KR20210062372A - Method and server for predicting user interaction for web applications - Google Patents

Abstract

The present invention provides a method of predicting user interaction for a web application, which uses an artificial neural network. According to one aspect of the present invention, a method of predicting user interaction for a web application includes the steps of: collecting event information including a navigation event and a click event for URL search generated when searching for a web application through a network from a user terminal; generating user behavior analysis data by applying a pre-processing process including data filtering and data vectorization for performance improvement to the collected event information; generating prediction result data including an event to occur after the new event by inputting new event occurrence information into a user interaction prediction model when a new event occurs according to a web application search in the user terminal; and adjusting a service state to be provided to the user terminal by the server based on the prediction result data. The user interaction prediction model is learned using a recurrent neural network (RNN) based on user behavior analysis data.

Description

How to predict user interactions for web applications and servers {METHOD AND SERVER FOR PREDICTING USER INTERACTION FOR WEB APPLICATIONS}

The present invention relates to a technique for collecting user interaction data generated in a web application and predicting the next requested resource based on this. More specifically, event data that occurs when a user interacts with a web application is collected and It relates to a method and a server for predicting the next event by training a prediction model based on it.

Understanding user interaction in web applications is very important to improve the quality of experience (QoE) of users when browsing the web. Web browsing consists of a series of clicks triggered by the user and navigation to other web applications accordingly. Given prior knowledge of such user interaction, the browser is the content that the user is expected to access next. It can be downloaded in advance and stored temporarily, so that it can be used immediately upon request of a user in the future. This prefetching technique ultimately improves the quality of experience for web applications by reducing the long delay required to download resources for web applications involving multiple multimedia resources. For example, Google's Chrome browser preemptively performs tasks such as prefetching domain name service (DNS) and prefetching TCP and TLS before a request from an actual user occurs in order to load web pages faster. It provides a predictive service that operates as a function. Today, browsers perform prefetching using clues specified in the source code by web application developers, which are based on simple rules using static information such as document structure, past navigation history, and metadata related to devices and users. Is based. On the other hand, navigation prediction based on the relationship between user interaction and web application can be used as a more valuable clue in improving prefetching performance by reducing the reaction time of individual navigation.

However, prediction of user interaction for web applications in real environments is difficult to properly utilize due to limitations in data collection and low prediction accuracy of existing models. Existing technologies that utilize user interaction information collected at the server level do not properly reflect the behavior of users interacting while browsing various web applications, and there is a limitation in the scope of application to specific web application services. On the other hand, the existing technology that collects detailed user interaction data such as clicks, calls, and application execution history at the client side requires modification of the source code of the browser or application, or its application range is limited to mobile platforms. This measurement limitation hinders the possibility of distribution and utilization of data collection techniques, making it difficult to properly collect user interaction data in real life. Meanwhile, the existing technology for web navigation prediction has been developed focusing on web search or online advertisement corresponding to typical web applications. These technologies can be used for traditional web applications, but they are not suitable as general-purpose technologies that can be applied to today's web applications that support various forms. In addition, in the case of a web environment in which the user's action range is limited to installed applications, in the case of a web environment in which any application can be explored, the variety of applications targeted for navigation prediction is very large, so simply applying the existing machine learning is There is a limitation in showing low prediction accuracy. Therefore, in order to accurately predict user interactions for web applications, a method of collecting user interaction data without limitation of the target application or platform, as well as an additional method that can improve the navigation prediction performance in consideration of the diversity of web applications is needed.

In relation to the user interaction prediction model, a recurrent neural network is composed of both forward and backward connections, unlike a feedforward neural network in which activation of individual neurons flows only in one direction. This reverse connection makes it possible for cyclic neurons to accept not only inputs but also outputs from previous times, so that they behave like a kind of memory, creating a structure suitable for processing time series data. The long short-term memory (LSTM), which improved the problem of the initial RNN, maintains both short and long-term states, leading to faster convergence during training compared to the initial RNN structure, and better detecting long-term dependencies existing on the data. GRU, which was recently proposed by lowering the complexity of LSTM, is widely used for learning time series data such as natural language processing by lowering the complexity of learning while producing almost the same performance as LSTM. This GRU is suitable to be used as a base prediction model for user interaction prediction in web applications that use a series of interaction events given over time, such as clicks and navigation, as input data. However, since the GRU-based prediction model alone shows low prediction accuracy, which is not suitable for application to an actual environment, an additional data processing technique is required to improve the performance of the base model.

Korean Patent Laid-Open Publication No. 10-2015-0009811 (Name of the invention: A method of improving web server performance using public web server and web requirement prediction)

The present invention provides a method for predicting user interaction for a web application using an artificial neural network as to solve the problems of the prior art described above.

Regardless of the target application or platform, by collecting user interaction data for web applications, stably expressing and recognizing each event, and learning through a GRU-based predictive model applied with data processing techniques to improve accuracy, Provides a method of predicting user navigation in an accurate level.

However, the technical problem to be achieved by the present embodiment is not limited to the technical problem as described above, and other technical problems may exist.

As a technical means for solving the above-described technical problem, the method for predicting user interaction for a web application according to the first aspect of the present disclosure includes a navigation event and a click event for URL search that occur when a web application is searched through a network. Collecting event information including a user terminal; Generating user behavior analysis data by applying a pre-processing process including data filtering and data vectorization for performance improvement to the collected event information; Generating prediction result data including an event to occur after the new event by inputting new event occurrence information into a user interaction prediction model when a new event occurs according to a web application search in the user terminal; And adjusting a service state to be provided to the user terminal by the server based on the prediction result data. The user interaction prediction model is learned using a recurrent neural network (RNN) based on user behavior analysis data.

As a technical means for solving the above technical problem, the user interaction prediction server for a web application according to the second aspect of the present disclosure includes: a memory storing a user interaction prediction program for the web application; It includes a processor that executes a program stored in memory. The processor collects event information including a navigation event and a click event for URL search, which occurs when browsing a web application through a network by executing a program, from the user terminal, and performs data filtering and data vectorization to improve performance. The included pre-processing process is applied to the collected event information to generate user behavior analysis data, and when a new event occurs according to web application browsing in the user terminal, the event to occur after the new event by inputting the new event occurrence information into the user interaction prediction model Generates prediction result data including, and adjusts a service state to be provided to the user terminal in the server based on the prediction result data. The user interaction prediction model is learned using a recurrent neural network (RNN) based on user behavior analysis data.

As a technical means for solving the above-described technical problem, the method for predicting user interaction for a web application in a user terminal according to the third aspect of the present disclosure is for searching a URL that occurs when searching for a web application through a network. Collecting event information including a navigation event and a click event; Generating user behavior analysis data by applying a pre-processing process including data filtering and data vectorization for performance improvement to the collected event information; Generating prediction result data including an event to occur after the new event by inputting new event occurrence information into a user interaction prediction model when a new event occurs according to a web application search in the user terminal; And adjusting a state of providing a service to be provided to the user terminal with respect to the external application service providing server based on the prediction result data. The user interaction prediction model is learned using a recurrent neural network (RNN) based on user behavior analysis data.

According to any one of the above-described problem solving means of the present application, performance of the prediction model may be further improved by utilizing a GRU-based prediction model to which a data processing technique for improving accuracy is applied.

Because it is designed based on JavaScript, one of the core technologies that compose most web applications, it is possible to collect events for user interaction such as clicks and navigation without modifying the browser or application source code for general web applications. In addition, by expressing and recognizing the clicked object based on the document object model (document object, model), stable object recognition is possible even for different devices or client software.

In addition, a variety of web applications are targeted for navigation prediction by applying a performance improvement technique that combines and processes a variant URL (uniform resource locator) for one web page in a GRU-based prediction model and expresses navigation and click events in the same vector space. Even in existing environments, navigation events can be predicted with high accuracy. This is applied as a base technology for the web prefetching technique to provide a high quality service with high user experience quality.

1 is a block diagram showing the configuration of a user interaction prediction server for a web application according to an embodiment of the present invention.
2 is a block diagram showing the configuration of a user interaction prediction system for a web application according to an embodiment of the present invention.
3 is a flow chart showing a process step of a method for predicting user interaction for a web application according to an embodiment of the present invention.
4 is a diagram showing the overall configuration of a method for predicting user interaction for a web application according to an embodiment of the present invention.
5 is a diagram illustrating a difference in performance when a preprocessing process is performed in a method for predicting user interaction for a web application according to an embodiment of the present invention.
6 is a diagram illustrating a difference in performance when a preprocessing process is performed in a method for predicting user interaction for a web application according to an embodiment of the present invention.

Hereinafter, embodiments of the present application will be described in detail with reference to the accompanying drawings so that those of ordinary skill in the art may easily implement the present application. However, the present application may be implemented in various different forms and is not limited to the embodiments described herein. In addition, in the drawings, parts irrelevant to the description are omitted in order to clearly describe the present application, and similar reference numerals are attached to similar parts throughout the specification.

Throughout this specification, when a part is said to be "connected" with another part, this includes not only the case that it is "directly connected", but also the case that it is "electrically connected" with another element interposed therebetween. do.

Throughout this specification, when a member is positioned “on” another member, this includes not only the case where a member is in contact with the other member, but also the case where another member exists between the two members.

Hereinafter, an embodiment of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram showing the configuration of a user interaction prediction server for a web application according to an embodiment of the present invention.

As illustrated, the user interaction prediction server 100 for a web application may include a communication module 110, a memory 120, a processor 130, a database 140, and an input module 150.

The communication module 110 communicates data with the connected user terminal 100 and the linked server 200, respectively. The communication module 110 may be a device including hardware and software necessary for transmitting and receiving a signal such as a control signal or a data signal through a wired or wireless connection with another network device.

The memory 120 stores a user interaction prediction program for a web application. The user interaction prediction program for web applications collects event information from the user terminal, including navigation events and click events for URL search, which occurs when browsing web applications through the network, and data filtering and data to improve performance. User behavior analysis data is generated by applying a pre-processing process including vectorization to the collected event information, and when a new event occurs according to web application search in the user terminal, the new event occurrence information is input into the user interaction prediction model, after the new event. Prediction result data including an event to occur is generated, and the server adjusts a service state to be provided to the user terminal based on the predicted result data.

The memory 120 stores various types of data generated during execution of an operating system for driving the user interaction prediction server 100 for a web application or a user interaction prediction program for a web application.

In this case, the memory 120 collectively refers to a nonvolatile storage device that continuously maintains stored information even when power is not supplied, and a volatile storage device that requires power to maintain the stored information.

In addition, the memory 120 may perform a function of temporarily or permanently storing data processed by the processor 130. Here, the memory 120 may include a magnetic storage media or a flash storage media in addition to a volatile storage device requiring power to maintain stored information, but the scope of the present invention is limited thereto. It does not become.

The processor 130 executes the program stored in the memory 140, but controls the entire process according to the execution of the user interaction prediction program for the web application. Each operation performed by the processor 130 will be described in more detail later.

The processor 130 may include all types of devices capable of processing data. For example, it may refer to a data processing device embedded in hardware having a circuit that is physically structured to perform a function represented by a code or command included in a program. As an example of a data processing device built into the hardware as described above, a microprocessor, a central processing unit (CPU), a processor core, a multiprocessor, and an application-specific application-specific device (ASIC) Integrated circuit) and processing devices such as a field programmable gate array (FPGA) may be covered, but the scope of the present invention is not limited thereto.

The database 140 stores or provides data necessary for a user interaction prediction system for a web application under the control of the processor 130. The database 140 may be included as a separate component from the memory 140 or may be built in a partial area of the memory 140.

2 is a block diagram showing the configuration of a system for predicting user interaction for a web application according to an embodiment of the present invention.

The method for predicting user interaction for a web application may be implemented in the server 100 or the user terminal 200. In the method for predicting user interaction for a web application, event information such as navigation or click is collected in the user terminal 200. Data analysis based on the collected event information may be performed in the server 100, and prediction result data may be provided to the user terminal 200. The user terminal 200 may directly predict an event through the user interaction prediction model and adjust the provision state of a service to be provided to the user terminal.

3 is a flow chart showing a process step of a method for predicting user interaction for a web application according to an embodiment of the present invention.

The processor 130 collects event information including a navigation event and a click event for URL search, which occurs when browsing a web application through a network, from the user terminal (S110). The event information collected at this time may include at least one of user identifier information, event type information, searched URL information, clicked object information, or a timestamp indicating an event occurrence time.

The processor 130 may collect navigation and click events that occur as a user searches for a general web application through an event collection tool in a client browser of the user terminal 200 and transmits them to the server 100. At this time, each clicked object is expressed based on the location and characteristics on the document object model tree, and is recognized collectively regardless of the device or browser used for measurement.

In this case, the document object model is an interface that allows a program or script to access components in a web page and change content or style, and is an interface that is not dependent on a specific platform or language.

The server 100 stores event information (user interaction data) collected from the user terminal 200 in a database and uses it for learning a user interaction prediction model.

The processor 130 generates user behavior analysis data by applying a preprocessing process including data filtering and data vectorization for performance improvement to the collected event information (S120).

The pre-processing process may include a URL grouping process in which URLs satisfying a preset condition are recognized as one integrated group. The URL grouping process removes the query string part of the URL preset for the navigation event, combines the URL from which the query string was removed for the click event and the string indicating the path to the clicked HTML element, and then sets a preset at page load time. URLs requesting the same multimedia resources occupying more than a proportion may be recognized as one unified group.

In addition, the pre-processing process may include a web embedding process in which the event information is arranged in the same target space and then vectorized in a low dimension. The web embedding process uses a word-to-vector (word2Vec) skip-gram to receive a navigation event and a click event included in the event information as inputs in an array form, and redesign the context between the inputs. , Using a stochastic gradient descent technique, the objective function is repeatedly optimized in a direction to reduce the error of vector expression, and the finally found embedding may be stored in the form of a dictionary.

The processor 130 may filter out unnecessary data through a pre-processing process and transform it into a form that is easy for predictive model learning. In the preprocessing process, variant URLs for one web page are integrated into one group to improve prediction accuracy, and navigation and click events are closely related to the same vector space by considering the degree of proximity between events to reduce data sparsity. And express it so that it can be matched. Based on all processed data (user behavior analysis data), a GRU-based user interaction prediction model is trained, and the learned prediction model is an application to be explored next when a new event, user interaction data array, is given as input. Can be used to predict

When a new event occurs according to web application search in the user terminal, the processor 130 inputs new event occurrence information into the user interaction prediction model to generate prediction result data including an event that will occur after the new event (S130).

The user interaction prediction model uses a GRU (Gated Recurrent Unit) of a recurrent neural network, and learns a relationship with an event that occurs next by using as input data a series of events given in chronological order, such as clicks and navigation, It may be predicting user events that will occur in the future.

The user interaction prediction model is an input layer capable of accepting an input of variable length according to context, a center composed of multiple layers of hidden layers composed of multiple GRU cells, and fully-connected to output prediction probability for each prediction target. It may include a layer and a softmax layer for deriving a probability value for each prediction target.

The GRU-based user interaction prediction model receives an array of user interaction events, such as navigation and clicks that occurred recently based on a specific point in time, as an input in chronological order, and learns the relationship between them by outputting the next application to be searched. The processor 130 filters invalid URLs such as a browser setting page for navigation and click event data collected in a preprocessing process performed prior to predictive model training, and events adjacent to each other in time among the data classified by each user's identifier. Clusters into individual groups. The processor 130 excludes data on user identifiers, web applications, and click objects that have a very low frequency of appearance in the data set during preprocessing. The processor 130 creates a dictionary by assigning a unique integer to all the URLs and click objects observed for the organized data set, so that the data expressed in text can be interpreted in the predictive model. Convert to form.

The preprocessed data is used as the input of the GRU-based user interaction prediction model, and the input layer of the GRU-based user interaction prediction model includes both the input data and the length of the input in order to accommodate the input of variable length depending on the context. Handle it. After the input layer, a hidden layer composed of multiple GRU cells is stacked to form the center of the prediction model, a fully-connected layer is added to output the prediction probability for each prediction target, and a softmax layer is finally added to predict each prediction. Derive the output value for each object. The user interaction prediction model is designed as a deep neural network structure suitable for web navigation prediction by utilizing the GRU cell structure widely used in the existing natural language processing field.

The processor 130 provides the prediction result data to the user terminal (S140). The processor 130 adjusts a service state to be provided from the server to the user terminal based on the prediction result data. In this step, a web prefetching step of prefetching content according to the prediction result data to the user terminal may be included. In addition, the processor 130 caches it in advance based on the predicted result data, or performs a server push that determines and transmits content that is easy to improve the quality of the user's experience in the server. It is also possible to use it throughout the service.

4 is a diagram showing the overall configuration of a method for predicting user interaction for a web application according to an embodiment of the present invention.

In the client browser of the user terminal 200, the event collection tool collects navigation and click events that occur as the user searches for a general web application and transmits them to the server 100. At this time, each clicked object is expressed based on the location and characteristics on the document object model tree, and is recognized collectively regardless of the device or browser used for measurement.

The server 100 stores the collected event information (user interaction data) in a database and uses it for learning a user interaction prediction model. Data is transformed into a form that is easy for predictive model training by filtering out unnecessary data through a pre-processing process. To improve prediction accuracy, variant URLs for one web page are integrated into one group, and navigation and click events are closely matched to the same vector space by considering the degree of proximity between events to reduce data sparsity. Is expressed. Based on all processed data (user behavior analysis data), a GRU-based user interaction prediction model is trained, and the learned prediction model is an application to be explored next when a new event, user interaction data array, is given as input. Can be used to predict

In the present invention, by designing a user interaction event collection tool using an event handler based on JavaScript, which is one of the core technologies constituting a web application, the constraints of the target application or platform are relaxed to increase the determinism of the data collection technique. The following shows the specific operation process of the event collection tool.

When a new web page is loaded, the event collection code written in JavaScript is inserted into the script code of the HTML (hypertext markup language) document. This insertion process does not require modification of the source code of the web application, and can be controlled by the internal engine of the browser or implemented in the form of a browser extension program. In an embodiment of the present invention, the result implemented as a Google Chrome browser extension program was used.

The event collection code defines handlers for two JavaScript events: navigation and click. In this case, information on the identification of the user who triggered the event, the type of the event, the URL to be searched, and a timestamp indicating the occurrence time of the event are collected for the navigation event. For the click event, the same data as the information collected for the navigation event (however, the timestamp is expressed as the relative time after the navigation of the web page), and additional information to recognize the clicked object is additionally collected. In this process, in order to protect the privacy of each user, the user's identifier uses a 32-digit hexadecimal value randomly generated and assigned for each browser account. Event handlers defined in this way are added to the <body> element of the loaded document to collect events. Table 1 below shows examples of data collected through an event collection tool.

[Table 1]

For the click event, the path to the clicked HTML element on the document object model tree is additionally collected. The object clicked due to user interaction is recognized as a specific node located on the document object model tree, and all existing in the path from the <body> element to the clicked HTML element in the document object model tree corresponding to one document. An integrated string is created based on the nodes to recognize the clicked object. In addition to the location of the clicked HTML element to distinguish objects with the same tag name, the ID and class information of the element are included in the string.

HTML elements that exist in the path to the clicked HTML element are separated by a special string called'@'.

(Example: @BODY@DIV@SPAN...)

At this time, to classify elements with the same tag name, CLASS information is indicated as'.' and ID information is indicated as'#'.

(Example: @BODY@DIV#LayoutWrapper...@SPAN.Carousel)

In front of each unified string, the URL of the searched web page is added to indicate which web application it is. At this time, if the clicked object is a child node of the <frame> or <iframe> tag, the URL of the frame is added instead.

(Example: https://www.example.com/@BODY@DIV#LayoutWrapper...@SPAN.Carousel)

The collected event data is accumulated and stored in the browser temporary storage space, and when moving to a new web application, the data stored up to that point is encoded in JSON format and transmitted to the server. The server stores the collected user interaction data in a database and uses it for predictive model training.

In the present invention, a user interaction prediction model is designed based on the GRU. The GRU-based model receives an array of user interaction events, such as navigation and clicks that occurred recently based on a specific point in time, as an input in chronological order, and learns the relationship between them by outputting the next application to be searched. In the preprocessing process prior to predictive model training, invalid URLs such as browser setting pages are filtered for the collected navigation and click event data, and the data classified by each user's identifier clusters adjacent events in time into individual groups. In addition, data on user identifiers, web applications, and click objects with a very low frequency of appearance in the dataset are excluded. By creating a dictionary by assigning a unique integer to all the URLs and click objects observed in the data set organized in this way, the data expressed in text is converted into a numeric form that can be interpreted by the predictive model. The preprocessed data is used as an input of the GRU-based prediction model, and the input layer of the GRU-based prediction model processes both the input data and the length of the input to accommodate the input of variable length according to the context. After the input layer, a hidden layer composed of multiple GRU cells is stacked to form the center of the prediction model, a fully-connected layer is added to output the prediction probability for each prediction target, and a softmax layer is finally added to predict each prediction. Derive output values for each object. As such, a deep neural network structure suitable for web navigation prediction was designed using the GRU cell structure, which was widely used in the field of natural language processing.

In the present invention, in order to improve the accuracy of the GRU-based prediction model, two performance enhancement techniques, called URL grouping and web embedding, are applied. The following shows a detailed operation process of each prediction performance improvement technique.

For the navigation event, the query string part of the specific URL is removed, the URL from which the query string is removed for the click event and the string representing the path to the clicked HTML element are combined, and the generated strings are each It is used as an identifier for the event of

For example, in a situation where the following three URLs exist, the query string consists of the type (where) of the previous page, the query of the current page, and the oquery of the previous page. These factors correspond to characteristics commonly used in search engines to analyze the relationship between web pages.

(URL 1) https://www.search.com/result.html?where=blogpost&query=machinelearning

(URL 2) https://www.search.com/result.html?where=searchbar&query=deeplearning&oquery=machinelearning

(URL 3) https://www.search.com/result.html?where=searchbar&query=gru&oquery=deeplearning

In the above example, the pages have different URLs, but all multimedia resources that occupy a major part of the page load time must be requested and downloaded to the server in the same manner. When URL grouping is used, these URLs are recognized and predicted as one unified group (ie, https://www.search.com/result.html), so predicted resource sharing becomes possible.

In web embedding, user interaction events are mapped and expressed in a low-dimensional vector space. Web embedding takes the entire array of events in the dataset as input and allocates a close vector of 64 dimensions for each event identifier. At this time, unlike the existing word embedding, which could only process words or phrases of the same type, in web embedding, two different objects, such as navigation and click, are aligned in the same target space and then vectorized in a lower dimension. It goes on. In the present invention, the number of skips is 2 and the size of the skip window is 1 so that the context between input symbols can be redesigned using the skip-gram model of Word2vec to implement web embedding. Train a two-layer neural network.

For example, as a dataset for web embedding, the following arrangement of navigation and click events can be considered. In other words, the user navigates to a web application with URL_a, clicks on three objects Object_x, Object_y, and Object_z, then navigates to a web application with URL_b, and then navigates to a web application with URL_c again ((Navigation to URL_a), (Click to Object_x), (Click to Object_y), (Click to Object_z), (Navigation to URL_b), (Navigation to URL_c), ...}

As the first process in the above example, (context, target) pairs for adjacent events are created as follows. This is because the size of the skip window is 1, so the left and right objects are defined as context based on a specific object. {([Navigation to URL_a, Click to Object_y], Click to Object_x), ([Click to Object_x, Click to Object_z) ], Click to Object_y), ([Click to Object_y, Navigation to URL_b], C to Object_z), ...}

At this time, since the skip-gram model predicts the context symbol from the target symbol, the dataset is converted into the following (input, output) pairs: ((Click to Object_x, Navigation to URL_a), (Click to Object_x, Click to Object_y) , (Click to Object_y, Click to Object_x), (Click to Object_y, C to Object_z), ...}

After that, the objective function is optimized in the direction of repeatedly reducing the error of vector expression by using stochastic gradient descent (SGD) technique, and the finally found embedding is stored in the form of a dictionary. In the actual learning step, each event given as an input is converted into a corresponding vector representation by searching the stored dictionary, and then a GRU-based prediction model is used as an input.

5 is a diagram illustrating a difference in performance when a preprocessing process is performed in a method for predicting user interaction for a web application according to an embodiment of the present invention.

In this embodiment, the effect of the present invention was verified by distributing an event collection tool to real users and learning a user interaction prediction model through the interaction data collected through it. The GRU-based prediction model consists of 4 layers, 128 units, uses the He initialization technique, uses 10 time steps per iteration, the activation function is ReLU, the optimization algorithm is Momentum, and the learning rate is 0.01, A batch size of 25 was used. The performance of the user interaction prediction model was evaluated through F-measure.

FIG. 5 shows the effect of the click event on the overall performance improvement when the click event is used for prediction, and it can be seen that the performance improvement of 9.2 to 19.8% on average, and 13.7% on average, depending on the used prediction model configuration. Whereas most of the existing prediction techniques utilize only navigation events between web pages, the present invention further improves prediction performance by utilizing click data generated in one application together for prediction.

6 is a diagram showing the effect of the URL grouping and web embedding techniques developed in the present invention on performance improvement according to the number of prediction candidate groups in the method for predicting user interaction for a web application according to an embodiment of the present invention. It can be seen that the F-measure is improved by an average of 39.% when the URL grouping is applied, and up to 85.6% when the number of candidate groups is small, compared to the simple training of the prediction model using GRU. In addition, when web embedding is additionally applied, it shows an additional average of 10.6% performance improvement. In other words, when the two techniques are combined and applied, the average performance improved by 56.6% compared to the basic GRU model, and the predictive performance of F-measure 0.798 was achieved through 52.4% performance improvement for the 3 candidate groups corresponding to the realistic number of candidate groups. I did. These results indicate that the present invention is suitable for use in predicting user interactions for web applications in an actual environment in consideration of the practicality and accuracy level of the present invention.

An embodiment of the present invention may also be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer-readable media can be any available media that can be accessed by a computer, and includes both volatile and nonvolatile media, removable and non-removable media. Further, the computer-readable medium may include a computer storage medium. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data.

Although the methods and systems of the present invention have been described in connection with specific embodiments, some or all of their components or operations may be implemented using a computer system having a general-purpose hardware architecture.

The foregoing description of the present application is for illustrative purposes only, and those of ordinary skill in the art to which the present application pertains will be able to understand that it can be easily modified into other specific forms without changing the technical spirit or essential features of the present application. Therefore, it should be understood that the embodiments described above are illustrative and non-limiting in all respects. For example, each component described as a single type may be implemented in a distributed manner, and similarly, components described as being distributed may also be implemented in a combined form.

The scope of the present application is indicated by the claims to be described later rather than the detailed description, and all changes or modified forms derived from the meaning and scope of the claims and their equivalent concepts should be construed as being included in the scope of the present application.

100: server
110: communication module
120: memory
130: processor
140: database
150: input module
200: user terminal

Claims (14)

In the method for predicting user interaction for a web application performed by a server,
(a) collecting event information from a user terminal including a navigation event and a click event for URL search, which occurs when a web application is searched through a network;
(b) generating user behavior analysis data by applying a pre-processing process including data filtering and data vectorization for performance improvement to the collected event information;
(c) generating prediction result data including an event to occur after the new event by inputting the new event occurrence information into a user interaction prediction model when a new event occurs according to a web application search in the user terminal; And
(d) adjusting a service state to be provided to the user terminal in the server based on the prediction result data; including,
The user interaction prediction model is learned using a recurrent neural network (RNN) based on the user behavior analysis data,
How to predict user interaction for web applications. The method of claim 1,
The event information collected in step (a) comprises at least one of user identifier information, event type information, searched URL information, clicked object information, or a timestamp indicating an event occurrence time,
How to predict user interaction for web applications. The method of claim 1,
The pre-processing process applied in step (b) includes a URL grouping process for recognizing URLs that satisfy a preset condition as one integrated group,
The URL grouping process removes the query string part of the URL preset for the navigation event, combines the URL from which the query string is removed for the click event and the string indicating the path to the clicked HTML element, and then writes it at page load time. Recognizing URLs requesting the same multimedia resources occupying more than a set proportion as one integrated group,
How to predict user interaction for web applications. The method of claim 1,
The pre-processing process applied in step (b) includes a web embedding process in which the event information is arranged in the same target space and then vectorized in a low dimension,
The web embedding process uses a word-to-vector (word2Vec) skip-gram to receive navigation events and click events included in the event information as inputs in an array form, and redesign the context between the inputs. And, by using a stochastic gradient descent technique, the objective function is repeatedly optimized in the direction of reducing the error of the vector expression, and the finally found embedding is stored in the form of a dictionary.
How to predict user interaction for web applications. The method of claim 1,
The step (d) includes a prefetching step of prefetching content corresponding to the prediction result data to the user terminal,
How to predict user interaction for web applications. The method of claim 1,
The user interaction prediction model uses a GRU (Gated Recurrent Unit) of a recurrent neural network, and learns the relationship with the next event by using as input data a click event and a navigation event occurring in a chronological order. Predicting user events,
How to predict user interaction for web applications. The method of claim 6,
The user interaction prediction model is an input layer capable of accepting an input of variable length depending on the context, a center composed of multiple layers of hidden layers composed of multiple GRU cells, and fully-connected to output prediction probabilities for each prediction target. Including a layer and a softmax layer for deriving a probability value for each prediction target,
How to predict user interaction for web applications. In the user interaction prediction server for a web application,
A memory in which a program for predicting user interaction for a web application is stored;
And a processor that executes a program stored in the memory, the processor by execution of the program,
Collecting event information including a navigation event and a click event for URL search that occurs when browsing a web application through a network from a user terminal,
Generating user behavior analysis data by applying a pre-processing process including data filtering and data vectorization for performance improvement to the collected event information,
When a new event occurs according to a web application search in the user terminal, the new event occurrence information is input into a user interaction prediction model to generate prediction result data including an event to occur after the new event,
Based on the prediction result data, the server adjusts the service state to be provided to the user terminal,
The user interaction prediction model is learned using a recurrent neural network (RNN) based on the user behavior analysis data,
User interaction prediction server for web applications. The method of claim 8,
The pre-processing process includes a URL grouping process for recognizing URLs satisfying a preset condition as one integrated group,
User interaction prediction server for web applications. The method of claim 8,
The pre-processing process includes a web embedding process in which the event information is arranged in the same target space and then vectorized in a low dimension,
User interaction prediction server for web applications. In a method for predicting user interaction for a web application in a user terminal,
(a) collecting event information including a navigation event and a click event for URL search that occurs when browsing a web application through a network;
(b) generating user behavior analysis data by applying a preprocessing process including data filtering and data vectorization for performance improvement to the collected event information;
(c) generating prediction result data including an event to occur after the new event by inputting the new event occurrence information into a user interaction prediction model when a new event occurs according to a web application search in the user terminal; And
(d) adjusting a provision state of a service to be provided to the user terminal with respect to an external application service providing server based on the prediction result data; including,
The user interaction prediction model is learned using a recurrent neural network (RNN) based on the user behavior analysis data,
How to predict user interaction for web applications. The method of claim 11,
The pre-processing process includes a URL grouping process for recognizing URLs satisfying a preset condition as one integrated group,
How to predict user interaction for web applications. The method of claim 11,
The pre-processing process includes a web embedding process in which the event information is arranged in the same target space and then vectorized in a low dimension,
How to predict user interaction for web applications. A non-transitory computer-readable recording medium on which a computer program for performing the method according to claim 1 is recorded.

Images