FR3078795A1 - NEURAL NETWORK SYSTEMS AND METHODS FOR APPLICATION NAVIGATION - Google Patents

Abstract

The present disclosure relates to systems and methods for application navigation using neural networks. The disclosed systems and methods may perform operations including providing an application including application states that work on data objects to generate pages, determine a current application state, and a current data object that predict a next page using a predictor-state neural network of an application and a predictor neural network-data object, and the provision of instructions to display an indication of the next predicted page. The next page prediction may include predicting a next application state using the prediction-application state neural network and a first input vector, and predicting a next data object using the neural network. predictor-of data object, a second input vector and the next application state.

Description

NEURAL NETWORK SYSTEMS AND METHODS FOR APPLICATION NAVIGATION

TECHNICAL FIELD [001] This revelation relates generally to the use of neural networks for application navigation. In particular, this revelation concerns the use of multiple neural networks to predict the next page of an application.

BACKGROUND [002] It may be difficult to navigate in an unfamiliar application. A user looking to perform tasks may not know how to access the necessary interfaces of the application. An untrained user can perform a task in a way that increases the time, number of steps, or the probability of an error involved in implementing the task. Existing processes to guide users through an application may prove to be ineffective and incorrect. For example, training guides may reflect how developers think a user should interact with an application, which may differ from how skillful users actually interact with the application. And training guides can quickly become obsolete as the app is updated or the context in which the app is used changes.

[003] There is therefore a need for an application navigation system which can reflect the way in which users in fact interact with an application, and which can continue to provide correct recommendations when the application is updated or when the context, in which the application is used, changes. According to an application development mode, the applications (or the components of the applications) can be built from application states which work on data objects to generate the pages which

-2 are displayed to a user. Application states can define the appearance and functionality of pages, while data objects can store actual data. As disclosed herein, systems and methods for application navigation can exploit this mode of application development to provide improved application navigation systems that overcome the shortcomings of existing approaches.

SUMMARY The embodiments of this disclosure describe systems and methods for application navigation using neural networks. These neural networks can be linked, and can be configured to predict a page of the application by predicting the state of the application and a data object, where the predicted data object could be managed by the state of the predicted application to generate the predicted page. Prediction may depend on a history of previously visited application pages.

An embodiment of the present disclosure may include at least one processor and at least one non-transient memory containing instructions. When executed by at least one processor, the instructions can cause the system to implement operations. Operations may include providing an application that includes application states that work on data objects to generate pages. Operations can also include predicting a new page using an application state predictor neural and a data object predictor neural network. Prediction may further include predicting a next application state using the application state predictor neural network and a first input vector, and predicting a next data object using the predictor neural network data object, a second input vector and the next application state. Operations may include

-3 also to provide instructions for displaying an indication of the next predicted page.

[006] In some embodiments, an output layer of the application state predictor neural network can include output nodes corresponding to the application states.

In various embodiments, predicting the next application state using the application state predictor-neural network and the first input vector may include hashing the first input vector. In some embodiments, common elements of the first input vector and the second input vector can be hashed once and be reused to predict the next data object. In various embodiments, the first input vector and the second input vector can be the same vector. In some embodiments, an output layer of the data object predictor neural network may include output nodes corresponding to the data objects. In various embodiments, the first input vector may include a history of the application. In some aspects, the history of the application includes elements indicating previous states of an application and previous data objects associated with the pages visited previously. In certain embodiments, the first input vector can include elements indicating at least one of an entity, a user associated with the entity, a role of the user, a level of authorization of the user, a day of the week, a date or a time. In various embodiments, the first input vector may include an element indicating a value of an actual data object.

In some embodiments, the operations can also include the generation of a current page according to a current application state using a real data object. In some embodiments, providing instructions for displaying the indication of the next predicted page may include providing instructions for displaying the current page, modified to indicate the next predicted page. In some modes of

Realization, the provision of instructions for displaying the indication of the next predicted page may include the provision of instructions for dynamically generating a current page in accordance with an actual application state using a real data object, the current page modified to indicate the next predicted page.

In some embodiments, providing instructions for displaying an indication of the next predicted page includes providing instructions for displaying indications of multiple predicted next pages including the next predicted page. In various aspects, the instructions for displaying the indications of the multiple predicted next pages may include the instructions for indicating the relative probabilities of the multiple predicted next pages. In some embodiments, providing instructions to display the indication of the next predicted page may include providing instructions to dynamically refresh a current page, to include a graphical element indicating the next page. predicted. In various embodiments, the graphical element is an inserted window including an icon which corresponds to the next predicted page. In some ways, the icon can be selectable to transition to the next page. In various embodiments, the provision of instructions for displaying the indication of the next predicted page may include the provision of instructions for dynamically updating a current page, by modifying an existing graphic element of the current page in order to indicate the next predicted page. In certain aspects, the modification of the existing graphic element of the current page includes at least one change among the positioning, the shape, the size, the color of the existing graphic element; or at least one change among the positioning, the font, the size, an accentuation or the color of text associated with the existing graphic element. In some embodiments, the application may be a single application page.

BRIEF DESCRIPTION OF THE DRAWINGS The accompanying drawings illustrate several embodiments and, together with the description, serve to explain the principles disclosed herein. In the drawings:

FIG. 1 describes a system for application navigation including a client and a server providing an application.

[011] FIG. 2A describes, by way of example, a schematic application including application states.

[012] FIG. 2B describes, by way of example, the illustrative schematic generation of a page by an application state of an application using a data object.

[013] Figs. 3A and 3B illustrate exemplary user trajectories through the pages of an application.

[014] FIG. 4 describes a prediction element configured to use linked neural networks to predict a next page of an application.

[015] FIG. 5 describes an exemplary input vector.

[016] FIG. 6 describes a sequential flow illustrating an exemplary method 600 for forming a component of a prediction.

[017] FIG. 7 describes a flow diagram illustrating a method of application navigation using neural networks.

[018] Figs. 8 to 10 describe exemplary user interfaces for an asset management application.

[019] FIG. 11 is a block diagram of an exemplary device with which the systems, methods and devices of the present invention can be implemented.

-6 DETAILED DESCRIPTION [020] Exemplary embodiments will now be discussed in detail with reference to the accompanying drawings. In some cases, the same reference numbers will be used in the drawings and in the following description to refer to the same or similar parts. Unless defined otherwise, technical and / or scientific terms have the meaning commonly understood by a person skilled in the art. The disclosed embodiments are described in sufficient detail to allow those skilled in the art to implement the disclosed embodiments. It will be understood that other embodiments can be used and that changes can be made without departing from the scope of the disclosed embodiments. Consequently, the supports, methods and examples are only illustrative and do not necessarily claim to be exhaustive.

[021] A computer system can provide an application to users. This app can include multiple pages. A page can be a user interface displayed to the user, including data and graphics. Graphic elements may include controls and indicators. As a non-exhaustive example, a page can be a web page. As a further example, a page can be a view (eg, a representation of information displayed to the user) according to a model-view-controller design mode or a similar software design mode (eg mode model-view-adapter design or model-view-model view mode). These pages can allow a user to read information or enter information or commands. Different pages can be used to read or enter different information or commands. The application can allow a user to perform tasks by reading and / or entering information or

- 7 commands using the pages. In some cases, accomplishing a task may require interaction with multiple pages of the application. As a non-exhaustive example, a user can retrieve information from a database using a first page, interact with the application to generate a second page, and then enter data or commands based on the information retrieved using the second page.

[022] Users may find it difficult to accomplish a task using such an application. A user may not be familiar with the application and / or the task at hand. As a result, they may not know how to navigate the application to the correct pages to complete the task. Alternatively or in addition, there may be a preferred way of performing the task, which reduces the time, the number of steps, or the probability of error. This preferred way may include visiting a particular set of pages in a particular order. A user unfamiliar with the application and / or the task may not know this preferred way of performing the task. Instead, the user can use another way of performing the task that takes longer, requires more steps, or has an increased likelihood of error.

The methods and systems envisaged can provide recommendations for using an application. For example, the methods and systems envisioned may infer the task that a user is trying to perform and recommend a next page to perform this task. This next page can be a page in a preferred way of executing the task. In certain aspects, the methods and systems envisaged can provide a recommendation by modifying a page of the application currently displayed. For example, the page currently displayed can be modified to include an additional window. This additional window can include one or more controls. Interaction with these controls may cause the application to transition to a recommended page. As

-8further example, if the interaction with an element of the currently displayed page causes the application to display the recommended page, then this element can be modified (e.g. by changing the placement, shape, size or the color of the item; or by changing the placement, font, size, accent, or color of text associated with the particular item). This change can draw the user's attention to the item, acting as a recommendation. In this way, the methods and systems envisaged can recommend actions making a transition from the currently displayed page to the recommended page.

[024] Artificial neural networks can be used to generate the recommendation. In some embodiments, the methods and systems contemplated may divide the overall prediction process into several subsidiary steps, each performed by a neural network in a chain of linked neural networks. As a non-exhaustive example, the recommendation process can be divided into two stages. In a first step, a first neural network can recommend a state based on the first data entry. In a second step, a second neural network can recommend a data object based on the recommended state and the second data entry. These artificial neural networks can be trained using real data from the application. In certain embodiments, the first and second data entries can be the same data entry.

The methods and systems envisaged are improved on the basis of the manuals, assistants, tutorials and other existing recommendation systems. For example, the processes and systems being considered may provide updated recommendations that follow changes in usage patterns or changes to the application. In addition, the recommendations are not limited to situations anticipated by the developers of the application. In contrast, textbooks, assistants, tutorials, and

- 9 other existing recommendation systems are made obsolete by changing usage patterns and application updates, and are limited to situations anticipated by developers.

[026] These improvements are enabled, at least in part, by the specific architecture disclosed here. The universe of the potential pages of an application can be extremely important, limiting the applicability of conventional neural networks. But the systems and methods envisaged use linked neural networks to divide the global prediction problem into smaller sub-problems. This allows for precise recommendations using artificial neural networks trained with actual usage data.

[027] FIG. 1 illustrates a system 100 for application navigation, comprising a client and a server providing an application over a network, compatible with the embodiments disclosed. Client 101 can be a user's computing device (eg, portable device, mobile device, laptop, landline, workstation, terminal, etc.) Client 101 can be configured to run a web browser (eg CHROME®, OPERA®, MOZILLA®, INTERNET EXPLORER®, SAFARI®, EDGE®, etc.) Server 103 can be a computer device configured to respond to requests from client 101. The network can be any type of network (including infrastructure) which provides communications, information exchange, and / or facilitates the exchange of information, such as the Internet, a Local Area Network or others appropriate connections. The client 101 can be configured to request a page from the server 103 in step 105. In response, the server 103 can be configured to provide one or more files describing the layout and content of the page in step 107. client 101 can then generate this page.

The client 101 or the server 103 can generate updates of the page based on the information exchanged during steps 109 and 11L. In certain embodiments, the

-10client 101 can contact the server 103 to request a new page in step 109. For example, a web browser running on the client 101 can send a POST request to the server 103. The server 103 can be configured to responding to this POST request, in certain embodiments, by generating a new page and by sending this new page to the client 101 in step 111. The client 101 can then load this new page. In various embodiments, the client 101 can dynamically update the page without loading a new page from the server 103 (eg, the application can be a single page application). In such embodiments, the application can be implemented using a JavaScript® web browser structure such as ANGULARJS®, EMBER.JS®, METEOR.JS®, EXTJS®, REACT® or similar structures known to man. business. The web browser running on client 101 can request the data and / or instructions in step 109 (eg using an AJAX® call). The server 103 can be configured to respond with data and instructions (eg, as JSON® objects) in step 111. The client 101 can then dynamically update the page using the instructions and data received.

[029] FIG. 2A illustrates an exemplary diagram of a request including the states of the application (eg state 201 - state 215), compatible with the embodiments disclosed. As noted, the application states can be linked, so that a user can transition from one application state to another. For example, a user can start the application 120 in the application state 201, make a transition from the application state 203 (e.g. by interacting with a graphic element in a page generated by the state 201), transition from application state 205 (eg, through another interaction), then return to application state 201. In this way, the states of the linked application can form one or more graphics. In certain aspects, the states of the application can make the application (or one or more components of the application).

- 11 [030] FIG. 2B describes an exemplary diagram illustrating the generation of page 235 by the state of the application 233 of the application 120 using the data object 231. In certain embodiments, the states of the application can define the appearance and functionality of the pages. For example, an application state can define one or more page templates, including markup language and script commands. In some aspects, each application state can have a unique identifier. This identifier can be a text string (eg "application / directory / dostuff" or "dostuff_state") or a numeric value. An application state (eg, application state 233) can take a data object (eg, data object 231) as an input. In various aspects, the data object can be chosen from a given set of data objects. Each data object can have a unique identifier. This identifier can be a text string (eg "re: do stuff" or "root / dir / identifier") or a numeric value. The application state can work on the data object selected to generate the page. As a non-exhaustive example, an application state defining a page template can be configured to apply the data object to the page template, by executing script commands using one or more values of the data object and presenting the results according to the markup language.

[031] Figs. 3A and 3B illustrate the exemplary user trajectories through the pages of the application 120, compatible with the embodiments revealed. As previously revealed, client 101 can be configured to generate these pages using application states and data objects. For example, client 101 can be configured to display page 501 when a first application application state 120 operates on a first data object. In this example, client 101 can be configured to display page 505 when a second application application state 120 operates on the first data object. And client 101 can be configured to display page 503 when the first

-12 application application state 120 operates on a second data object. Since pages can depend on both the application state and the data object, the application state graph and the page graph may differ.

[032] Users can make transitions between pages when they interact with the 120 application. For example, as shown in Figs. 3A and 3B, a user can transition from page 301 to page 303, or to page 305. The pages can include one or more graphical elements allowing the user to transition to another page. For example, a page may include a drop-down menu, address labels, an address bar, or other appropriate controls known to those of skill in the art. Since users performing different tasks may interact differently with the application 120, the system 100 may use transitions between pages to predict the next page for a user. For example, as shown in Fig. 3A, when the page 303 is the active page and the page 301 the previous page, the system 100 can estimate that the next page should be the page 307. On the other hand, as indicated in FIG. 3B, page 303 is the active page and page 305 the previous page, the system 100 can estimate that the next page should be page 309.

[033] FIG. 4 illustrates a prediction component 400 configured to use the linked neural networks to predict a next page of an application, compatible with the embodiments revealed. In some aspects, a prediction component 400 can be configured to divide the prediction of the next page into two sub-problems. The prediction component 400 can be configured to use the state predictor 420 to predict one or more future application states based on a first input vector. The prediction component 400 can be configured to use the data predictor 430 to predict one or more next data objects operated by the one or more next states based on the one or more next application states and a second input vector.

- 13 In some embodiments, the first and second input vector can be the same (eg, input vector 410, as shown in Figure 4). The system 100 can be configured to predict the page based on the predicted application state and the predicted data. In some embodiments, the prediction component 400 can be configured to generate multiple page predictions. For example, the prediction component 400 can be configured to generate a page prediction for each predicted data object and application state.

[034] In certain embodiments, the prediction component 400 can be a component of the system 100. For example, the system 100 can be configured to implement the application 120 and the prediction component 400. As a further example, the component 400 can be an application component 120. In various embodiments, the client 101 and / or the server 103 can be configured to implement the prediction component 400. In certain aspects, the client 101 can be configured to predict the state of the application and / or data object. In various aspects, the server 103 can be configured to predict the state of the application and / or the data object. In some aspects, the client 101 and the server 103 can interact to predict one or more application states or the data object. For example, the client 101 can provide at least part of the first entry or the second entry to the server 103. The client 101 can then receive from the server 103 one or more states of the predicted application, the data and the page.

The input vector 410, described in more detail below, can include several elements, compatible with the embodiments revealed. These elements are not limited to a particular data type, but may include, without limitation, objects, arrays, strings, characters, floating point numbers, or other data types.

[036] The state predictor 420 can include a neural network which generates the state output 423 from the input vector 410. The neural network can include an input layer, one or more hidden layers and an output layer. The elements of the input vector 410 can correspond to the nodes of the input layer. In some embodiments, the input layer may include between 5 and 100 knots, or between 20 and 50 knots. In various embodiments, the neural network can include a single hidden layer. The hidden layer (s) can include between 5 and 50 knots. According to methods known to those skilled in the art, the nodes in each of the one or more hidden layers and the output layer may have values dependent on the weights and values of the nodes in a previous layer. The output layer may include nodes corresponding to at least some of the application states 120. The nodes in the output layer may have probability scores which indicate whether the application state corresponding to the node is the state next application. In some respects, probability scores can indicate a probability that the corresponding state is the next state of application. In various aspects, the highest score can correspond to the most probable state, the next highest score can correspond to the next most probable state, etc.

[037] The state feature extractor 421 can convert the input vector 410 to a value for each of the nodes in the input layer, compatible with the disclosed embodiments. In various aspects, the state characteristic extractor 421 can be configured to calculate a hash on the value of each element in the input vector. The state feature extractor 421 can assign the hash value for the corresponding node in the input layer. For example, when the element is a URL (eg "StateA / StateC / StateE") then the hash of the URL (eg 8858d721629ce6f9c8ff608ce5cff8dl for MD4) can be assigned to the corresponding node in the layer of Entrance. In some embodiments, hashing an element can

- 15 be calculated on the raw binary values of the element. In some aspects, only part of this hash (eg, a subset of the hash bits) can be assigned. The hash function may be the Fowler-Noll-Vo hash function, or another hash function known to those skilled in the art. Assigning the hash value to the node can include converting the hash value to a floating point number with a predetermined number of bits. This floating point number can be normalized within a predetermined range (eg, 0 to 1).

[038] State output 423 can be configured to select one or more predicted application states. In some embodiments, the selected application state can be one or more of the most predicted application states. For example, state output 423 can select a predetermined number of the most predicted application states. As indicated above with regard to FIG. IB, an application state can have an identifier, such as a text string (eg "StateA.StateC.StateE" or "StateA / StateC / StateE") or a numeric value. In some embodiments, the state output 423 may include the identifier (s) for the one or more predicted application states.

[039] The data predictor 430 can generate one or more predicted data objects for each predicted state received from the state predictor 420. The data predictor 430 can include a neural network that generates the data output 433. The neural network may include an entry layer, one or more hidden layers, and an exit layer. In some embodiments, the input layer may include between 5 and 100 knots, or between 20 and 50 knots. In various embodiments, the neural network can include a single hidden layer. The hidden layer (s) can include between 5 and 50 knots. The neural network can generate data output 433 using one or more state output elements 423 and an input vector. The input vector can be the input vector 410 (such

-16 as shown) or another vector (eg a vector that differs from the input vector 410 of one or more elements). The elements of the input vector can correspond to the nodes of the input layer of the neural network. According to methods known to those skilled in the art, the nodes in each of the one or more hidden layers and the output layer may have values dependent on the weights and values of the nodes in a previous layer. [040] As indicated above with regard to FIG. 2B, application 120 can be configured to select data objects (eg, data object 231) from a set of data objects. In some embodiments, the output layer of the data predictor 430 may include a node corresponding to each object in the set of data objects. The nodes in the output layer of the predictor 430 can have probability scores that indicate whether the data object corresponding to the node is the next data object. In some respects, probability scores can indicate a probability that the corresponding data object is the next data object in the application. In various aspects, the highest score may correspond to the most likely data object, the next highest score may correspond to the next most likely data object, etc.

[041] The object data characteristic extractor 431 can convert the input and output vector of a state or several states of the state predictor 420 to a value for each of the nodes in the layer of input, compatible with the disclosed embodiments. In some aspects, the input layer of the data predictor 430 includes input nodes corresponding to the output of one state or more states of the state predictor 420 and the elements of the input vector 410. In various aspects, the data object feature extractor 431 can be configured to calculate a hash on the value of each element and assign the value of the hash to the corresponding node in the input layer. In some embodiments, the hash of an element can be calculated on the values

- 17 raw binaries of the element. In some aspects, only part of this hash (eg, a subset of hash bits) can be assigned. In some aspects, the hash function can be the Fowler-Noll-Vo hash function. Assigning the hash value to the node can include converting the hash value to a floating point number with a predetermined number of bits. This floating point number can be normalized within a predetermined range (eg, 0 to 1).

[042] In certain embodiments, when the input vector for the state predictor 420 shares common elements with the input vector for the data predictor 430, the prediction component 400 can be configured to hash these common elements once, and then assign the values obtained to the corresponding nodes in the input layers of both the state predictor 420 and the data predictor 430. For example, when the input vector for the data predictor 430 is equal to the input vector 410, the prediction component 400 can be configured to hash the input vector 410 once, and to reuse the hashed values of the input vector 410 as inputs for the data predictor 430.

[043] Data output 433 can select one or more predicted data objects. In some embodiments, the selected data objects may be the most likely predicted data object or multiple predicted data objects. For example, data output 433 can select a predetermined number of the most likely predicted data objects. In some embodiments, the data object may have an identifier, such as a text identification string (eg, "2 Seaport Lane") or a numeric identification value. In some embodiments, the data output 433 can include the identifier (s) for one or more predicted states.

[044] The predicted page 440 can indicate the predicted page based on the entry vector 410 (and, in certain embodiments, based in addition on a second entry of

- 18 vector which differs from the data predictor 430, as described above). In some embodiments, the predicted page 440 may include one or more state pairs and data object pairs. For example, predicted page 440 may include the most likely application state and the most likely data object. As described above, when the state predictor 420 exits multiple state objects, the predicted page 440 may include a data-state pair for each of the state objects and the most likely data object for this state object. Likewise, when the data predictor 430 exits multiple state objects, the predicted page 440 may include a data-state pair for each of the state objects and for each of the data objects for that state object. In some embodiments, the predicted page 440 may be based on one or more pairs of application states and data objects. For example, the predicted page 440 may include data or instructions for creating, or dynamically updating, the page 235 to indicate the predicted page. For example, instructions or data can configure the application 120 to display a link on page 235 to the predicted page, or modify a graphic element on page 235 to indicate that interaction with this element would bring the application 120 to transition to the predicted page.

FIG. 5 illustrates an exemplary input vector 410, compatible with the embodiments revealed. As shown in Fig. 5, the input vector 410 can include the current data values 510, the application context 520 and the application history 530. As indicated above, the system 100 can be configured to predict one or more pages following by applying the input vector 410 to the input predictor 420 and by applying the input vector 410 (or an input vector which differs in some embodiments) to the data predictor 430.

[046] Current data values 510 may include one or more elements with values based on a current data object. For example, the application can be in a

- 19current state. In this current state, the application can be managed under a current data object. In some respects, one or more values can be associated with this current data object. For example, this current data object can have one or more parameters. These parameters can take values. For example, an email subject can include a sender setting, a recipient setting, and a content setting, each of which takes a value.

[047] The application context 520 can include one or more elements indicating a context in which the user interacts with the application 120. In certain embodiments, the application context 520 can also specify who uses the application 120. For example, when server 103 provides application 120 to multiple different entities (eg, different companies or individuals), the user can be associated with that entity (eg, the user can be a employee of the entity). The application context 520 can then include one or more elements indicating the entity associated with the user. In some aspects, the application context 520 may include one or more elements indicating at least one of the user associated with the entity (e.g. a user name), a role of the user ( eg "Associate," "Intern," etc.), or user authorization (eg, "User," "Admin," "Root," etc.) In some embodiments, the application context 520 can also specify when the user interacts with application 120. For example, application context 520 can then include one or more elements indicating at least one of the days of the week, the date or the time of interaction.

[048] The application history 530 can indicate the trajectory of the user through the application 120. In certain embodiments, the application history 530 can indicate the pages most recently visited by the user. . The 530 application history can be limited to a predetermined number of most recently visited pages. This predetermined number can vary between 1 and 20, or between 5 and 15. The 530 application history can

- be structured as a queue, with the indication of the most recently visited page in the initial position and the indication of the least recently visited page in the last position. When a new page is visited, the least visited page recently left the t

last position and an indication of the new page is pushed to the initial position. In some embodiments, elements of the 530 application history can be initialized to default values (eg, a value of zero or "null"). When the user interacts with the application 120, these default values are gradually replaced. In this way, the application history 530 can document the user's path through the application 120.

[049] In some embodiments, the application history 530 can indicate each page with a pair of elements: an element indicating an application state (eg the status indication 531) and a element indicating a data object under which it is managed by the application state to generate the page (eg the data object indication 533). In some aspects, an application state may not work under a data object to generate a page (eg a menu or an initial application start screen may not take a data object). In these aspects, a default value (eg, a value of zero or "null") to be assigned to the data object indication 533 to indicate that the application state has not taken a data object. As previously described, the application 120 can be configured to assign unique identifiers to reports and data objects. These identifiers can be text strings or numeric values. An item in application history 530 indicating a state (eg, state indication 531) may include such a unique identifier, or a value based on the unique identifier. Likewise, an item in application history 530 indicating a data object (eg, data object indication 521) may include such a unique identifier or a value based on the unique identifier.

[050] FIG. 6 describes a flow diagram illustrating an exemplary method 600 for forming a prediction component 400, compatible with the embodiments revealed. In some embodiments, the method 600 may include the steps of receiving training data, forming a state predictor, forming a data object predictor, and providing the prediction model for the application 120. As one skilled in the art would recognize, this particular sequence of steps is not intended to be restrictive. The steps of method 600 can be combined or divided, and the method 600 can include additional steps or fewer steps, without departing from the embodiments envisaged.

[051] After starting in step 601, the training system can be configured to receive training data in step 603. In some embodiments, the training data can include both information for predicting the next page and information to identify the actual next page. For example, training data can include data entry in state predictor 420 (e.g., content of input vector 410), input data in data predictor 430 (if differ from the content of the input vector 410), an identifier for the next real state, and an identifier for the next real data object.

[052] In some embodiments, the training system can be configured to receive training data from a system that provides the application 120 to users. For example, the system 100 can be configured to store and / or constitute a record of the interactions of the user with the application 120 in the training system. For example, the client 101 and / or the server 103 can be configured to provide such registration. In some aspects, the training system can receive training data directly from the system that provides the application 120 to users. In various aspects, the training system can indirectly receive data from

-22formation. For example, the training system can receive training data from a database. This database can in turn receive training data from the system that provides the application 120 to users. In certain aspects, the training system can receive the training data, this training data being generated. In various aspects, the training system can receive training data repeatedly or periodically.

[053] After step 603, the training system can be configured to form a state predictor in step 605. In some embodiments, the training system can be configured to use the training data to generate predicted states. In some aspects, the training system can be configured to generate input vectors and apply these input vectors to a state predictor neural network (eg, state indicator 420). The state predictor neural network can be configured to generate application state predictions based on training data. The training system can be configured to compare these application state predictions to the actual application states chosen by the user to generate an error signal. According to methods known to those skilled in the art, the training system can use this error signal to update the weights of the nodes of one or more hidden layers and the output layer. According to methods known to those skilled in the art, the training system can be configured to generate application state predictions from the training data, to generate error signals using the state states. predicted and actual application, and to update the weights until one or more training criteria are met. In some aspects, the training criteria may include a precision criterion based on (i) the number of times the predicted state corresponds to the actual state and (ii) the number of entries in the training data. In various aspects, the training criteria can include a criterion of mean square error based on a comparison of the state

-23 predicted and current state, weighted by the number of application states and the number of observations in training data.

[054] After step 603, the training system can be configured to form a data object predictor in step 607. In some embodiments, the training system can be configured to form a data predictor state in step 605 and then forming a data object predictor in step 607. In various embodiments, the training system may alternate between forming a state predictor and forming a data predictor. For example, the training system may receive a batch of training data, train the state predictor using the batch of training data, and then train the data object predictor using the output of the training prediction. status and batch of training data. The training system can then repeat this training process with the next batch of training data.

[055] In some embodiments, the training system can be configured using the training data to generate the predicted data objects. In some aspects, the training system can be configured to generate the input vectors and apply these input vectors to a neural data predictor network (eg, status indicator 430). The neural data predictor network can be configured to generate predictions of data objects based on training data. The training system can be configured to compare these predicted data objects to the actual data objects chosen by the user to generate an error signal. According to methods known to those skilled in the art, the training system can use this error signal to update the weights of the nodes of one or more hidden layers and the output layer. According to methods known to those skilled in the art, the training system can be configured to generate predictions of data objects from the training data, to generate error signals using the actual data objects, and to set at

-24 days the weights until one or more training criteria are met. In some aspects, the training criteria may include a precision criterion based on (i) the number of times the data object corresponds to the actual state and (ii) the number of entries in the training data. In some aspects, the training criteria may include a mean square error criterion based on a comparison of the predicted data object with the actual data object, weighted by the number of data objects in the predetermined set data objects and the number of observations in training data. [056] After step 605 and step 607, the training system can be configured to provide the state predictor and the data object predictor to the system 100 in step 609. For example, the system can provide the state predictor and the data object predictor to the client 101 and / or the server 103. In some embodiments, the training system can provide the state predictor and the object predictor data to system 100 as objects (eg data structures). In various embodiments, the training system can be configured to supply the weights for the state predictor and the data object predictor to the system 100. In various embodiments, the training system can be configured to provide updates to an existing state predictor and the system object data object predictor 100.

[057] In certain embodiments, the training system and the system 100 can be executed on the same computer devices. For example, both the system 100 and the training system can be run on the server 103. In various embodiments, the system 100 can be used for training. For example, the system 100 can be configured to operate in a training mode, relying on user input for training data, and then enter prediction mode when the training criteria are met. After step 609, method 600 can end in step 611.

-25 [058] FIG. 7 describes a flow diagram illustrating a method of application navigation 700 using neural networks, compatible with the embodiments revealed. According to method 700, the prediction component 400 can receive an input vector, predict a next state, predict a next data object, and indicate a next page to the user. As one skilled in the art would recognize, this particular sequence of steps is not intended to be restrictive. The steps of method 700 can be combined or divided, and method 700 can include additional steps or fewer steps, without departing from the embodiments envisaged.

[059] After starting in step 701, the prediction component 400 can be configured to receive an input vector in step 703. In certain aspects, the prediction component 400 can be configured to receive the elements of the input vector directly or indirectly from the application 120. For example, in certain embodiments the prediction component 400 can be integrated into the application 120, and can create the input vector from the values stored by l application 120. As a further example, system 100 can be configured to implement prediction component 400 and application 120 as separate applications. The application 120 can then supply values for the elements of the input vector directly to the prediction component 400, or can supply these values to a memory or a database. The prediction component 400 can then retrieve these values from the memory or from the database.

[060] As revealed above in FIG. 5, in some embodiments the input vector may include the current data values 510, the application context 520 and the application history 530. Receiving the input vector may include receiving context information application by the user. For example, as noted above, the 520 application context can include one or more elements

-26 indicating who is using the application 120 (e.g. the entity associated with the user, a user name, a user role or a job title, or a user authorization or job responsibility ) or when the user interacts with the 120 application (e.g. at least one day of the week, the date or time of the interaction). The prediction component 400 can be configured to receive this information once (eg when the user logs into the application 120) and reuse this information to generate the input vector as necessary.

[061] After step 703, the prediction component 400 can be configured to predict at least one next state of application in step 705. As indicated above with regard to FIG. 4, the prediction component 400 can be configured to convert the values of the input vector into numbers using the state characteristics extractor 421. In some embodiments, the prediction component 400 can perform this conversion using a hash function. For example, the prediction component 400 can be configured to generate a hash of the raw binary values of each element of the input vector. The hash function may be the Fowler-Noll-Vo hash function, or another hash function known in the art. The hash values can be applied to the input nodes of the state predictor 420. The prediction component 400 can propagate these hash values by advancing them to the masked layer (s) and the output layer to generate probability scores . These probability scores may depend on the hashed input values and the weights of the state predictor 420 (which can be determined according to the method 600). State output 423 can be configured to select one or more of the most likely predicted application states. For example, state output 423 can be configured to select as the predicted application state the state application corresponding to the node of the output layer with the highest probability value. As an additional example, state output 423 can be configured to select

-27the application states corresponding to a predetermined number of nodes of the output layer with the largest probability values.

[062] After step 705, the prediction component 400 can be configured to predict at least one next data object in step 707. In certain embodiments, for each state selected by state output 423, the prediction component 400 can be configured to generate one or more of the data objects. As indicated above with regard to FIG. 4, the prediction component 400 can be configured to convert the identifiers of one or more application states selected by the state output 423 and the values of the input vector into numbers using the feature extractor 431 data object. In some embodiments, the feature extractor for data object 431 can use a hash function to perform this conversion. For example, the prediction component 400 can be configured to generate a hash of the raw binary values of each element of the input vector. The hash function may be the Fowler-Noll-Vo hash function, or another hash function known in the art. The hashed values can be applied to the input nodes of the data predictor 430. The system 100 can propagate these hashed values by forwarding them to the hidden layer (s) and the output layer to generate probability scores. These probability scores may depend on the hashed input values and the weights of the data predictor 430 (which can be determined by method 600). Data output 433 can be configured to select one or more of the most likely predicted application states. For example, data output 433 can be configured to select as the predicted data object the data object corresponding to the node of the output layer with the highest probability value. As an additional example, data output 433 can be configured to select the states

-28 of application corresponding to a predetermined number of nodes of the output layer with the largest probability values.

[063] After step 707, the prediction component 400 can be configured to predict at least one next page to the user in step 709. This indication can depend on the one or more predicted pairs of states of application and data objects. For example, the prediction component 400 may provide data or instructions to indicate the predicted page (s). For example, the instructions or data can configure the application 120 to display one or more links to the predicted pages, or modify one or more graphic elements of an existing page to indicate that interacting with the one or more elements would bring the application 120 to transition to the predicted page (s) (eg highlight, underline, animate, change element color, etc.). As an additional example, data or instructions can configure the application 120 to display explicit instructions indicating the next page. These instructions can be displayed on an inserted page, or on a separate page. In some aspects, the data or instructions can be provided to the client 101, which can use them to dynamically update an existing page. In various aspects, the data or the instructions can be supplied to the server 103, which can use them to create a new page to be supplied to the client 101. After step 709, the method 700 can end in step 711.

[064] The previous revelation describes the embodiments of a system for application navigation using neural networks. Such a system can generally be used with supplied applications as described in FIG. 1. In addition, such a system can be used for navigation in an application supplied by the client 101. For example, an application supplied by the client 101 can be configured to report any usage information including the status information. application and object information of

Given to the server 103. In this non-exhaustive example the application cannot obtain the data objects (or the information contained in the data objects) from the server 103. According to the usage information obtained from several users , the server 103 can be configured to generate a state predictor and a data object predictor. The server 103 can be configured to provide this application state predictor and the data object predictor to a prediction component of the application running on the client 101.

[065] Exemplary application: Asset management application [066] The processes and systems envisaged can provide recommendations to users of an asset management application. This exemplary application illustrates the applicability of the embodiments envisaged in the art and is not intended to be limiting.

[067] According to FIG. 1, a server can be configured to provide the asset management application to customers. Users can interact with the app through clients. According to FIG. 2A, the asset management application states can be linked, so that a user can transition from one state to another. According to FIG. 2B, the asset management application can include the application states that define the appearance and functionality of the pages of the asset management application. These application states can be defined using templates including markup language and script commands. For example, the application can include the application states defined using ANGULAR UI®, RESPOND®, or a similar JavaScript® structure. According to FIG. 2B, the asset management application can operate on data objects representing properties. These data management objects may include fields describing the characteristics of the goods, such as ownership, brand, name of the goods, location of the goods, inventory of the goods, pricing of the inventory, policies , promotions, booking classes and more

-30 characteristics of goods known to those skilled in the art. In accordance with the disclosed embodiments, an application state can be executed on a data object representing a policy aimed at generating a page. As previously revealed, users can interact with one or more pages to perform tasks. In this example, a task could include modifying the inventory available within an asset. The user could select a property using a first page, access a second page allowing the user to add the inventory to this property, and then go to a third page allowing the user to modify the categories of pricing and reservation for this newly added inventory.

[068] In accordance with Figs. 3A and 3B, the asset management system can be configured to display pages, and the probability that a user will transition from one page to another may depend on the user's prior trajectory by the through the asset management application.

[069] According to FIG. 4, in some embodiments, the asset management application may include a predictor 400 configured to use neural networks to predict the next page with which the user is to interact. This prediction element can include an input vector 410, a state predictor 420, a data predictor 430 and a page prediction. As a non-exhaustive example, neural networks can be implemented using JAVA®.

[070] In accordance with FIG. 5, in some embodiments, the input vector may include the current data values 510, the application context 520 and the application history 530. In some aspects, the values of the current data 510 may include the values of the current data object, as illustrated in the following exemplary pseudo-code:

[0711 input vector <= get (propertyCode);

-31 [072] input vector <= get (propertyDetails.extemalInventory);

[073] input vector <- get (propertyDetails.interfacedAvalon);

[074] input vector <= get (propertyDetails.interfacedRms);

[075] input vector <= get (propertyDetails.overrideSellStrategy);

[076] input vector <= get (propertyDetails.regionConversion);

[077] input vector <= get (propertyDetails.roomClass);

[078] input vector <= get (propertyDetails.tieredPricing);

[079] input vector <= get (propertyDetails.visibility);

[080] input vector <= get (propertyDetails.status);

[081] input vector <= get (propertyDetails.marketingCity);

[082] input vector <= get (propertyDetai! S.brandConversion) [083] input vector <= get (propertyDetails.classSharing);

[084] Thus the current data values 510 may include information about the inventory, pricing and location of the current data object [085] In various aspects, the application context 520 may include contextual parameters concerning the user of the asset management application, as indicated in the following exemplary pseudo-code [086] input vector <= get (login);

[087] input vector <= get (day);

[088] input vector <= get (month);

[089] input vector <= get (chainCode);

[090] input vector <= get (brandCode);

[091] Thus, the application context 520 can include the user's login information, the user's day and month and the information describing the owner and the brand.

-32 [092] In various aspects, the 530 application history can include previous application states and user ID parameters, as illustrated in the following exemplary pseudo-code:

[093] input vector <= [get (prior_states), get (prior_data_objects)];

[094] Thus, the application history 530 can include the pairs of previous application states and the data objects executed while the application was in these states.

[095] In accordance with FIG. 6, in some embodiments the prediction element of the asset management application can be formed. As revealed in Fig. 6, a training system can be configured to receive training data. In some aspects, this training system can be the system implementing the asset management application. In various aspects, this training system can be a separate system that receives training data from the system implementing the asset management application. Training data may include the data communicated above, and may additionally include the current application status and the current data object. As revealed in Fig. 6, the training system can be configured to train a state predictor and an object predictor. The state predictor and the object predictor can be provided to the system implementing the asset management application.

[096] In accordance with FIG. 7, in some embodiments, the prediction component of the asset management application can be configured to receive an input vector, predict a next state, predict a next data object, and then indicate one or more next pages predicted to the user. In some embodiments, this prediction can be made when a user transitions from a first page to a second page in the asset management application. In this example, the user could transition to one of the next page (s)

-33 predicted from the second page. As revealed above with respect to FIG. 7, the forecasting component of the asset management application can be configured to provide an indication of the predicted page to the user. In some respects, this indication can be provided as a dynamic update of the second page. In various aspects, a modified second page can be provided to the customer to be displayed to the user.

[097] FIG. 8 illustrates an exemplary graphical user interface 800 for an asset management application, compatible with the embodiments disclosed. As shown in Fig. 8, the prediction component can be configured to provide instructions to modify the graphical user interface 800 to include the insert 801. The insert 801 can include the icons corresponding to the next predicted pages (in this example, the goods 803a and region 803b). As indicated, these icons may vary in size and / or color. In some respects, these variations may indicate differences in probability. For example, when the next most likely page is goods 803a, the icon for goods 803a may be larger than the icon for region 803b. In some embodiments, the graphical user interface 800 can be configured to modify the icons on the events on which the mouse pointer is placed. For example, when the cursor is placed on one of the icons, this icon can grow in size, change shape or change color. In some embodiments, selecting one of these icons may cause the asset management application to transition to the corresponding page.

[098] FIG. 9 illustrates an additional exemplary graphical user interface 900 for an asset management application, compatible with the embodiments disclosed. As shown in Figure 9, the prediction component can be configured to provide instructions for modifying the graphical user interface 900 to include a

-34 textual indication 903 of the relative risk of the next predicted page corresponding to the graphic element 901. In certain embodiments, this textual indication can be provided, in addition to indicating the relative probabilities of the pages using the relative size of the corresponding icons.

[099] FIG. 10 illustrates an additional exemplary graphical interface 1000 for an asset management application, compatible with the embodiments disclosed. As shown in Fig. 10, the prediction component can be configured to provide instructions for modifying the graphical user interface 1000 to include an insert 1001. Inset 1001 can include a text description of a predicted next page. [0100] FIG. 11 is a block diagram of an exemplary device with which the systems, methods and devices of the present invention can be implemented. The example device 1100 can comprise at least one processor (eg 1103) and at least one memory (eg memories 1105a and 1105b). The processor 1103 may include a central processing unit (UCT), a graphics processing unit (UGT), or other similar circuit capable of performing one or more operations on a data stream. The processor 1103 can be configured to execute the instructions which can, for example, be stored on one or more of the memories 1105a and 1105b. Memories 1105a and 1105b can be volatile memory (such as RAM or the like) and / or non-volatile memory (such as a flash memory, a hard disk or the like). As explained above, the memories 1105a and 1105b can store instructions for execution by the processor 1103. As illustrated later in FIG. 11, the user interface device 1100 may include at least one network interface controller (CIR) (eg CIR 1115). The CIR 1115 can be configured to facilitate communication on at least one computer network (eg network 1117). Communication functions can therefore be facilitated through one or more CIRs, which risk (s) of being wireless or wired and

-35 possibly include an Ethernet port, radio frequency transmitters and receivers, and / or optical transmitters and receivers (e.g. infrared). The specific design and implementation of the one or more CIRs depend on the computer network 1117 on which the user interface device 1100 is intended to operate. For example, in certain embodiments, the user interface device 1100 can include one or more wireless or wired CIRs designed to operate on a GSM network, a GPRS network, an EDGE network, a network Wi-Fi or WiMax, and a Bluetooth® network. Alternatively or at the same time, the user interface device 1100 can include one or more wireless or wired CIRs designed to operate on a TCP / IP network. As described in FIG.

11, the user interface device 1100 can include and / or be operatively connected to a storage device 1121. The storage device 1121 can be volatile (such as RAM or the like) or non-volatile (such as memory flash, hard drive or similar). The I / O module 1119 can allow communications between the processor 1103, the memories 1105a and 1105b, the CIR 1115, and / or the storage device 1121. The processor 1103, the memories 1105a and 1105b, the CIR 1115 and / or the storage device 1121 can comprise separate components or can be integrated into one or more integrated circuits. The various components within the user interface device 1100 can be coupled to one or more communication buses or signal lines (not shown).

The above description has been presented for the purpose of illustration. It is not exhaustive and is not limited to the precise forms or to the embodiments disclosed. The modifications and adaptations of the embodiments will result from the examination of the specification and the implementation of the disclosed embodiments. For example, the described implementations include hardware and software, but systems and methods compatible with this disclosure can be implemented with hardware

-36 Patent application n ° 18 51948 A5013309FRP00Mm only. In addition, while some components have been described as being coupled together, these components can be integrated with each other or distributed as appropriate.

In addition, while the illustrative embodiments have been described herein, the scope includes all embodiments having equivalent elements, modifications, omissions, combinations (e.g., aspects to through various embodiments), adaptations and / or alterations based on this disclosure. The instructions or the operational steps stored by a computer-readable medium can be in the form of computer programs, program modules, or codes. As described herein, computer programs, program modules, and code based on the written description of this specification, such as that used by the processor, is easily the responsibility of a software developer. Computer programs, program modules, or code can be created using a variety of programming techniques. For example, they can be designed in or through Java, C, C ++, assembly language, or any other programming language. One or more of these programs, modules or code can be integrated into an existing device system or communication software. Programs, modules or code can also be implemented or replicated as firmware or circuit logic.

The characteristics and advantages of the disclosure emerge from the detailed specifications. Furthermore, insofar as numerous modifications and variations will result openly from the study of this revelation, it is not intended to limit the revelation to the exact construction and to the operation as illustrated and described, and in

Consequently, all appropriate modifications and equivalences can be invoked, while remaining reasonably within the scope of the revelation.

Other embodiments will follow from examining the specification and implementing the embodiments disclosed herein. It is intended that the specifications and examples are to be regarded as examples only.

Claims (15)

Claims: 1. A system for application navigation using neural networks, comprising: at least one processor; and at least one non-transient memory comprising instructions which, when executed by said at least one processor, cause the system to implement operations comprising: the provision of an application comprising application states which work on data objects to generate pages; predicting a next page using an application state predictor neural network and a data object predictor neural network, the prediction comprising: predicting a next next application state using the application state predictor neural network and a first input vector; and predicting a next data object using the data object predictor neural network, a second input vector and the next application state; and instructional instruction to display an indication of the next predicted page. 2. The system according to claim 1, in which an output layer of the application state predictor neural network comprises output nodes corresponding to the application states. The system of claim 1 or claim 2, wherein predicting the next state of application using the state-predictor neural network. -39 of application and the first input vector, includes the hash of the first input vector. 4. The system according to any one of the preceding claims, wherein common elements of the first input vector and the second input vector are hashed once and reused to predict the next data object. 5. The system according to any one of the preceding claims, wherein the first input vector and the second input vector are the same vector. The system according to any of the preceding claims, wherein an output layer of the data object predictor-neural network includes output nodes corresponding to the data objects. 7. The system according to any one of the preceding claims, in which the first input vector comprises a history of the application and / or includes elements indicating at least one of an entity, a user associated with the entity. , a user role, a user authorization level, a day of the week, a date, a time, or a value of a current data object. 8. The system of claim 7, wherein the application history includes elements indicating prior application states and prior data objects associated with the pages previously visited. 9. The system according to claim 1, in which the operations furthermore comprise the generation of a current page in accordance with a current application state using a current data object, and in which the contribution of instructions for displaying the indication of the next predicted page includes the provision of instructions for displaying the current page, modified to indicate the next predicted page. The system of claim 9, wherein providing instructions to display the indication of the next predicted page includes providing instructions to dynamically generate the current page in accordance with the current application state. using the current data object, the current page modified to indicate the next predicted page. The system according to any one of the preceding claims, wherein providing instructions to display the indication of the next predicted page includes providing instructions to display the indications of multiple predicted next pages including the next page. predicted. The system of claim 11, wherein the instructions for displaying the indications of the predicted next multiple pages include the instructions for indicating the relative probabilities of the predicted next multiple pages. 13. The system of any of claims 9 to 12, wherein providing instructions for displaying an indication of the next predicted page includes providing instructions for refreshing the current page to include an item -41 graphic indicating the next predicted page or by modifying a graphic element of the current page to indicate the next predicted page. 14. The system of claim 13, wherein the graphical element and a window 5 inserted including an icon corresponding to the next predicted page, the icon being preferably selectable for making a transition to the next predicted page. 15. The system of claim 13 or claim 14, wherein the modification of 10 the existing graphic element of the current page includes changing at least one positioning, shape, size, or color of the existing graphic element; or changing at least one positioning, font, size, accent, or color of text associated with the existing graphic element.