GB2524730A - Predicting application user behaviour and interests in real-time using predictive analytics as a service - Google Patents

Abstract

This invention enables website/application vendors to optimise the experience for their users by increasing the prominence of sections of the application or content with the application based on the predicted interests of the user. By using statistical models based on past behaviour and the attributes of individual users this invention uses predictive analytics to infer the interest of any aspect of the application for any specific user. The user interest can also be predicted whilst the user is interacting with the application and the user's current activity within the session is used as a factor in determining the items the user will be subsequently interested in.

Description

I

Predicting Application User Behaviour and Interests in Real-Time using Predictive Analytics As A Service

Field of Invention

This invention relates generally to a network connected service that predicts users needs and interests when using a web, mobile or PC client application.

Background of the Invention

Increasing amounts of retail is being done online. Everything from weekly groceries to holidays can be brought over the Internet. As online retailers are not constrained by physical space they can carry hundreds, thousands or tens of thousands of stock lines. The problem they now face is that when visitors come to their site the visitor can't quickly find what they are looking for. As there are usually many alternate sites on the Internet for the visitor to choose from, if the visitor does not find what they want quickly they wifl move on to another site. In order to be successful the website must find ways to assist the user to find what they need, in the shortest time and least number ot clicks possible.

This same problem is manifested in other online applications. The last 2 years have seen an explosion in streaming media. Sums, TV shows, live events and music can all be accessed online on a pay-per-use, subscription or advert funded model. AU such sites have potentially millions of streaming media items for the user to access. Again the value to the user is only achieved if they can find the items that they are looking for quickly and easily. Site owners must find ways to not oniy assist the user to find what they are looking for, but also keep them engaged with additional items for as long as possible.

Similarly for news, magazine or content heavy editorial sites the problem facing site owners is how to keep users engaged and browsing their site tor as long as possible and how to increase repeat visits. They must constantly allow the user to discover the contenc items that are of interest to them.

Disclosure of Invention

The proposed invention abstracts the data storage and processing of predictive analytics such that it can be used by any website owner without the need for them to understand the data science or complex bespoke progranuring required. The website owner adds one line of javascript to their site and au activity is lagged to the Predictive Analytics Service (PAP) provided by the invention. The PAP will then anaiyse all activity iogged. For any visitor arriving on the customer wehsite the application logic of the customer website can query the service to retrieve a list of the items that analysis shows will he of most interest to the visitor. Then as the visitor browses the site the service will refine its predictions based on the items that visitor has already interacted with.

The invention comprises of a service ohat is accessible over a computer network (such as the pubic internet) . The client application wiii call methods on the service to iog all relevant user activity. The client application refers to any software-based application that a user interacts with to complete a task.

If the application contains a shop, catalogue or database of items, each item viewed by the user can also be passed into the service. The application vendor can specify goals within their application that they want application users to achieve e.g. a product purchase or item download. Details of the goal and any associated items are also passed to the service. The application can also pass any relevant attributes about the visitor/user. The attributes are not restricted and can be defined dynamically by the application vendor.

Examples include age, gender, interesos, preferences, and income range.

The service will analyse data iogged by the appiication to find patterns of user behaviour.

The service will also enrich the visiror data by importing data from parties or the customer's own data sources. The enrichment process will identify additional visitor/user attributes that can be used in the statisticai analysis process to generate a predictive interest rank.

The appiication vendor or website owner can interrogate the service to retrieve predictive analytics. For any page or section or item within the application the service wiii return an overali interest rank that will indicate the total interest in that page/section/item relative to au other pages/sections/items in the application/website. For any given user/visitor the service will return the predicted interest rank tor any given page/item or the most popular pages/items for that visitor. It will also be possible to interrogate the service whiist the user/visitor is interacting with the application/website. In this instance the service will return the pages/items most likeiy to be of subsequent interest to the user. The interest prediction will be based on statisticai analysis ot past behaviour of users/visitors and the known attributes of the given visitor and that visitor' s current interaction with the application/website.

The results of the analysis can be viewed by application/website owners through a dashboard type reporting tool. The reporting tool wili show historical information on the relative interest of all pages/items of the application/website. It is aiso possible to see au visitors to the application/website with their usage history. It is also possibie to see all visitor/user attributes, how those attributes correlate to interest in pages/items and how the interest has changed over time.

The benetit to the owner ot the websie, application or other eiectronic store front is that they can serve visitors to their site with content items that would otherwise not have been tound and they can keep visitors on their site for longer or increase repeat visits by providing the most relevant content. This in turn increases the revenue for the site owner (through what ever revenue model they employ) . This is achieved without any knowledge of large-scale data analytics, statistical analysis or customer application development. This allows smaii and medium sized sites to have access to the same technology, and therefore have the same competitive advantage, as large sites. But without the need for investment in bespoke programming teams and without the need to understand the dam science that underlies it.

The benefit to the site visitor is that they can find items they need, or are interested in, faster. They spend less time navigating items that are not what they require. They are more likeiy to discover items that they are interested in, but were not immediately searching for.

Brief Description of Drawings

For a more complete explanation of the present invention and the technical advantages thereof, reference is now made to the following description and the accompanying drawing in which: Figure 1 shows a logical architecture of the invention.

Figure 2 shows logic flow diagram of The invention recording user activity within an application and predicting user interest within a session Figure 3 shows the topologicai networl< options tor vendor applications to connect to the Predictive Analytics Service

Best Mode for Carrying Out the Invention

Embodiments of the present invention and their technical advantages may be better understood by referring to Figure 1.

A session is defined as a user's (Figure 1:131) continuous interaction with a application/webslte (Figure 1:132) to complete a task. A session may have discrete start and end events or may be defined through periods of inactivity.

A page or screen is defined as any distinct section or screen within the application.

Application content items are discrete entities within the vendor application. For example: * In a. shopping application this would represent items for saie.

* In a news application these would represent news stories.

* In a music streaming application these would he individual songs.

Visitor attributes are any additional data that the application vendor stores about each application user. For example: * Demographic (age, geographic location, income) * Product related (licence type, customer type) * Business reiated (job function, business type, market segment) * Other, as defined by the application vendor.

It will be possible for the application vendor to define a number of Goals within the application. These will be some target point within the application that the vendor wishes the user to attain. For example: * Product purchase * Item download * Account signup.

The application will log activity with the invention by calling an Application Programming Interface (API) exposed to the application (Figure 1:133). The API will be accessed over a computer network (Figure 1:104).

The application will send visitor attributes over the API.

The API will store all data passed in from the application in a persistent data store (Figure 1:139).

The application will retrieve details of user interests via the API (Figure 1:135).

The Predictive Analytics Service (lAS) (Figure 1:106) may access 3rd party systems (Figure 1. 107 to discover further information about each visitor.

For exampie: * Lookup I? address to determine geographic location * Lookup company to determine business sector and business size * Lookup individual to determine user demographics * Lookup electoral roli intormation for a visitor * Lookup credit information for a visitor The data from 3rd party systems will be stored in a persistent data store (Figure 1:109) along with any attribces passed in from the application.

The Predictive Analytics Service (PAP) (Figure 1:106) may access the application vendors' own data repositories (Figure 1:108) to retrieve additional attributes about the visitor. These attributes will be stored in a persistent data store (Figure 1:109) along with the data from 3rJ party sources and attributes passed directly from the application.

The Predictive Analytics Service (PAS) (Figure 1:106) will analyse activity data logged by the application as a background activity.

For each visitor to the site the Predictive Analytics Service (PAP) looks up additional data about the visitor. From the IF address it is possible to determine the country from which they accessed the site, the company trom which they accessed the site, or if the visitor is accessing the site from a domestic internet provider.

The PAP will also lookup the visitor with 3rd party service that will provide information about other sites the visitor has accessed.

It is also possible for the site to pass into the PAP any other attributes that are of relevance to the site owner. It is not necessary for the PAP to know in advance what attributes will be passed by the site. It is also not necessary for the site to pass the same set of attributes for all visitors.

The PAP analyses data for each site daily. First, for each visitor the interest rank is calculated for every item available on the site. The interest rank is a function of how often any given site item has been accessed by a visitor and how long ago they accessed it. To compute this a weighted moving average is calculated using exponential smoothing.

tv1t* wjt1 f -Where: st = the interest rank for any given item x = the unique count of visits for that user of the given item at time t t = a point in time where the unique visit count was taken (e.g. daily) w = the weighting factor appiied to each visit count The weighting factor will be such that the sum ot all w values (wn) = 1. This allows more prominence to be placed on the items that were viewed more recently.

The weighting factor will start at a constant value for all visitors to all sites. Bowever the relevance with respect to time will be different for all sites e.g. relevance will decay over rime much faster on a current attairs news site compared to a specialist sports equipment retailer. To account for this the predicted interest wiil be compared every day to the actuai interest. The error value will be automatically compared to the actual value.

The size (and sign( of the error value will be used to automatically adjust the weighting factor.

Then for each item the service compiles a list of items that were the next items viewed.

E.g. visitor views items in the following order A->B->A->C->D Indicating that for item A the next items were B and C For item B the next item was A For item C the next item was D Again exponential smoothing is used to compute a weighted moving average to give a prediction of the likelihood that any item will be the next item viewed for every item on the site.

The weighting factor will be similarly adjusted from the error value.

Once an interest rank is generated for every item, and a next item array is generated for every item, the attribce analysis can begin.

For each attribute a correlation is calculated to look for the presence of a linear relationship between the attribute and a high interest rank in any given item.

Consider a set of a paired observations (xl,yl), (x2,y2).. (xn, yn) CS(x, y) = LX xy1 -(t) 3) Where correiation between visitors with attribute A having interest in item B x = visitor interest rank in item B y = views of item B by visitors with attribute A The correiation coefficient is calculated as: r = CS(x,y)/,J[CSx,x)CS(y,y)] Such that: * r wiii be between -l and +1 * Positive r: y increases as x increases. Negative r: y decreases as x increases * If x and y are perfectly lineariy correiated r = -l or + 1 * r ciose to zero indicates that x and y are not linearly correiated Where a significant iinear reiationship has been established (correiation coefticient tested for signiticance using a hypothesis test) the attributes are passed through a multipie iinear regression.

The dependent variable is interest rank of item B The independent variables are the views of item B for each different attribute.

The interest rank, attribute correlation and next page array can then be interrogated to establish at any poinc what a visitor will be interested in

next. For example:

When the visitor first arrives at the site the PAS will already know: * The time of day * Day of the week * The country the visitor is accessing the site from * The company from which they are accessing the site (or it it is a domestic user) * It may also be possible to access their past browsing history and obtain a list of other sites they have visited It also knows the first page that they have landed on.

The site may also pass in other attributes known about the visitor.

From these attributes it can then look up which items have the highest correlation with interest rank. This can be conjoined with the list of items in the next page array for the initiai landing page to produce a sub set ot items in order of highest interest for this user. These can be presented to the user betore they have interacted (clicked on) with the site in any way.

Once the user has selected their item of interest the PAS can refine the interest prediction as follows: Look at the attributes of other visitors that are highly correlated with interest in the selected item. Find the union set of attributes between these and that of the visitor. Then find the items that are highly correiated with interest amongst this combined set of attributes. Merge that with the next item array of the selected item to produce a sub set ot items in order ot highest interest for this user.

The process can be repeated and refined with each item the user selects.

Figure 3 shows a number of possible network configurations for vendor applications to connect to the Predicive Analytics Service.

Figure 3:303 shows a user interacting with a web browser based application.

The browser application is accessed from a web server (Figure 3: 302) . The web server connects to the Predictive Analytics Service.

Figure 3:303 shows a desktop application. As the user interacts with the application the application connects directly to the Predictive Analytics Service.

Figure 3:304 shows the invention working in a machine-to-machine scenario. In this arrangement the application is not working in direct response to a human user input. However, the predictive analysis service will still use past activity and session attributes to predict the likelihood of future scenarios.

Figure 3:305 shows a mobile client application connecting over a radio network (Figure 3:306).

Claims (3)

1. Claims 1. Within any usage session in an appiication/website predict the next pages/items nf interest to a. visitor given the known attributes about the visitor and their activity within thac session.
2. 2. For any visitor/user ot an applicacion predict the level of interest in any page/item within the appiication.
3. 3. Provide the functionality in claim 1 and ciaim 2 to application/website vendors as a service that they access over a network connection.