, and this server is programmed setting up pseudo-best neighborhood at random based on described grading, and based on described neighborhood commending contents is returned.

On the other hand, the function that assesses the cost to computer iterations, this cost function are represented from the entity of the neighborhood of home entertainment system and the relevant grading similarity scoring sum of content item that watch with the user of described home entertainment system.Described neighborhood is used for to user's content recommendation of described home entertainment system.

Can come to understand best the details about its structure and operation of the present invention with reference to the accompanying drawings, in the accompanying drawings, the similar similar part of label indication, wherein:

Description of drawings

Fig. 1 is the block diagram according to non-limiting system of the present invention; And

Fig. 2 is the process flow diagram of this logic.

Embodiment

At first with reference to figure 1, show and always be designated as 10 system, this system comprises content provider server 12, such as but not limited to wired front end (cable head end) server that content can be provided to main user household system 14 with aptitude manner.Primary home system 14 can comprise the one or more multimedia display equipment such as televisor and/or computing machine, and the one or more multimedias or the content data storage device that can be provided in content displayed on the display device, for example DVR and Disc player (for example digital video disk (DVD) or Blu-ray player etc.).Under hard-core situation, user system interface can be realized by set-top box.Replace set-top box or except set-top box, user system interface can also be connected by the wide area network of the Internet connection device such as wired or wireless modulator-demodular unit or other type and realizes.Therefore, primary home system 14 and below can be via the Internet and/or the TV cable and/or the broadcasting link of land and satellite mode with the communication between the server of describing, and can be two-way, for example, primary home system 14 can receive from the content of server 12 and will transmit back server 12 to the grading of specific multinomial content.

As shown in Figure 1, not only primary home system 14 can be communicated by letter with server 12, and neighbours' domestic system 15 also can be communicated by letter with server 12.Similar to primary home system 14, neighbours system 15 can receive the content from server 12, and will transmit back server 12 to the grading of specific multinomial content.The subset definition near best (" pseudo-best ") that the objective of the invention is neighbours system 15 is " neighborhood " of primary home system 14, can be used to primary home system 14 content recommendations from the grading of neighborhood.Should be understood that identical processing can be used for each neighbours system 15, wherein, neighbours system 15 can be the primary home system in the following algorithm.

In shown non-limiting implementation, content server 12 can determine that server 16 communicates by letter with neighborhood, and neighborhood is determined that server 16 is actually and carried out following logic to determine the computing machine near best neighborhood.According to principle well known in the art, can be sent to recommendation server 18 near best neighborhood, recommendation server 18 can return to primary home system 14 with commending contents by using this neighborhood.Though show three servers 12,16,18, also can use more or less server.

Forward Fig. 2 now to,, set up cost (cost) function C at piece 20 places.According to the present invention, provided preferred cost function by formula (1).Cost function is used for based on from the grading of neighborhood system 15 and set up pseudo-best neighborhood randomly.In the formula (1) rightmost and expression from neighborhood system 15, relevant grading similarity scoring with the content item that the user watched of master's house front yard entertainment systems 14 and (it can use based on the method for cosine or other similarity based method), and middle be actually a normalized factor.In other words, the cost function C that optimizes be for given neighborhood N, at all scoring sums of the content item in user's receiver (bin) of main system 14, and its scope (because normalization) is for changing to 1 from 0, and wherein, desirable best neighborhood has been found in 1 indication.

$\left(1\right),C=f\left(N\right)=\frac{1}{\left|N\right|}\underset{s\∈\hat{S}}{\mathrm{\Σ}}(\frac{1}{\underset{c^{\′}\∈\hat{C}}{\mathrm{\Σ}}\left|\mathrm{sim}(c,c^{\′})\right|}\×\underset{c^{\′}\∈\hat{C}}{\mathrm{\Σ}}\mathrm{sim}(c,c^{\′})\×r_{c^{\′},s})$

Above the formula textual representation be: C (will be maximized)=1/N multiply by (passed through grading all film s with) { 1/ (for all the candidate neighbours 15 in the current neighborhood, from each candidate neighbour's 15 grading and come the similarity between the grading of autonomous system 14 absolute value and) multiply by [(for all the candidate neighbours 15 in the current neighborhood, from each candidate neighbour's 15 grading and come between the grading of autonomous system 14 similarity and) multiply by grading to the film of institute's addition from candidate neighbours].

In case defined cost function, then to move to piece 22 be initial " neighborhood " of main system 14 with the initial subset definition with all neighbours systems 15 to logic.Can select subclass at random.Advance to piece 24, calculate above-mentioned cost function for the first time.

At piece 26 places, select an operational character, utilize this operational character, neighborhood is by disturbance (perturb).In a kind of nonrestrictive implementation,, can select one of four operational characters at piece 26 places.A system 15 of non-neighborhood can enter neighborhood (" moving "), " moves " simultaneously also to comprise a system 15 is shifted out current neighborhood.Second operational character (" exchange ") relates to not system in current neighborhood 15 and 15 exchanges of the system in current neighborhood, and the 3rd operational character (" exchange-exchange ") relates to system in current neighborhood 15 and the system in neighborhood 15 exchange twice continuously.The 4th operational character (" exchange-move ") need be carried out to exchange and then to carry out then and move.

Advance to piece 28, after carrying out disturbance, recomputate the cost function by the operational character at piece 26 places.Advance to decision diamond 30, judge whether the value of cost function has increased than preceding value,, then accept disturbance at piece 32 places if increased.But if do not increase at decision diamond 30 place's cost functions, then logic moves to decision diamond 34, utilizes by the probability of formula (2) definition and accepts or do not accept disturbance.Based on this probability and lot at random (random draw) that a numeral and this probability are compared, can accept disturbance or not accept disturbance at piece 32 at piece 36, do not accept disturbance then disturbance in fact be cancelled so that neighborhood continues to remain on piece 26 places it carried out the member relation that it had before the disturbance.

$\left(2\right),P=e^{-\frac{\mathrm{\ΔC}}{T}}$

P=e ^{-(change among the C)/T}, wherein " T " is the annealing of rule of thumb determining " temperature " (annealing " temperature ").

Whether enough low decision diamond 38 decision probabilities (for example, being lower than the value of rule of thumb determining) obtained near best neighborhood with indication.If probability is enough low, then process finishes so that current neighborhood is outputed to main system 14 as the recommendation basis at state 40.Otherwise logic loops is returned piece 26.

If desired, the increasing cost function evaluation can be used, also look-up table can be used.Except those operational characters described here, other operational character can also be used, and short annealing can be used.And, can use the cost function solution based on the user of standard to define initial neighborhood, and can use other cost function.

Be appreciated that now the present invention can be with doing current engine,, and need not to be carried out to relatively at user/project similarity assessment for example to the improvement of current similarity engine based on cosine.Can use this random algorithm to obtain near best (being also referred to as " pseudo-best ") solution at this.Compare with existing method, can obtain the neighborhood of different sizes, these neighborhoods can be advantageously used in different user to obtain better solution.

As substituting of simulated annealing, can use genetic algorithm is that so-called ant group (ant colony) optimizes, and other method.

Though be shown specifically and described specific the neighborhood optimization system and the method that are used for commending contents at this, should be understood that the theme that the present invention comprises is only limited by claim.

Claims (9)

1. one kind is used for to domestic system (14) content recommendation comprising so that described content is presented at method on the display device of described domestic system (14):

Cost function below using defines other user's neighborhood:

$C=f\left(N\right)=\frac{1}{\left|N\right|}\underset{s\∈\hat{S}}{\mathrm{\Σ}}(\frac{1}{\underset{c^{\′}\∈\hat{C}}{\mathrm{\Σ}}\left|\mathrm{sim}(c,c^{\′})\right|}\×\underset{c^{\′}\∈\hat{C}}{\mathrm{\Σ}}\mathrm{sim}(c,c^{\′})\×r_{c^{\′},s})$

Based on described neighborhood, provide at least one commending contents to the user of described domestic system (14).

The method of claim 1, wherein described cost function expression from the entity of described neighborhood and the relevant grading similarity scoring sum of content item that watch with described user.

3. the method for claim 1, wherein described cost function is by normalization.

4. method as claimed in claim 2, wherein, described grading similarity scoring is based on the similarity scoring of cosine.

5. the method for claim 1, wherein use random device to optimize described cost function.

6. method as claimed in claim 5, wherein, described random device is a simulated annealing.

7. method as claimed in claim 6, wherein, use at least two operational characters to realize described simulated annealing, described at least two operational characters are to select from the group that comprises following operational character: non-neighborhood entity is moved to (" moving ") in the described neighborhood, the neighborhood entity is shifted out described neighborhood (" moving "), will be in entity in the described neighborhood and the not exchange of the entity in described neighborhood (" exchange "), to exchange twice (" exchange ") with the entity in described neighborhood is not continuous at the entity in the described neighborhood, and carry out exchange and and then carry out mobile subsequently.

8. system, this system receives grading from the potential neighbours' (15) that always comprise object user the neighborhood entity, and this system returns to described object user based on described grading with commending contents at least in part, and this system comprises:

At least one server (12,16,18), described at least one server are programmed so that pseudo-best neighborhood is set up at random based on described grading in small part ground, and come returned content to recommend based on described neighborhood at least in part.

9. system as claimed in claim 8, wherein, server (12,16,18) use cost function defines described neighborhood, and wherein, an expression of described cost function is from the entity of described neighborhood and the relevant grading similarity scoring sum of content item that watch with described user.

Images