RO131695A2 - Customized prediction method concerning categories of internet pages to be visited by a user - Google Patents

Abstract

The invention relates to a customized prediction method based on artificial intelligence algorithms, concerning the categories of Internet pages to be visited by a user, method that can be employed by Internet browsers and providers of web services for suggesting pages thereto, aiming at accelerating and simplifying navigation, with benefits in navigation from personal computers, but especially from mobile devices. According to the invention, the prediction method comprises the following stages: a first stage of pre-processing (3) in which, starting from a current navigation sequence of a user and from his navigation history, one or several acyclic vectors are created with the categories of visited pages, in the visiting order; a second stage, of processing (4), which predicts the category of the next Internet page to be accessed by the user, by combining two different prediction techniques: one based on a recommendation process (12) and the other - a statistic method based on comparing probabilities of access of various categories of pages, using three different factors for making predictions: the user's navigation history, the current moment in time and the user's current navigation pace; there is also a third stage in which the pre-processed information is added to the user's navigation history. The prediction method outputs the probable categories (O1) for the current moment, the probable categories (O2) for the current navigation pace, the probability (O3) of continuing the sequence and the predicted category (O4), by combining the previous categories (O1, O2, O3).

Description

The invention relates to a method of prediction, based on artificial intelligence algorithms, of the probable categories of which web pages will be visited by a user, a method that can be used by Internet browsers and service providers. web site for offering page suggestions, in order to speed up and simplify navigation. Also, the invention can also be used to provide complex web services, based on information from multiple sources, obtained in advance of the time when they are required, as well as to increase security, by detecting an unauthorized user. The category of a web page means the domain to which the page belongs, for example: news, movies, music, science and technology, etc.

Various prediction methods used in different applications are known. Some of them are specialized for mobile devices, for tasks such as predicting the areas of documents that have to be pre-loaded on mobile devices, predicting the words that they want to be typed based on a few ambiguous characters or the functionalities of the phones, such as the ones presented. In Patent Documents Predicting and Retrieving Data for Preloading on Client Device, Inventors Chang J, Yuan S, Wang B, Tsai E, Jiang A (US2013304798A1, published 14.11.2013), A Communication Terminal Having a Predictive Editor Application, Inventor Haestrup Jan (EP1031914A2, published 30.08.2000), Messaging System, inventors Ezra D and Kaluaratchie DC (GB2470585A, published 01.12.2010).

Most methods, however, are either specialized for desktop devices or adapted for both types of devices, for tasks such as improving search engine results, predicting the next link to be visited by the user, improving web browsing through pre- rendering web pages or by creating a caching system on the server side, predicting the navigation paths consisting of several steps, predicting the websites that are wanted to be accessed using only the first characters printed in a browser, evaluating the possible actions of users in the case of IPTV systems, such as those presented in Patent documents Modeling, Detecting and Predicting User Behavior with Hidden Markov Models, Heck inventor Larry P. (US2009164395A1, published 25.06.2009), System and Method for Auto-sensed Search Help, inventor Watson Eric B. (US2006259861A1, published 16.11.2006), Predicting Use r Navigation Events Based on Chronological History Data, inventors Burkard T, Hamon D and Jain A (US8655819B1, published 18.02.2014), Accelerating User Interfaces by Predicting User Actions, inventors Fredricksen E R., Buchheit P and Rennie J G. (W02006085314A2 , published June 30, 2004), Access to NetWork Content, inventors Burkard T and Jain A (US2014095966A1, published 03.04.2014), Predicting User Navigation Events, inventors Hamon D, Burkard T and Jain A (US8566696B1, published 22.10.2013), Predicting User Navigation Events in an Internet Browser, inventor Hamon D (US8744988B1, published 03.06.2014), A New Book-like Internet Browser for Electronic Information, inventors Seet Chern H and Ho Seng B (W00206917A2, published 24.01.2002), Dynamic Prefetching Method and System for Metadata, Cetin G, Gumus U and Kupusoglu O inventors (EP2194471A1, published 09.06.2010), Web Site Connecting System Using Keyword and '2 0 1 5. - 066072 1 -08-2015

Method Threreby, Park Ho-nam inventor (W02007027002A1, published 08.03.2007), Enabling Predictive Web Browsing, White C A. and Jones C D. inventors (US2012137201A1, published 31.05.2012), User Intention Modeling for Web Navigation, Zheng inventors C, Xiaoming S and Liu W (US2003212760A1, published 13.11.2003), Dynamic Page Generation Acceleration Using Component-Level Caching, Anindya D inventor (US6622168B1, published 16.11.2003), Access to Network Content, Burkard T, Jain A inventors and Bentzel C (US2012324043A1, published 20.12.2012), Adaptive Prefetching for Computer Network and Web Browsing with a Graphic User Interface, inventors Jiang Z and Kleinrock L (US6385641B1, published 07.05.2002), Access to Network Content, inventor Hamon Dominic (US8954524B1 , published 10.02.2015), Web Brower Graphics Management, Eves D and Timms A inventors (W00019336A1, published 06.04.2000), Navigating Organizational Structures, Forman GH and Suermondt HJ inventors (US2003139901A1, published Cat. 24.07.2003), Accelerating User Interfaces by Predicting User Actions, inventors Fredricksen E R., Buchheit P and Rennie J G. (W02006012144A2, published 02.02.2006).

These methods have many disadvantages. A disadvantage is that, in all cases, for the prediction, simplistic methods are used that use little information (bigrame or at most n-grams of page visits) and simplistic assumptions (Naive Bayse) based on which limited information can be derived ( generally only the prediction of the next page or the prediction of a sequence of pages, but with increasing accuracy).

Most of the above methods (with the exception of US2003212760A1) consider only the current page as a browsing history, without referring to the previous browsing history, and this very short history does not allow to obtain a good accuracy. Another disadvantage is that these methods use the information on the current page to predict only the next page that will be accessed and not a certain navigation. Most of the time, the prediction is based on parsing the current page to extract existing links to other web pages and the prediction considers that these links can continue to be browsed, which requires more processing time to extract links from the page and hence another disadvantage. Exceptions are US2003212760A1, US6622168B1. Some of the mentioned methods use sensitive information for clients, which raises trust issues from them, representing another weakness.

In addition, the method used in US2003212760A1, which appears as an exception in the list above, has the disadvantage that it needs to operate a long list of information for which the history (s), its access time, time spent on each page, the way in which it reached a certain number: following a click on a link, through a search using a motor or directly writing the url in the browser, what search engine was used, the words searched, the results obtained, which of the results returned by the search engine was clicked) which means that, when implemented on a mobile device, all this vast information is saved on the mobile device (occupying memory) and then processing is done locally (with impact on the duration battery life and device performance), or, if the information is transmitted on a server, there will be a lot of data traffic, with a negative impact on the service life ab. landing gear and device performance, including browsing speed.

X * —- 1II 15, - Ο Β 6 Ο 7 ->

1 -08-2015

Another major disadvantage is that, often, the prediction for a user is made based on the previous actions of other users, so it is not individualized for each user, but takes into account the actions of the majority.

The technical problem solved by the invention is the acceleration and simplification of navigation, by offering, at the right time, relevant suggestions, easy to access, which is beneficial to both the user and the platform used, but is very important when using mobile devices. (phones, tablets). In this case, accessing pages containing the information you want at a given time can be done by using search engines, but this involves manually entering several terms, and possibly repeating the operation several times, until finding what you are looking for. By the proposed method, this step is eliminated, the user being able to directly receive page suggestions from the predicted category. In addition, by the way it is designed, the method can be easily implemented, in an optimized manner, both on mobile devices (with minimized power consumption and a reduced volume of additional data traffic), as well as on personal computers.

The prediction method, according to the invention, solves the mentioned technical problem and removes the mentioned disadvantages, by providing, with high probability and in an efficient way, the intention of navigating a user, represented by the categories of the web pages to be visited by for this purpose, intended for use on personal computers and mobile devices (phones, tablets), in order to speed up and simplify navigation, by offering, at the right time, relevant suggestions of web pages, easily accessible by the user. In this way, it is possible to preload a series of pages in this category and offer them as options to continue the browsing session, thus allowing to speed up and simplify navigation.

The method of prediction, according to the invention, consists of the following phases:

a. Pre-processing, in which, starting from the user's current browsing sequence and from its browsing history, the imic or more acyclic vectors with the categories of pages visited are created, in order of visits;

b. Processing, in which the category of the next category of web pages to be accessed by the user is predicted. This prediction is done by combining two different prediction techniques: one based on the use of a recommendation process, and the other, a statistical method based on comparing the probabilities of the different categories of pages to be accessed; and using three different factors to make the prediction: the user's browsing history, the current time point and the user's current browsing rate;

c. Adding pre-processed information to the user's browsing history.

The prediction method presented according to the invention has the following advantages:

- It has high accuracy because:

o It is applied to each user individually, being based on the processing of the navigation information specific to it and not by mediating the information coming from a large group of users;

o It combines two different prediction techniques: one based on the use of a recommendation process, and the other, a statistical method based on comparing the probabilities of the different categories of pages to be accessed;

^ '2 0 1 5, - 0 0 6 0 7 - kt 2 1 -08-2015 ^V o Combines three different factors to make the prediction: the history of navigation, the current moment of time and the current pace of navigation.

- Use only non-intrusive, slightly sensitive information, ie browsing time, web page accessed and its category, and not information about gender, age, income, passions, location, etc.

- Requires as input information only the time of navigation, the url of the page accessed and its category, so it allows easy implementation on mobile devices, where the method is implemented on the server and receives this information, without significant impact on data traffic, battery life and mobile device performance.

The invention is presented in relation to the following figures, which represent:

FIG. 1, the main stages of the process covered by this patent;

FIG. 2, the flow diagram of the processing involved by the process covered by this patent;

FIG. 3, the flow chart of determining the probable categories according to different time points;

FIG. 4, the flow chart of determining the probable categories at the current moment of time;

FIG. 5, the flow chart of determining the probable categories according to the navigation rhythm;

FIG. 6, flow chart for determining the average and the deviation of the user's navigation rate;

FIG. 7, the flow chart of determining the probable categories according to the current navigation rhythm;

- fig. 8, the flow chart for determining the probabilities of continuing the sequence started with different categories.

in FIG. 1 shows the architecture of the proposed prediction method, according to the invention. Thus, the processing module 4 will receive, for each individual user, the current navigation sequence 1 containing the pages visited, the time points of the visits and the categories of these pages. The time of visits will help us evaluate two of the factors mentioned above: the time of navigation and the user's browsing pace, while the pages visited give us the user's history, the third factor mentioned.

Due to the fact that the web pages visited by the users have a great diversity, which impedes learning, we will reduce this diversity by replacing the web pages with the categories they belong to, which is why this information is also retained for processing.

in addition, due to the fact that a typical browsing sequence may contain cycles "- a user may return to a specific topic (page category) viewed in the past within the same sequence -, while for learning we need acyclic sequences (we try to learn the index of a certain category in a sequence and as a result it can have only one value), they will be pre-processed 3 to eliminate the eiclicicity.

At processing module 4, in addition to the pre-processed information related to the current navigation sequence, the user's history 2 also comes, containing previous navigation sequences.

^ “2015, -00607- $ (l 1 -08-2015

Based on this information, the method identifies the probable categories to be visited by the user considering the current time Ol, the categories most likely to be visited according to its current browsing rhythm 02, the categories most likely to be visited according to history of navigation 03 and the prediction of the category of the next page to be visited by the customer, taking into account all the factors considered in the analysis 04.

Next, check whether or not the navigation sequence is completed 5. If the navigation sequence has not been detected, the whole process is resumed in order to continue the analysis and prediction 10, 11, 12, 13. otherwise, the sequence must be saved in the user's history. To do this, first determine the space needed to store the current sequence 6. Due to the fact that the memory allocated to this process can be limited, before saving the sequence, check whether there is sufficient space available or not 7.

If enough space is available, then the sequence is added to log 9 and the method ends its operation. If there is not enough space, then a series of entries already existing in history 8 are deleted first to create the necessary space, and then the new sequence is stored in the newly released space 9. The deletion of the sequences has the purpose on the one hand. to keep history 2 within reasonable limits from the point of view of the occupied dimension and, on the other hand, to provide predictions as close as possible to the current behavior of the current user.

Having (for each user) current navigation sequences 1, in processing 4 we will use an independent vector and matrix for each individual user (the vector representing the current navigation sequence 1, and the user history matrix 2). The matrix containing the history of user 2 will have on the two dimensions the following information:

the web page sequences visited by the respective user in the past and - the possible categories for the different web pages.

The information stored in the matrix will represent sequence indexes of the different categories of pages visited. In other words, in the matrix we will capture the order in which the user visited different categories of pages. Due to the fact that in such matrices, only one value can be saved for each tuple (in our case <sequence, sequence category>) while a typical sequence may contain cycles "- a user may return to a particular topic (page category) viewed in the past within the same sequence - we must take care of each sequence in the matrix to be acyclic. As a result, it is necessary to preprocess the initial navigation sequences 1 so that all associated information can be saved (even if there are cycles). Thus, the current navigation sequence 1 is preprocessed and transformed into one or more current navigation vectors. If more such vectors are obtained, it means that cycles have been identified and in this case it is not maintained for processing 4 than the last vector. The other vectors will be saved 9 in the array representing user history 2 if there is enough storage space 7, or if not, before this operation, 8 a series of columns in that history array 2 are removed to make room for the current sequence save 9. Once processing 4 of the current navigation vector is completed, the same thing happens with this vector: 9 is added to the history matrix 2 if possible 7, and if not, then first 8 columns from history 2 are removed and then the respective vector 9 is saved.

“2 O 1 5, - 0 0 6 0 7 2 1 -08- 2015

The pre-processing algorithm 3, realizes, according to the known methods, the transformation of the current navigation sequence into a series of acyclic vectors, containing the categories of the pages visited (obtained from an existing service). Also, three fictitious categories are added (start, continuous, stop) for the purpose of delimiting the sub-sequences.

After pre-processing 3, the vector to be processed is obtained. The details of processing 4 are shown in Fig. 2. Processing 4 involves, on the one hand, information on the current navigation vector, and on the other, information on the user's browsing history 2. Navigation history 2 is used to determine the probable categories to be visited by the user. according to various time points 14, as well as for determining the probable categories to be visited by it according to its navigation rate 15. Once this information is obtained, they are combined with the current navigation vector to determine which are the probable categories. to be visited at the current time of time 10, respectively the probable categories according to the current navigation rhythm 11. Also, the current navigation vector and the user's browsing history 2 are used in a module to determine the probability of continuation of the current sequence. 12.

Once the results have been obtained from the modules for determining probable categories at the current time 10 (Ol), for determining probable categories according to the current navigation rate 11 (02) and for determining the probability of continuation of sequence 12 (03), they are combined for to obtain a single set of values for each category in part 16. finally, based on these values, the prediction of the category of the next page will be made, which will be visited by the user 04 within the category 13 prediction module.

in FIG. 3 is detailed the module for determining the probable categories according to different time points 14. Within it, first a matrix (M) is constructed with the categories of pages per line and the time slots on columns 17 and initialized with 0 each matrix entry 18. In this context, a time slot may have more granularities, depending on the amount of data available and the need for prediction. The user history 2 is analyzed below and for each entry in part 19, the time point at which the respective category 20 was accessed is determined, and then the respective time point is included in one of the defined time slots 21. Based on this classification will increase the element in the matrix (Mjj) corresponding to the category i and the time slot to which it belongs (j) 22. Once the analysis of the entire user browsing history 2 has been completed, the probabilities of each category i can be determined for each time slot j - p (i | j) - (by dividing M; j by Lj Mjj) 23, as well as the probabilities of each time slot j - p (j) - (by dividing Σι Mjj by Ση Mjj) 24.

category i | slot j) =

mg

IjMt, p (, SÎOtj ') ZtMțj l / M ,,

Based on these probabilities, the categories most likely to be accessed at the present time of time Ol are determined in the module for determining the probable categories at the current moment of time 10. The diagram of this process is shown in fig. 4. Once the probabilities of each category i for each time slot j 23 are obtained, these values

A- -2015, -006072 1 -08- 2015 must be filtered so that only those specific values for the current slot remain. As a result, we first determine the time of the last category accessed 25 based on the current navigation vector, and then the current sequence falls into one of the defined time slots 26. Having identified the current time slot, we will extract from the identified probabilities. for each category i and for each time slot j 23 only those probabilities that meet in the current time slot 10. These probabilities are provided as the output of the method Ol.

in FIG. 5 is presented in detail the block for determining the probable categories to be visited according to the navigation rhythm 15. To do this, first a matrix is constructed having on line the categories of available pages, and on the columns the identified navigation rhythms. For example, 4 different browsing rhythms were considered, using the average and the standard deviation (deviation) of the rhythm of the previous sequences of the user calculated on the basis of his / her history 2. However, depending on the needs, a large number of such can be defined of beats, using other metrics besides the two suggested above. The 4 navigation rates considered are:

1. High average time between categories, high deviation -> pause navigation (Type 1)

2. High average time between categories, small deviation -> slow navigation (Type 2)

3. Small average time between categories, large deviation -> chaotic navigation (Type 3)

4. Low average time between categories, small deviation -> accelerated navigation (Type 4)

Once the matrix with the page categories on the lines and the 4 column browsing rhythms is constructed, it is initialized with the value 0 28. The average and the deviation of the user's navigation rate are determined for each entry in each sequence in the browsing history 29. This module will be detailed in FIG. 6. Further, for each entry in the navigation history 30 it must be determined whether or not the respective entry is the first entry in the sequence 31, because for the first entry the navigation rate cannot be determined as it is considered that there is no previous entry and then for we have no way to calculate the average or the deviation of the navigation rate. If it is the first entry 31 yes, then it is passed to the next entry (nothing is calculated) 32. Otherwise, 31 No, the average and the deviation of the current navigation rate of the user 33 already calculated in step 29. Once these values are obtained, they are are compared with 2 thresholds (one for mean 34 and one for deviation 35, 36). The values of the 2 thresholds can be determined heuristically, can be chosen pseudo-randomly or can be determined based on the user's history 2.

If the values are chosen pseudo-randomly, values recognized as valid for different browsing rhythms can be considered (for example, the average time spent on a web page - ie the delay between 2 consecutive pages - is 15 seconds and the deviation is of 3 seconds). The advantage of this method is that it does not require too much knowledge about the user or the pages visited and can apply immediately a new user has been registered. However, such a method cannot work effectively for everyone and all kinds of web pages; it must be adapted to the different situations encountered in practice.

If the thresholds are determined based on user history 2, then the values of the mean and the deviation of the navigation delay can be calculated and then these values are considered as thresholds. The advantage of this method is that the values are very well adapted to the user.

.λ, -2 Ο 1 5 BD 6 0 7 - λ / ω 2 1 -08-2015 ^F depending on the values of the 2 thresholds and the comparisons of the average 34 and the deviation 35,36 with these thresholds, the current navigation rate is established it is part of one of the 4 types and then the value corresponding to the current category and the current navigation rate 37, 38.39, 40 is increased. Further, it is checked whether all entries 41. have been processed. to be processed 41 No, then proceed to the next entry 32 and the process resumes. otherwise 41 Yes you can calculate the probabilities of each category i for each navigation rhythm j (p (i | j)) by dividing the value of My by Zj Mij 42. Also, here you can determine the probabilities of each navigation rhythm j (p (j)) by dividing Day My by Eij My.

category i | browsing rate j)

Mg p (browsing rate j) = = —r / in fig. 6, as already mentioned, describes how to calculate the average and the deviation of the current navigation rate of the user 29. These calculations can be performed considering the entire user navigation sequence or only the last part of this sequence, given that the user's browsing rate may change during a navigation sequence depending on various factors, such as the type of information sought or possible boredom of the user that may occur after a certain time. As a result, before starting the calculation, the amount of information related to the previous navigation used for these calculations is determined, that is, the window from which we extract the information related to the previous visits of the user from the current sequence. This window can be defined either as the entire current sequence of the user (regardless of the rhythm changes that may occur within the sequence), or as the most recent part of this sequence, which can be defined in turn or as a time interval ( for example, visits from the last 30 seconds, the last 5 minutes, or the last hour from the current time, either as the number of page categories visited (for example, the last 3, 5, 10 categories of pages visited that are part of the current sequence ), or as a combination of these two factors (number of page categories visited in a given time frame: for example, the last 10 categories visited in the last 5 minutes). in FIG. 6 is detailed the variant that takes into account a fixed number of categories of pages visited (window size, window_size) to highlight what problems such an approach raises, the other variants representing simple adaptations of this approach. Thus, for the first sequence j (= 0) in navigation history 43, once the window size (window_size) is established, two vectors will be created having a number of components equal to the window size (window_size): t and delta 44. The vector t represents the vector of the times for accessing different categories of pages, while the delta vector represents the time difference between the current and the previous access. Also, in this step, two other m and dev vectors will be constructed, having a number of components equal to the sequence size. M represents the mean vector of the time differences before each entry in sequence j (taking into account the window size and the entries that fit into this window), and dev represents the standard deviations of the time differences calculated for the same entries. Dev will be calculated based on the vectors m and delta, according to the formula:

- ^ - 21 15, -08 6 8 7 2 1 -08-2015 devi =% k (delta _k - m _t ) ² windowjsize -1 where k represents the indices of the entries in the selected window and varies from O to window size -1.

In addition to these vectors, the algorithm for determining the average user's current navigation and deviation 29 also uses three variables: s, represents the sum of the time differences in the window and will be used to calculate the average of the time differences in the window: m; = s / window_size; i, represents the index of the element in the sequence and is used for switching from one entry to another and j which represents the index of the current sequence in the user's history

2.

After the construction of the four vectors (m, dev, t and delta) 44, they are initialized with the value O and the same happens with the values of s and i 45 Q was already initialized in step 43). Next, for each entry i 46, determine the time point at which the respective category was accessed and save this value in t [i% window_size] 47 where operator% represents the operation of obtaining the rest of the division. In other words, as long as its value is smaller than the size of the window, the moments of time are saved even in the order in which they were read. Once its value i reaches the value of the window size (i = window size), it means that the oldest value in the window must be deleted because that entry is no longer part of the window (the indexes in the window are between O and window_size -1 and then when the entry with window_size index is found, it means that the most recent window_size entries are the sequence entries with values between 1 and window_size). As a result, the information related to the current entry must overwrite the data related to the oldest entry in the window (entry with index 0), and this is done using the modulo operator (%), since the rest of the window size division (the current index in the sequence) at window size, it is 0 and so the information about the current input will overwrite the existing information at position 0 in the vector t.

Once the time of the current entry is saved, it is checked whether this entry is the first in sequence 48, because if it is the first entry, the metrics required for it cannot be determined. If it is the first entry 48 Yes, then it goes to the next entry by increasing the index i 53. Otherwise 48 No, it is checked whether or not the value of i is greater than window_size 49. This check is done with the purpose of optimizing the calculation of the value s (sum time differences in the window). Thus, if its value i is greater than the value of window_size 49 Yes some values already saved with the new values calculated as a result of the current entry will have to be updated. One of these values is s (the sum of the time differences in the window). The value of s will have to subtract the first value of the delay and instead add the new value calculated for the current entry. The old delay value is eliminated using the formula s = s - delta [i% window_size] 50. Further, the new delay must be added, but in order to do so, this new value must first be calculated and saved using the delta formula [ i% window_size] = t [i% window_size] -t [(il)% window_size] 51. Once the value of the new delay is taken, it is added to the sum of the time differences in the window (s): s = s + delta [i% window_size] 51. at this time you can calculate

0 15, -00 6 0 7 - (Λ

1 -08- 7015 average and current navigation rate deviation using the formulas above. These values will be saved at position i in the vectors m and dev.

If the value of i is not greater than the value of windowsize 49 No, it means that we will not have to update values, but only to add the new information that is part of the window. As a result, it will skip step 50 and proceed directly to step 51, in which the new delay is calculated and saved as the time difference between the current and the previous entry, and then its value s is calculated and based on it, the values of the mean and the navigation deviation from the current time point. Given that its i values are smaller than window_size, the module (%) does not influence the operations performed in step 51, so that both cases when i is larger than window size and when i is more little.

Once the average values and the current navigation deviation are calculated, one can go to the next entry 52 to repeat the process. For this, it is checked whether or not the current entry was the last one in the current sequence 53. If not 53 No, then the process resumes from step 46.

Otherwise 53 Yes, the current sequence processing is completed and the next sequence to be processed is attempted. In the next step 54 it is checked if the sequence whose processing has just ended was the last one in the history of user 2. If yes 54, then the processing ends as the whole history of user 2 has been processed. otherwise, proceed to the next sequence 55 and then the process is resumed from step 44.

in FIG. 7 is presented the way in which the probabilities of each category of pages are determined, so as to take into account the current navigation rate 02. Thus, once the probabilities of each category i for each navigation rhythm 42 are determined, they are filtered according to the current navigation rate 56, to resume only those probabilities that are adapted to this rhythm 11 (02). Determining the current navigation rhythm 56 is done with the help of the module for determining the average and the navigation deviation of the user 29, with the observation that in this case we have only one sequence (so j = 0 always) and in addition we are only interested in the values the latest values of the mean and the user's navigation deviation.

in FIG. 8 is described the way of calculating the categories most likely to be visited according to the navigation history 03, within the module for determining the probability of continuation of the sequence 12, using the navigation history of user 2. As already mentioned, the information related to The user's browsing history is stored in an array, so that the order in which the user visited various categories of pages is surprised. As a result, the information saved in this matrix will represent sequence indexes of the different categories of pages visited. Using this matrix 2 and the partial navigation sequence 1, it will be determined the probability of each category of pages to be visited further with the help of a recommendation process. In other words, a recommendation process is used which, based on a partial sequence existing at a given time and the known history of a user (his previous browsing sequences), to learn the order in which he visits different categories of pages and then to suggest possibilities for further (partial) navigation sequences. The recommendation procedures are based on the realization of a user-object matrix, where the ratings given by each user to the objects used are usually saved. This matrix is usually rare, since the large number of objects and users makes the existence practically impossible. ^ .- 2 0 15, -B86B7- (d 2 1 -08- 2015 a rating for each tuple <user, object> (which implies for each user to have used and evaluated each of the objects in the matrix.) The classic variant of such a recommendation process has been adapted to take into account the needs / information available in our invention, so that instead of having more users , we will have more browsing sequences, and as objects we will have the page categories identified; in our case, instead of the classic cutilizer tuples, object> we will have tuples <sequence, category from the sequence>. In other words, we will use the matrix of user history 2 having on the two dimensions the web page sequences visited by the respective user in the past and the possible categories for the different p web agini. Using this matrix, what the recommendation process teaches is different from the classic version (where it detects a rating for a particular object). Thus, the referral process actually learns the order in which different categories of pages have been browsed, and can subsequently be used to suggest the order in which various categories for new, unfinished sequences will be visited.

This module takes the current navigation vector (which is incomplete) and the user's browsing history 2 and sends it to a recommendation procedure 57. The recommendation process can be any classic or improved recommendation process based on new discoveries in the field, previously made adaptations. not imposing any restrictions on its functionality, making it possible to use any such process. Next, the recommendation process will provide a value for each previously unseen category (which is not part of the partial navigation sequence) 58, including for the continuous and stop categories additionally introduced in the pre-processing algorithm. This value represents the probable order of each category analyzed in the incomplete sequence given for evaluation.

Knowing how many categories have already been visited in the current incomplete sequence, it is possible to determine, for each category that remains unseen, the difference between the following index of the sequence (length of sequence + 1) - the index on which a new category would be visited and the predicted value by the recommendation process. This difference is transformed into a probability of visitation 59 by dividing it by the total number of categories + 1 and subtracting the value obtained from value 1. This is because the longest sequence that can be built can have a maximum total number of categories +2 elements (start and continue or stop): start (beginning of the sequence), all categories (each one once because otherwise we would have cycle and the sequence would break into two sub-sequences) and then either continuous (it was found cycle) or stop (navigation sequence ended), and the maximum difference of indices in this sequence is the total number of categories + 1. Thus, using this methodology it is possible to determine the probability that an unvisited category of pages will still be the next visited page. However, the problem of estimating the probabilities for all categories of pages is not yet solved, as there are also the categories of pages already existing in the sequence (visited), whose probability of re-visiting must be estimated 60. To evaluate these probabilities two information are used: the probability of the continuous category (the probability of completing a cycle in the next step) and the probabilities of the categories already visited (except for the Start category) of being in the first position in the user's history 2.

In other words, we first estimate the probability of completing the cycle in the next step, and then we must evaluate which category is most likely to close the cycle from those already visited. As the first information is provided by the recommendation process, it remains for us to evaluate the second information. Due to the li algorithm

2075, -00607 -, (7®

1 -08- 2015 pre-processing, whenever a new cycle is identified in the navigation sequence, it is broken into two sub-sequences, the second starting with the category that led to the creation of the cycle. As a result, it is possible to evaluate the probability of each category of generating cycles by determining the number of situations in which this happened and dividing this number by the total number of cycles identified throughout the user's browsing history 2.

Given these probabilities, they are filtered so that only the probabilities of those categories that have already been visited in the sequence remain, after which the obtained values are multiplied by the probability of completing the cycle (determined based on the results of the recommendation process). the multiplication of the two probabilities takes place due to the fact that both events (the occurrence of the cycle and the category that creates the respective cycle) must occur simultaneously and thus their probabilities of occurrence must be combined (through multiplication). After this last step we have all the probabilities of continuation of the current sequence 03 available and these can be provided at the output of module 61.

Once the values of the probabilities of each category are obtained, depending on the factors considered (current time, module 10, output Ol, current browsing rate, module 11, output 02 and user history, module 12, output 03), these values must be combined. so that the process provides a single final probability for each category 16. Basically, the process must provide the probability value of each category and given the current time, current browsing rate and user history: p (category | current time of time, current browsing rate, user history). The combination of probabilities can be done either by directly multiplying the probabilities corresponding to the same category, either by summing them, or by summing their logarithms, or by using a weighted sum so that different importance can be given to each of the three categories of probabilities considered. If the three probabilities are directly multiplied, it is considered that the three factors (current time, current browsing rate and user history) are independent and then the probabilities involved by each of these factors factor, resulting in:

p (category | current time point, current browsing rate, user history) = p (category | current time point) * p (category | current browsing rate) * p (category | user history).

Due to the fact that they work with probabilities having values between 0 and 1, some of them being extremely small, it is possible that by multiplying them they can reach very quickly the value 0 (due to the exceeding of the precision of the computers). That is why, in some cases, it is possible to resort to an artifice: the logarithm in advance of the above formula. Thus, the product is transformed into sum and as a result a summary of the logarithm of probabilities (instead of a direct multiplication of them) can be made to avoid these problems. Sometimes, to simplify the calculations, instead of summing the logarithm of the probabilities, the values of the probabilities are summed, even if the results are not as correct. Also, using the sum (logarithm) of the probabilities instead of their product, one can change the importance given to each factor by adding one coefficient of importance to each factor separately. These coefficients do not appear by chance, their presence being explained by the possible interactions between different factors, interactions that cannot be surprised by the initial model in which the independence of the presented factors was assumed. As a result, if ^ -2015, -00 6 0 7 - $ 5

1 -08- 2015 there are influences between factors, those factors with higher influences will have higher coefficients within the sum. Also, if one does not want to take into account any factor, it can receive the coefficient O and thus its prediction is ignored.

The last step of the application, category 13 prediction, has the role of determining and providing at the output the most probable category taking into account the user's history, the factors that are taken into account (current time, current browsing rate and user history), as well as and any coefficients of importance given to each of these factors. In this step, the values of the probabilities of all categories (obtained from the combination of probabilities module 16) are compared and the one that has the maximum probability 04. The example of applying the proposed method will be provided at the output.

In a non-limiting example of the application of the invention, the proposed method can be used for browsing on a desktop (personal computer) system. In this case, the method will be implemented on the computer itself and will analyze the browsing history of the user, making predictions related to the next / next categories of pages that will be visited by him. For this, the method will be implemented in the browser, for example in a browser extension (written in JavaScript).

The input data is represented by a navigation sequence of the form:

2015-03-02 11:29:11 impbeacon {location: www.google.com, impId: O7be8c3b-3O2f457e-864f-ab4953100636, vertical: Intemet_and_Telecom / Search_Engine, pk: 0, appInstanceUid: 3AA812EC-A5994-38B-9 B747296AB527} ES

2015-03-02 13:49:33 impbeacon {location: www.angrybirdsgames.com, impId: 8ebad827-881a-4f8d-8104-38164fddl689, vertical: Games, pk: O, appInstanceUid: 3AA812EC-A599-438B-9664- B747296AB527} ES

The procedure will determine the probabilities of each category separately for each time slot 14. For example, if you consider 3 time slots (morning, lunch, evening), then the probabilities in Table 1 for the 5 categories of pages considered (Internet_and_ Telecom) will be obtained. , Games, Adult, Finance and Other) + two more (Continue and Stop). Table 1. The probabilities of each category separately for each time slot

^ \ Slot category owns ^ Î Internet_and_Telecom Games Adult Finance Other continue Stop Morning 0.46 0.23 0 0.3 0 0.31 0.06 Lunch 0.38 0.08 0 0.15 0.38 0.44 0.23 In the evening 0.23 0.3 0.15 0 0.3 0.25 0.69

Also, the probabilities of each category will be determined for each browsing rhythm of user 15, obtaining the values in Table 2.

Finally, given a partial sequence containing the categories Start, IntemetandTelecom and Games, using the procedure of recommendation 12, the following indices will be determined for the remaining remaining categories: Adult - 5, Finance -4, Other - 3, Continue - 3, Stop - 5.

fu-ι O 15 D16 D 7 2 1 -08-2015

The probabilities obtained for each of these categories are Adult - 0.66, Finance -0.83, Other -1, Continue -1, Stop - 0.66. Also, for the 2 categories that can determine the cycle, the probabilities of IntemetandTelecom - 0.8 and Games - 0.2 are obtained. Then output 03 will be: Internet_and_Telecom - 0.8, Games - 0.2, Adult - 0.66, Finance -0.83, Other -1, Continue -1, Stop - 0.66.

Table 2. The probabilities of each category separately for each time slot

Category Rhythm navigaîe \ ^^ Internet_and_ Telecom Games Adult Finance Other continue Stop Tiplic 0.17 0.67 0 0 0.17 0.13 0.38 Type 2 0.43 0.14 0 0.14 0.29 0.31 0.31 Type 3 0.31 0.15 0 0.23 0.31 0.38 0.15 Type 4 0.46 0.08 0.15 0.15 0.15 0.19 0.15

Knowing that current browsing is done in the lunch slot, the Ol output of the process will be represented by Intemet_and_Telecom values - 0.38, Games - 0.08, Adult - 0, Finance -0.15, Other - 0.38, Continue - 0.44, Stop - 0.23 . If the user's navigation rate is of type T2, the output 02 of the procedure will consist of the values: Intemet_and_Telecom - 0.43, Games - 0.14, Adult - 0, Finance -0.14, Other - 0.29, Continue - 0.31, Stop - 0.31.

In order to obtain the final probabilities, the values representing the outputs Ol, 02 and 03 must be multiplied and then normalized the obtained values. As a result, the final odds will be: Intemet_and_Telecom - 0.294, Games - 0.005, Adult - 0, Finance - 0.039, Other - 0.248, Continue - 0.307, Stop - 0.106. Based on these values, it is observed that we will obtain the maximum value for the Continue category (30.7%), which means that the user will most likely return to one of the two categories already visited. Of the two categories, the one with the highest probability (close to that of the Continue category) is the Internet_and_Telecom category (29.4%) and as a result will be chosen as the category of the next page that will be visited by the user, this being output 04.

Example 2 application of the proposed method. In another non-limiting embodiment of the invention, the application may be used on a mobile device (telephone or tablet). Compared to the version on the personal computer, now the data are no longer processed locally, due to the limited resources available to the mobile device, but are sent to a server where the processing presented in the previous example is done. The mobile device browser will collect the browsing information (same as in the previous example) and then send it to a server, for example using the Java Script Object Notation, JSON format. The server can be deployed on a single machine, in which case the transmission mechanism on the server uses, for example, PHP language and a web server (for example, Tomcat), or it can be deployed in a cloud, if decides to keep more data for prediction (and where implicitly more computing power is needed to process them). In this case the transmission module on the server is implemented, for example, in Scale. At the end of the processing, the most probable category (for example, as in Example 1 Intemet_and_Telecom) is resumed on the mobile device, where it is used in conjunction with another module for pre-loading pages of the predicted category.

Claims (4)

1. Method of individual prediction of the categories of web pages to be visited by a user, intended for use on personal computers and mobile devices (phones, tablets), in order to speed up and simplify browsing, by offering, at the right time, relevant web page suggestions, easy to access by the user, characterized by having the following phases: a. Pre-processing 3, in which, starting from the user's current navigation sequence and from its browsing history, one or more acyclic vectors with the categories of pages visited, in order of visits; b. Processing 4, in which the category of the next web page to be accessed by the user is predicted. This prediction is done by combining two different prediction techniques: one based on the use of a recommendation procedure 12, and the other, a statistical method based on comparing the probabilities of the different categories of pages to be accessed 10.11; and using three different factors to make the prediction: the user's browsing history, the current time point and the user's current browsing rate; c. Adding pre-processed information to the user's browsing history. 2. The prediction method according to claim 1, characterized in that the user's current navigation rate is determined, within the processing module 4, by the method of determining the probable category according to the navigation rate 11, based on the mean and speed deviation. navigation, as belonging to one of the following 4 categories: pause navigation, slow navigation, chaotic navigation, and accelerated navigation. 3. The method of prediction based on a recommendation method, according to claim 1, characterized in that, in the processing module 4, the method of determining the probability of continuation of the sequence 12 takes into account the entire browsing history of the user and can generate predictions for an entire browsing sequence (not just for the next category to be accessed) along with the method of pre-processing the data 3 so that it can be used in such a recommendation process. 4. The prediction method according to claim 3, characterized in that the method of determining the probability of continuation of sequence 12 uses the transformation of the results provided by the recommendation process into probabilities: o for the unseen categories, by estimating the difference between the order provided by the recommendation procedure and the order number of the next category that should be visited within the partial sequence, and normalizing this difference by dividing by the maximum difference that can occur (the total number of categories + 1 ) o for the re-visited categories, by multiplying the probability of the Continue category (determined using the methodology above) with the probability of each category already visited to generate cycles.