WO2009059437A1

WO2009059437A1 - Method and computer system for indexing multimedia data objects

Info

Publication number: WO2009059437A1
Application number: PCT/CH2008/000114
Authority: WO
Inventors: Riley Crane; Didier Sornette
Original assignee: ETH Zürich
Priority date: 2007-11-07
Filing date: 2008-03-18
Publication date: 2009-05-14

Abstract

For indexing a plurality of multimedia objects accessible to a plurality of users via a telecommunications network, determined and assigned (S4) to each of the multimedia objects is a respective classification index which indicates an objective measure of attention paid by the users to the respective multimedia object. For determining the classification index (S4), established is a time series (S2) which indicates the number of accesses to the respective object within each of multiple consecutive time periods. Subsequently, the classification index is determined (S4) based on the power law of the relaxation of the number of accesses following a peak in the time series. The classification index provides an objective measure of the access dynamics, enabling search as well as cache and storage management for multimedia objects, without having to rely on content description, mere magnitude of view count, or user ratings.

Description

METHOD AND COMPUTER SYSTEM FOR INDEXING MULTIMEDIA DATA OBJECTS

Field of the Invention The present invention relates to a computer-implemented method and a computer system for indexing a plurality of multimedia data objects. Specifically, the present invention relates to a computer-implemented method and a computer system for indexing a plurality of multimedia data objects accessible to a plurality of users via a telecommunications network, whereby for each of the multime- dia data objects determined is a classification index which indicates an objective measure of attention paid by the users to the respective multimedia data object.

Background of the Invention

The emergence of the Internet as a vehicle for news, commerce and social activity, has resulted in so much user-generated content that information is now diffi- cult to manage and locate. Moreover, the overload of information results in the user's attention becoming scarce, and valuation becomes difficult. Existing approaches, such as those of Google and other search and/or recommendation engines, exploit static snapshots of the World Wide Web's evolving structure to identify relevance, but success is limited by the lack of rich link structure and text common to content on sites like YouTube and Flickr. Other methods solve this problem using metadata, voting systems, or more commonly, by serving up the most popular (i.e. most accessed) items. However, this approach fails to account for the famous long tail behavior of the web, i.e. the collective effect of numerous small, niche interests outweighing the demand of the mainstream, thus providing content that has mass appeal and masking the interests of the idiosyncratic many.

Financial markets, and more recently prediction markets, are the epitome of how information and self-consistency are aggregated into prices, seen as the result of the collective vote of the majority. This idea of valuation or relevance selection through an aggregation of information is also at the basis of the PageRank method used by Google and by most other recommendation methods and search engines on the Internet. In such approaches, the adjacency matrix of the World Wide Web graph at a given time is used to extract the relevant tree structure and provide recommendation via principal vector analysis, correlation-based methods, related spectral analysis, or opinion diffusion. In a nutshell, these algorithms find the "equilibrium", self-consistent, static ranking that best represents all past votes of users according to various selected criteria. However, these approaches require extensive inputs and learning in order to deliver useful and relevant results, which include better understanding of users and items, incorporating contextual information into the recommendation process, support for multiple ratings criteria, and accounting for the inherent variability in the databases.

Summary of the Invention It is an object of this invention to provide a computer-implemented method and a computer system for indexing a plurality of multimedia data objects accessible to a plurality of users via a telecommunications network, such as the Internet, whereby a classification index indicates in each case an objective measure of attention paid by the users to the respective multimedia data object. In particular, it is an object of the present invention to provide a computer-implemented method and a computer system for indexing a plurality of multimedia data objects without having to rely on content description or other meta data associated with the multimedia data objects, and without being dependent on the mere magnitude of a view count or the rating provided by a user voting system. It is a further object of the present invention to provide a computer-implemented method and a computer system for indexing a plurality of multimedia data objects for subsequent selection of the multimedia data objects based on their respective classification index, e.g. in search applications, responsive to user queries, or in caching and storage management for improving performance and effi- ciency of resource usage. It is another object of the present invention to provide a computer-implemented method and a computer system for searching and accessing multimedia data objects based on a classification index assigned to the multimedia data objects. It is yet another object of the present invention to pro- vide a computer-implemented method and a computer system for managing storage and/or caching of multimedia data objects based on a classification index assigned to the multimedia data objects.

According to the present invention, these objects are achieved particularly through the features of the independent claims. In addition, further advanta- geous embodiments follow from the dependent claims and the description.

For indexing a plurality of multimedia data objects accessible to a plurality of users via a telecommunications network, determined and assigned to each of the multimedia data objects is a respective classification index which indicates an objective measure of attention paid by the users to the respective multimedia data object.

According to the present invention, the above-mentioned objects are particularly achieved in that, for each of the multimedia data objects, a time series is determined which indicates for multiple consecutive time periods the number of ac- cesses by the users to the respective multimedia data object in the respective time period. For each of the time series, determined is a power law of relaxation (decline) of the number of accesses following a peak number of accesses in the respective time series. For example, the multimedia data objects include digital video objects, accessible to users on a social networking application on the Internet, and the number of accesses in a time period includes daily viewing counts. The classification index is determined for each of the multimedia data objects based on the power law of relaxation determined for the time series associated with the respective multimedia data object. The classification index is assigned to the respective multimedia data object. The classification index provides an objective measure of the access dynamics and attention paid by the users to a multimedia data object. This objective measure enables searching for multimedia data objects as well as cache and storage management of multimedia data objects without having to rely on content description or other meta data associated with the multimedia data objects, and without being dependent on the mere magnitude of a view count or the rating provided by a user voting system. In an alternative application, this objective measure makes it possible to assess the impact of a marketing or advertising campaign to the access rate associated with a specific multimedia data object. Preferably, the classification index is stored as a searchable index enabling users to query for multimedia data objects by including the classification index in the query data, and receiving with the query result multimedia data objects associated with the specified classification index.

In an embodiment, multimedia data objects are selected based on the classification index for caching the multimedia data object on a server accessible efficiently for at least some of the users.

In a preferred embodiment, the classification index is determined by assigning an exogenous sub-critical index value, indicative of an exogenous stimulus with non-lasting attention, for a relaxation having a characteristics of -^- , by assign¬

ing an exogenous critical index value, indicative of an exogenous stimulus with lasting attention, for a relaxation having a characteristics of -^- , and by assign-

ing an endogenous critical index value, indicative of an endogenous stimulus with lasting attention, for a relaxation having a characteristics of -^- , t being

representative of time, and θ having a value of approximately 0.4. Preferably, the power law of relaxation is determined by finding the largest peak in the time series, and by finding for a plurality of time windows of different

lengths an optimal set of parameters A, t_c, α for a best fit of y(t) ■= to the

(t - *c)^a

time series in an optimally long time window following the largest peak.

In addition to a computer-based method and a computer system for indexing a plurality of multimedia data objects, the present invention also relates to a computer program product including computer program code means for controlling one or more processors of a computer, particularly, a computer program product including a computer readable medium containing therein the computer pro- gram code means.

Brief Description of the Drawings

The present invention will be explained in more detail, by way of example, with reference to the drawings in which:

Figure 1 shows a block diagram illustrating schematically an exemplary con- figuration of a computer system for indexing multimedia data objects which are accessible to a plurality of users via a telecommunications network. Figure 2 shows a flow diagram illustrating an example of a sequence of steps executed according to the present invention for indexing the multimedia data objects.

Figure 3 shows an example of a synthetic data set representing a time series of access count for a multimedia data object.

Figure 4 shows an example of a time-series of the daily view count following an exogenous shock for a popular video.

Figure 5 shows an example of an endogenous time-series for a video whose peak view. count is preceded by a significant number of views.

Figure 6 shows the histogram of the exponents associated with three different classes of power law relaxations.

Figure 7 shows an example of normalized daily view counts illustrating varying precursory growth before a peak and after-peak relaxation for the three different classes.

Detailed Description of the Preferred Embodiments In Figure 1, reference numeral 1 refers to a computer system comprising one or more computers, having one or more processors. As is illustrated schematically, the computer system 1 is accessible to a plurality of users' communication terminals 3 via a telecommunication network 2. For example, the communication terminals 3 include personal computers, mobile radio telephones and/or personal digital assistants. The telecommunication network 2 includes fixed networks and/or wireless networks. For example, the telecommunication network 2 includes the Internet, a local area network (LAN), an integrated services digital network (ISDN), a GSM network (Global System for Mobile Communication, a UMTS network (Universal Mobile Telephone System) or another mobile radio telephone system, and/or a wireless local area network (WLAN).

As is illustrated schematically in Figure 1, the computer system 1 includes a data server 10 connected to a data store 11, as well as various functional modules including a counter 100, an optional search engine 12, an optional cache mana- gement module 13, as well as a sampling module 141 and a classification module 142. For example, data server 10 and data store 11 correspond to YouTubeTM (YouTube.comTM website), at present hosting the largest video library worldwide and tracking view counts of millions of video data objects. Preferably, the functional modules are implemented as programmed software modules. The computer program code of the software modules is stored in a computer program product, i.e. in a computer readable medium, either in memory integrated in a computer of computer system 1 or on a data carrier that can be inserted into a computer of computer system 1. One skilled in the art will understand, howe- ver, that, alternatively, the functional modules can be implemented fully or partly by means of hardware components. In different embodiments, the search engine 12 and the cache management module 13 are implemented on the same computer (s) as data server 10, or on one or more other computers. Preferably, the sampling module 141 and the classification module 142 are implemented on an indexing server 14, connected to telecommunication network 2, with one or more computers separate from the data server 10. Alternatively, the sampling module 141 and the classification module 142 are implemented on the same computer (s) as data server 10. Data store 11 comprises a database and/or a file server, and includes a plurality of multimedia data objects 111, e.g. digital video objects, audio objects and/or photographs stored as data files with standardized formatting and coding of contents.

In the following paragraphs, described with reference to Figure 2 is a possible sequence of steps executed for indexing the multimedia data objects 111. In preparatory step SO, counter 100 keeps continuously track of the number of access to each of the multimedia data objects 111, e.g. the number of viewings of a video object, downloads of a music file, or access to a photograph. An access count data field is assigned to each of the multimedia data objects 111. In an alternative embodiment, the multimedia data objects 111 represent books.

In step Sl, sampling module 141 periodically reads the number of accesses (access count) to the multimedia data objects 111. For example, the access count is determined once or more times a day. For example, the sampling module 141 uses an API (application programming interface) provided by YouTubeTM to submit a data request query to the data server 10, and receives a structured (XML) document containing information such as the cumulative number of times a video has been viewed (dynamic), along with descriptive information (static) concerning the user who posted the video, the title, tags, length, category, rating, comments, etc. Once a multimedia data object is added to a tracking list, infor- mation about the access count is continuously tracked and sampled at least once per day; the tracking frequency is limited by the update time of the data server, for example, in the case of YouTubeTM, the update time is typically in a range between 4 to 10 hours. In the case of YouTubeTM, the sampling module 141 further tracks continuously any video which appears on any "featured" page of the site, with a sampling rate of approximately 30 minutes. In this way, tracked is the impact of preferred placement (i.e. featuring) of videos.

In step S2, sampling module 141 establishes for each or selected ones of the multimedia data objects 111 a time series which records for a defined time pe- riod, e.g. daily or hourly, the (incremental) number of accesses to the respective multimedia data object 111 during the respective time period. For example, sampling module 141 keeps a chronological table indicating for every consecutive time period the individual number of accesses to each or selected ones of the multimedia data objects 111, as illustrated in Table 1.

Table 1

For this purpose, the sampling module 141 converts the raw cumulative access count C(t) into incremental access counts per sampling period, e.g. incremental access count Vi on day i. For example, in the case of YouTubeTM, the cumulative view count for each video is sampled at least twice per day, but often times much more. Subsequently, removed are artifacts that arise from issues related to data transmission and server update errors. These issues arise because YouTubeTM updates their servers only every four to ten hours to reflect the current C(t). Furthermore, when data is requested for a particular video, the request is randomly assigned to one of YouTube's cloned servers, which may have a different C(t) than the most recent count because of lags in updating the system as a whole. The data is resampled into "bins" of one day, creating a coarse-grained time-series based on the underlying, higher-frequency sampling. The views per day is calculated from the difference of view counts of consecutive days, V₁ = Q₊₁ - Q. In step S3, the classification module 142 determines for the time series of each of the multimedia data objects 111, or the time series of the selected multimedia data objects 111, the power law of relaxation (decline) of the number of accesses following a peak number of accesses in the respective time series. The problem simply stated is: given a time-series of the incremental access count per day, whose underlying dynamics is thought to be governed by a power-law of the form

estimate the best set of parameters (A, t_c, α)i and best window-size i over which the fit is performed. At first, excluded from the multimedia data objects are those that do not have a minimum access count per day, e.g. at least hundred views per day. Furthermore, for measuring the relaxation signature, excluded are multimedia data objects having a time-series with fewer than a minimum number of days of data after the main peak, e.g. ten days. For each of the remaining time- series, determined is the largest peak and extracted are the three parameters of equation (1) for a ten day window following this peak. This is done by performing a least-squares fit on the logarithm of the data, which is equivalent to the maximum-likelihood estimates of α and A, assuming log-normal multiplicative noise on y(t) . For this window size, selected are fifty values of t_c in increments of 0.1 days in the 5 days before the peak in order to find the set (A, t_c, α)i that minimizes the RMS (root mean square) deviations of the data from the fit. Subsequently, the window size is increased by one day, and the process is repeated until the triplet parameter sets (A, t_c, α)i have been calculated for all of the days following the largest peak, which can be anywhere from ten to two hundred and twenty four days. The dynamics of the daily view count follow a complicated trajectory, being influenced by a number of factors ranging from random stochastic forces, propagation of a multimedia data object through a social network by email or other methods, featuring the multimedia data object on the data server 10, or any other website, newspaper, magazine, television show, to name a few. For measuring the relaxation of the dynamics after a peak, needed is an algorithm that automatically finds the best range over which to fit the data, so as to stop fitting once the system undergoes a subsequent shock. Furthermore, this algorithm is able to reject the data when the dynamics evolve from equation (1) to some other characteristics.

Having calculated the best parameter triplet (A, t_c, α)i for each window size, it remains to be determined which set i best represents the data over the largest number of days. In reality there is noise in the data so it is typically more ap- propriate to consider the following stochastic process for the rate of access to the multimedia data objects

y(t) = ^{A(l + η)} , (2)

where η is some stochastic noise, and δt represents the uncertainty arising from the coarse graining of the data into periods of one day, which is uniformly distributed between -0.5 and +0.5 days. The hypothesis that the multiplicative noise 1 + η is log-normal amounts to assuming that In(I+ η)«η is normally distributed ( N(0,σ)) with zero mean and standard deviation σ. The value of σ is estimated directly from the data by measuring the sample standard deviation from the best fit of In y(t) during the 10 days following the peak. A simple propagation of uncertainties enables the estimation of the bounds within which the data should fall. An example of this is shown in Figure 3 for synthetic data generated using (A=IOOOO, t_c =3.2, α=1.5) along with σ= 30%. Using these parameters, calculated is also the interval within which it is expected to find 99% of the data (three standard deviations) using propagation of errors and the assumption that the percent deviations are normally distributed. This makes it possible to estimate the appropriate window size i for extracting (A, ^, α)i. Subse- quently, determined is the probability that the first point (i+1) beyond the window belongs to the best fit (A, t_c, α)i calculated for i days. Specifically, calculated is the probability that the next two points beyond the window (i+1, i+2) belong to the best fit curve for a window of i days. Continuing in this way generates a list of probabilities that the next N points beyond the current fit window belong to the curve found by fitting only the points in the window. Now it is possible determine whether or not the points beyond the current fit window belong to the same curve found in the fit window. This is done by comparing the probability for these points to belong at the 1% level

P₁ - P₂ -P₃ ^{■ ■ ■} P_N > (0.01)" . (3)

If the quantity on the left is larger than 0.01N, then it is reasonable to expect that the data beyond the fit window is well described by the fit within the window.

This procedure is repeated, continually expanding the window, until it is found that four consecutive days beyond the window are not likely to belong to the fit found for a given window size. The value of four days was chosen after extensive testing, and gives a nice balance between allowing small fluctuations and rejecting clear regime changes. In step S4, based on the power law, i.e. exponent α, determined in step S3, the classification module 142 determines for each or selected ones of the multimedia data objects Ili a classification index. In other words, the classification module 142 determines the classification indices based on the dynamic signature (of re- laxation) of access counts. Subsequently, the classification index is stored assigned to the respective multimedia data object. Preferably, the classification indices are stored as searchable indices.

In a first class, a multimedia data object is classified as "exogenous sub-critical", when the network is not "ripe", i.e. an activity generated by an exogenous event at time t_c does not cascade beyond the first few generations, and the activity is proportional to the direct (or "bare") memory function:

For an exogenous sub-critical peak, the fraction of the total access (view) count F during the day of the peak is significantly larger than the total number of ac- cess (view) counts over the rest of the dynamics, because of the lack of precursory growth coupled with quickly diminishing access numbers after the peak. For example, in the first class the fraction of the total access count F is in the range 80% ≤ F < 1.

In a second class, a multimedia data object is classified as "exogenous critical", when the network is "ripe" for a particular multimedia data object, and the bare response is renormalized as the spreading is propagated through many generations, and the activity is expected to be described by:

For an exogenous critical peak, the fraction of the total access (view) count F during the day of the peak is smaller than for the previous case, with an access (view) count during the peak day roughly comparable with the total number of access (view) counts recorded during the rest of the dynamics, because the event manages to penetrate the social network and spread. For example, in the second class the fraction of the total access count F is in the range 20% ≤ F ≤ 80%.

Figure 4 shows the time-series of the daily view count following an exogenous shock for the very popular video on YouTubeTM "Hahaha"

(http://youtube.com/watch? v=5P6UU6m3cqk) which had over 25 million views. The decay (inset) follows a power-law over more than 70 days. A power law fit reveals an exponent of p«0.8 for the relaxation dynamics of the view counts ~ l/t^p , suggesting that this video received an exogenous shock which spread through the network of viewers (p=l~θ and θ«0.2).

In a third class, a multimedia data object is classified as "endogenous critical", if in addition to being "ripe", the burst of activity is not the result of an exogenous event, but is instead fueled by endogenous (word-of-mouth) growth, the bare response is renormalized in a different way giving the following time- dependence for the access count before and after the peak of activity:

For an endogenous critical peak, expected are significant precursory growth as well as slow decay, implying that the fractional access (view) count F contained in the peak is small compared to the total number of accesses (views). For example, in the third class the fraction of the total access count F is in the range 0 ≤ F < 20%. Figure 5 shows an example of an endogenous time-series for a video whose peak view count was preceded by a significant number of views. The precursory behavior before the view count peak is approximately the mirror image of the aftershock dynamics following the peak.

Of course, the fractional threshold values of 20% and 80% do not play any special role. These values have been changed between 10% and 40% for the lower value and between 60% and 90% for the upper value, without significant qualitative change in the distributions shown in Figure 6.

Figure 6 shows the histogram of the exponents α of the power law relaxation ~ \/t^a of the view counts following a peak belonging respectively to the first class (α=l+θ, dashed line), the second class (α=l-θ, dotted line) and the third class (α=l-2θ, continuous line). Even using such a blunt rule for grouping together videos in terms of the fraction F, one immediately sees that the exponents cluster into three distinct classes, providing evidence that the proposed classifi- cation model is able to capture the behavior for individual multimedia data objects, such as videos. Notwithstanding its simplicity, the classification in terms of the fraction F is associated with very different values of the exponents. The most probable mode of the distributions of exponents for each class give respectively the values α=1.4, α=0.6 and α=0.2 for the first, second or third class, respec¬

tively. These values are compatible with the predictions of the epidemic branch¬

ing model as described in Sornette, D. and A. Helmstetter, Endogeneous Versus

Exogeneous Shocks in Systems with Memory, Physica A 318 (3-4), 577-591

(2003), with a unique common value for θ equal to 0.4±0.1. The predicted val¬

ues for the exogenous sub-critical (equation (4)), exogenous critical (equation

( (5)) and for the endogenous critical (equation (6)) are shown by the vertical

dashed lines with their quantitative values with the choice of θ=0.4.

Having empirically extracted a value of θ=0.4, a further test of the model is pro-

vided by determining if the dynamics of videos with these exponents are consis¬

tent with the description of the model. This is checked by performing a peak-

centered, aggregate sum for all videos available on YouTubeTM with exponents

α near either 1.4, 0.6, or 0.2, with the intent of visualizing the time evolution.

Each time-series is first normalized to one, to avoid a single video from dominat-

ing the sum, and the final result is divided by the number of videos in each set,

so that the three classes can be compared. Here, the test of the epidemic model

lies in the precursory dynamics. The model predicts, and can be seen in Figure 7,

that videos whose post-peak dynamics is governed by small exponents have sig- nificantly more precursory growth. One also sees very little precursory growth for the two exogenous classes. Since by construction the grouping and classification is based on the exponent characterizing the relaxation after the peak, one is not surprised to visualize faster decays for those videos with exponents near 1.4 compared with those of 0.6 and 0.2.

The search engine 12 is configured to receive from users of the communication terminals 3 query requests for multimedia data objects. The query requests are received via telecommunications network 2 and comprise query data including a classification index (and possible other query parameters) to define the re- quested multimedia object(s). Based on the classification index, the search engine 12 searches and determines corresponding multimedia data objects in data store 11. Depending on the embodiment and/or application, the search engine 12 returns to the querying user (or application) one or more references (links) to the respective multimedia data object(s), or data server 10 returns to the user (application) the respective multimedia data object(s). Thus, the search engine 12 makes it possible for users (applications) to query for and subsequently receive multimedia data objects having a defined access dynamics and/or experienced user attention. The cache management module 13 is configured to select at least one of the multimedia data objects 111 based on the classification index for caching the multimedia data object on a data server 10 accessible efficiently for at least some of the users' communication terminals 3 (e.g. on a data server 10 located in tele- communications network 2 near the users' communication terminals). Thus, the cache management module 13 makes it possible to cache multimedia data objects depending on their respective access dynamics and/or experienced user attention.

It should be noted that, in the description, the computer program code has been associated with specific functional modules and the sequence of the steps has been presented in a specific order, one skilled in the art will understand, however, that the computer program code may be structured differently and that the order of at least some of the steps could be altered, without deviating from the scope of the invention.

Claims

1. A computer-implemented method of indexing a plurality of multimedia data objects (111) accessible to a plurality of users via a telecommunications network (2), the method comprising determining and assigning to each of the multimedia data objects (111) a respective classification index indicating an objective measure of attention paid by the users to the respective multimedia data object, wherein determining and assigning of the classification indices includes:

determining (S2) for each of the multimedia data objects (111) a time se- ries which indicates for multiple consecutive time periods the number of accesses by the users to the respective multimedia data object in the respective time period;

determining (S3) for each of the time series a power law of relaxation of the number of accesses following a peak number of accesses in the respec- tive time series;

determining (S4) the classification index for each of the multimedia data objects (111) based on the power law of relaxation determined for the time series associated with the respective multimedia data object; and storing the classification index assigned to the respective multimedia data object.

2. The method of claim 1, wherein the classification index is stored as a searchable index; and the method further comprises receiving from the us- ers query data including a classification index, and returning to the users a query result with at least one multimedia data object selected based on the classification index received from the users.

3. The method of one of claims 1 or 2, further comprising selecting at least one of the multimedia data objects (111) based on the classification index for caching the multimedia data object on a server accessible efficiently for at least some of the users.

4. The method of one of claims 1 to 4, wherein determining the classification index (S4) comprises assigning an exogenous sub-critical index value, indicative of an exogenous stimulus with non-lasting attention, for a relaxa- tion having a characteristics of -^- , assigning an exogenous critical index

value, indicative of an exogenous stimulus with lasting attention, for a relaxation having a characteristics of -^- , and assigning an endogenous critical index value, indicative of an endogenous stimulus with lasting attention, for a relaxation having a characteristics of —^ , t being represen¬

tative of time, and θ having a value of approximately 0.4.

5. The method of one of claims 1 to 4, wherein determining the power law (S3) of relaxation comprises determining the largest peak in the time series, and determining for a plurality of time windows of different lengths an op- timal set of parameters A, t_c, α for a best fit of y(i) = to the time

(t-t_c)^a

series in an optimally long time window following the largest peak.

6. The method of one of claims 1 to 5, wherein the multimedia data objects (111) include digital video objects; the video objects are accessible to users on a social networking application on the Internet; and the number of accesses in a time period includes daily viewing counts.

7. A computer system (1) for indexing a plurality of multimedia data objects (111) accessible to a plurality of users via a telecommunications network (2), by determining and assigning to each of the multimedia data objects

(111) a respective classification index indicating an objective measure of attention paid by the users to the respective multimedia data object, wherein the system (1) comprises:

a sampling module (141) configured to determine for each of the multimedia data objects (111) a time series which indicates for multiple consecu- tive time periods the number of accesses by the users to the respective multimedia data object in the respective time period; and

a classification module (142) configured to determine for each of the time series a power law of relaxation of the number of accesses following a peak number of accesses in the respective time series, to determine the classifi- cation index for each of the multimedia data objects (111) based on the power law of relaxation determined for the time series associated with the respective multimedia data object, and to store the classification index assigned to the respective multimedia data object.

8. The system (1) of claim 7, wherein the classification module (142) is con- figured to store the classification index as a searchable index; and the system (1) comprises a search engine (12) configured to receive from the users query data including a classification index, and to return to the users a query result with at least one multimedia data object selected based on the classification index received from the users.

9. The system (1) of one of claims 7 or 8, further comprising a cache management module (13) configured to select at least one of the multimedia data objects (111) based on the classification index for caching the multimedia data object on a server accessible efficiently for at least some of the users.

10. The system (1) of one of claims 7 to 9, wherein the classification module (142) is configured to determine the classification index by assigning an exogenous sub-critical index value, indicative of an exogenous stimulus with non-lasting attention, for a relaxation having a characteristics of -^- ,

by assigning an exogenous critical index value, indicative of an exogenous stimulus with lasting attention, for a relaxation having a characteristics of

-£_g , and by assigning an endogenous critical index value, indicative of an

endogenous stimulus with lasting attention, for a relaxation having a characteristics of -^_g- , t being representative of time, and θ having a value of

approximately 0.4.

11. The system (1) of one of claims 7 to 10, wherein the classification module (142) is configured to determine the power law of relaxation by determining the largest peak in the time series, and by determining for a plurality of time windows of different lengths an optimal set of parameters A, t_c, α for a

best fit of y(f) = — to the time series in an optimally long time win-

dow following the largest peak.

12. The system (1) of one of claims 7 to 11, wherein the multimedia data objects (111) include digital video objects which are accessible to users on a social networking application on the Internet; and the number of accesses in a time period includes daily viewing counts.

13. Computer program product comprising computer program code means for controlling one or more processors of a computer in a data processing system (1), such that the computer executes a method of indexing a plurality of multimedia data objects (111) accessible to a plurality of users via a telecommunications network (2), the method comprising determining and assigning to each of the multimedia data objects (111) a respective classification index indicating an objective measure of attention paid by the users to the respective multimedia data object, wherein determining and assigning the classification index comprises:

determining (S2) for the respective multimedia data object a time series which indicates for multiple consecutive time periods the number of ac- cesses by the users to the multimedia data object in the time period;

determining (S3) for the time series a power law of relaxation of the number of accesses following a peak number of accesses in the time series;

determining the classification index (S4) for the multimedia data object based on the power law of relaxation determined for the time series associ- ated with the multimedia data object; and

storing the classification index assigned to the multimedia data object.