Research questionnaire matching method based on AI behavior recognition

Abstract

The application discloses a survey questionnaire matching method based on AI behavior recognition and belongs to the technical field of artificial intelligence behavior analysis and online survey questionnaire accurate matching. The method comprises the following steps: S1, semantic demand setting; S2, real-time data acquisition and preprocessing; S3, data analysis and demand matching; S4, target questionnaire pop-up; and S5, report and suggestion output. The application adopts the above-mentioned survey questionnaire matching method based on AI behavior recognition, realizes simple rule setting, exempts from the early portrait cost, accurately matches the current situation of the user, and greatly improves the real-time performance and accuracy of the questionnaire matching through the semantic demand description, real-time page behavior acquisition and AI large model intention analysis technology.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence behavior analysis and precise matching of online survey questionnaires, specifically involving a survey questionnaire matching method based on AI behavior recognition. Background Technology

[0002] In various scenarios such as e-commerce, online services, and APP operation, surveys are key tools for collecting user feedback and optimizing products and services, and the accuracy of survey matching affects survey efficiency and data value.

[0003] However, existing technologies have the following shortcomings: specific rules need to be predefined, the configuration process is cumbersome, and incomplete rule settings can easily lead to matching errors; they rely on building user profiles in advance, which requires a large amount of work in the early stages and is also prone to data deviations; they are all based on fixed rules or historical data, which cannot accurately determine the user's current real-time situation, and the matching lacks timeliness and specificity.

[0004] Therefore, a new method is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a survey questionnaire matching method based on AI behavior recognition. This method uses semantic requirements description, real-time page behavior collection and AI large model intent analysis technology to achieve simple rule setting, eliminate the cost of preliminary profiling, accurately match the user's current situation, and greatly improve the real-time performance and accuracy of questionnaire matching.

[0006] To achieve the above objectives, this invention provides a survey questionnaire matching method based on AI behavior recognition, comprising the following steps: S1. Users input their contextual semantic requirements and mark their priorities. The semantic rules are then transformed into system technical rules through a rule generation algorithm. After word segmentation and stop word removal, the semantic feature vectors are extracted and semantic feature confidence is calculated using the LoRA lightweight fine-tuned BERT-base model. The results are then transmitted synchronously after being associated with priority tags. S2. Pre-configure collection rules, capture user page operation data in real time through event listening and DOM structure parsing technology, and generate structured behavioral data after data cropping, outlier removal, standardization, time series feature extraction and differential privacy protection processing, and transmit it in encrypted form. S3. Obtain the semantic feature vector of S1 and the structured behavioral data of S2. Combine historical event data and the current DOM structure, and obtain the weighted matching degree through feature vector mapping, AI large model intent recognition, dynamic weight optimization and cosine similarity calculation. This will assist in questionnaire title analysis and optional user follow-up questioning mechanism to select the best matching questionnaire. S4. Calculate dynamic thresholds based on business load, user activity, and time period information. When the weighted matching degree meets the standard, trigger pop-up rules based on user operation status, collect questionnaire filling status data in real time, and transmit it in encrypted form. S5. Decipher the questionnaire feedback data and behavior analysis results, statistically analyze the core indicators and the degree of correlation between behavior and feedback, predict users' potential needs through the XGBoost model, generate personalized reports, and optimize model parameters in reverse with the help of the federated learning framework to form a self-learning closed loop.

[0007] Preferably, in S1, the contextual semantic requirements support single-intent description or multi-intent combination description; The BERT-base model has a hidden layer dimension of 768, 10 fine-tuning rounds, and a learning rate of [missing information]. The formula for calculating the confidence score of semantic features is: ; In the formula, The confidence score is the semantic feature confidence score; the number of successfully matched intent slots is the number of key elements extracted from the user's semantic needs that were successfully identified and matched by the system; the total number of intent slots is the total number of all key elements obtained after the user's semantic needs were extracted; the model prediction probability is the confidence probability of the AI ​​model that processes semantic needs for the current semantic needs recognition result. The priority weight calculation formula is as follows: ; In the formula, The priority weight is assigned to the requirement; the priority set by the user is the priority of the requirement labeled by the user; the historical matching success rate is the proportion of the semantic requirement that was successfully matched with the questionnaire in past scenarios.

[0008] Preferably, in S2, the data collection dimensions include basic behavioral data, operation sequence data, and environmental feature data, with a collection frequency of 100ms / time. An isolated forest algorithm is used to remove outliers, constructing 50 isolated trees with a depth limit of 10. Each tree randomly samples 200 samples. An outlier score s ≥ 0.6 is considered an outlier. The formula for calculating the outlier score is: ; ; Data dimensions are standardized using Z-score, and time series features are extracted using a sliding window method with a window size of 5s. Laplace noise is used for differential privacy protection, and finally 128-dimensional structured behavioral data is generated and transmitted via TLS 1.3 protocol.

[0009] Preferably, the Laplace noise formula is: ; In the formula, This represents the original number of clicks; Privacy-preserving data obtained after adding noise; For privacy budget, ; The random noise value follows a Laplace distribution with a location parameter of 0 and a scale parameter of 1.

[0010] Preferably, the structured behavioral data is mapped into a 768-dimensional behavioral feature vector through a fully connected layer, which is then input into a large model to identify user intent and output confidence scores. The dynamic weight optimization strategy updates the behavioral feature weights every 500 samples. The weight update formula is as follows: ; In the formula, For the first The cross-entropy loss gradient of class features is used; if the gradient is positive, the weights are adjusted in the positive direction. For the first The updated weights of class-behavioral features; For the first The current weight of the class behavior feature before the update; For the Sigmoid function; The user follow-up questioning mechanism is triggered when the confidence level of intent is less than 0.7.

[0011] Preferably, in S3, the matching degree between the semantic feature vector and the behavioral feature vector is calculated using cosine similarity, and the weighted matching degree calculation formula is: ; In the formula, For weighted matching degree; Priority weights for requirements; The number of behavioral feature categories involved in the calculation; For the first Class behavioral characteristics in the current round Dynamic weights; For the first The matching degree corresponding to the class behavior features.

[0012] Preferably, in S4, the dynamic threshold The calculation rules are as follows: When the query rate per second is >100,000 or the user activity is ≥5 times / month, =0.6; Between 00:00 and 06:00 in the morning or in sensitive business scenarios, =0.75; In other scenarios =0.7; The pop-up window occupies 1 / 3 of the screen height and does not obscure the core operation buttons. It provides "Do not fill in for now" and "Close" options. The "Do not fill in for now" preference is valid for 7 days and will pop up 24 hours later in the same scenario. The questionnaire completion status data is encrypted using the AES-256 encryption algorithm.

[0013] Preferably, the user follow-up questioning mechanism is as follows: After collecting additional intent information from the user by displaying a pop-up radio button on the page, the matching degree is recalculated.

[0014] Therefore, the present invention employs the above-mentioned AI-based behavior recognition-based questionnaire matching method, and compared with the prior art, the technical solution of the present invention has the following beneficial effects: (1) The use of semantic scenario requirement description technology overcomes the technical problems of "cumbersome rule configuration and matching deviation", and achieves the technical effect of "simple rule setting and reduced configuration cost". (2) The use of real-time collection of user page behavior data for scene analysis technology overcomes the technical problems of "large workload and data deviation in user profiling", and achieves the technical effect of eliminating the cost of early profiling and improving data accuracy. (3) Use real-time behavioral data (combined with 10 historical events + current DOM structure) for matching to accurately match the user's current situation and solve the problem of insufficient real-time performance caused by reliance on historical data.

[0015] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating an embodiment of the survey questionnaire matching method based on AI behavior recognition of the present invention. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Unless otherwise defined, the technical or scientific terms used in the present invention should have the ordinary meaning understood by those skilled in the art.

[0018] Example 1 like Figure 1As shown, this embodiment provides a survey questionnaire matching method based on AI behavior recognition. It should be understood that the specific parameters, models and protocols mentioned in this embodiment are only examples to help those skilled in the art understand the present invention, and are not intended to limit the present invention.

[0019] The survey questionnaire matching method based on AI behavior recognition of the present invention includes the following steps: S1. Users can input contextual semantic needs through the interface. Single intents are supported, such as popping up a questionnaire when a user hesitates to place an order or when a user is indecisive. Multiple intents can also be combined and described in the following format: [User characteristics] + [Behavioral actions] + [Focus dimensions]. For example, 25-35-year-old working users + abandon payment + focus on logistics timeliness. Users can also mark priorities through drop-down boxes, with level 1 being the highest and level 5 being the lowest. After receiving the requirements, the algorithm generates rules to configure the AI ​​with corresponding personas and requirements, and transforms semantic rules into technical rules that conform to the system. For example, when a payment is abandoned, a questionnaire is triggered. The algorithm will automatically understand the text through the AI, analyze the website structure, and map it to the click logic of the corresponding page and button. The text is then split using a word segmentation tool. For example, if a user who works overtime abandons payment and focuses on the price, it will be split into overtime / user / abandon payment / focus on price. At the same time, meaningless words such as "user" are removed from the stop word list. Then, the LoRA-based lightweight fine-tuned BERT-base model is used, which has a hidden layer dimension of 768, 10 fine-tuning rounds, and a learning rate of [missing information]. By randomly masking 15% of the input token using a masked language modeling task, deep semantic features are extracted, generating a 768-dimensional standardized semantic feature vector. Simultaneously, the semantic feature confidence score is calculated using the following formula: ; In the formula, The semantic feature confidence score is defined as follows: the number of successfully matched intent slots is the number of elements successfully identified and matched by the system among the key elements (i.e., "intent slots") extracted from the user's semantic needs. For example, if the need "overtime + cancel payment + focus on price" is extracted into 3 slots, and all 3 are successfully identified, this value is 3; the total number of intent slots is the total number of all key elements (intent slots) obtained after extracting the user's semantic needs; the model prediction probability is the confidence probability of the AI ​​model processing the semantic needs for the current semantic needs recognition result, with a value range of 0-1. A demand is considered valid when the confidence level is ≥0.85. For example, if all three slots "overtime + forfeit payment + price focus" match successfully and the model predicts a probability of 0.92, then the demand is considered valid. After associating with the user-defined priority tags, the demand is synchronously transmitted to S2 via HTTP interface. The priority weight formula is as follows: ; In the formula, Weights are assigned to priority requirements; user-defined priorities are the user-labeled priority of the requirements; historical matching success rate is the proportion of questionnaires in which semantic requirements have been successfully matched in past scenarios; S2. Pre-configure collection rules, with collection dimensions including three core data categories; The first category is basic behavioral data, covering the number of clicks (C), dwell time (T), and number of abandoned operations (F), collected at a frequency of 100ms / time. The second category is operation sequence data, including click paths and operation intervals (Δt), such as dwell time on the details page and dwell time when clicking the checkout button. The third category is environmental characteristic data, divided into six time periods, such as the morning peak from 7-9 am, and device types, such as mobile / PC / tablet. Relying on HTML's onclick / onmouseover event listeners and DOM structure parsing technology, user page operation data is captured in real time. Unnecessary data such as CSS is first filtered out through front-end data trimming algorithms, compressed into the minimum data, and then transmitted to the server via the WebSocket protocol. The server uses the Isolation Forest algorithm to remove outliers. The calculation formula is as follows: ; ; In the formula, These are abnormal scores, with values ​​ranging from [0-1]. For the sample Average path length among 50 trees; It is the harmonic number; The target data sample to be detected; The number of random samples used when constructing the isolated tree (iTree); For the sample size The expected path length at that time; Construct 50 isolated trees with an iTree depth limit of 10. Randomly sample 200 samples from each tree and calculate anomaly scores. When s ≥ 0.6, it is considered an outlier. For example, a stay time of 1000s that far exceeds the normal threshold is considered an outlier. Then, the data units are standardized using Z-score, and the calculation formula is as follows: ; In the formula, This represents the data mean. Standard deviation; This is the original data; Next, a sliding window method with a window size of 5 seconds is used to extract time series features, such as the gradient of the interval between three consecutive clicks. Finally, differential privacy technology is used to add Laplace noise. The formula for adding differential privacy noise is expressed as: ; In the formula, This represents the original number of clicks; Privacy-preserving data obtained after adding noise; Budget for privacy; The random noise value follows a Laplace distribution with a location parameter of 0 and a scale parameter of 1. The privacy budget of this noise Ultimately, 128-dimensional structured behavioral data is generated and transmitted to S3 via a TLS 1.3 encrypted transmission channel. S3. Obtain the 768-dimensional semantic feature vector of S1 and the 128-dimensional structured behavioral data of S2. The semantic feature vector includes priority weights. At the same time, the analysis is carried out based on the real-time matching engine logic, combined with 10 historical event data and the current DOM structure. First, the behavioral data is mapped to a 768-dimensional behavioral feature vector through a fully connected layer. Then, it is input into the Llama2-7B large model optimized by prompt. The prompt template is user behavior data: {Number of clicks: C, Duration of stay: T, Operation path: ...}. Please identify the user's current intent and rate it (0-1 points). The model outputs the intent confidence score. The behavioral feature weights are updated every 500 samples using a Dynamic Weight Optimization (DWO) strategy. The weight update formula is: ; In the formula, For the first The cross-entropy loss gradient of class features is used; if the gradient is positive, the weights are adjusted in the positive direction. For the first The updated weights of class-behavioral features; For the first The current weight of the class behavior feature before the update; This is the Sigmoid function, which maps the loss gradient to the interval between 0 and 1. The matching degree between semantic feature vectors and behavioral feature vectors is calculated using cosine similarity. The calculation formula is: ; In the formula, For vectors The cosine similarity results range from -1 to 1; For vectors The modulus length; For vectors The modulus length; Then, the final weighted matching degree is obtained by combining the priority weights of the requirements. The calculation formula is as follows: ; In the formula, Weighted matching degree; Preliminary behavior-semantic matching score; Priority weights for requirements; The number of behavioral feature categories involved in the calculation; For the first Class behavioral characteristics in the current round Dynamic weights; For the first The degree of matching corresponding to class-behavioral features; Additionally, a questionnaire title analysis step is added to assist in determining the best matching questionnaire; when the intent confidence score is <0.7, a user follow-up question mechanism is triggered, and a single-choice pop-up window appears on the page with the reason for your abandonment of the operation (such as the price being too high or the logistics being too slow). After supplementing other intent information, the matching score is recalculated.

[0020] Finally, the survey questionnaire with the highest FinalScore is selected, and its data such as questionnaire ID, question content, and display style are encapsulated in JSON format and sent to S4. S4. After receiving the questionnaire data in JSON format, first read the system's real-time monitoring data on business load, user activity (interaction frequency in the past 30 days), and time period information. Then, calculate the current threshold using a dynamic threshold formula, which is: ; In the formula, This is the current threshold. Queries per second; Scenarios with low user activity; Sensitive business scenarios are those involving user privacy and transaction risks; User activity ≥ 5 times / month means that the user has interacted with the platform ≥ 5 times in the past 30 days, indicating that the user belongs to a high-frequency active group; When FinalScore ≥ T, a pop-up mechanism is triggered. First, the user's current operation state is detected. If the user is in an input state, such as filling in a shipping address, the pop-up will be delayed until 1 second after the operation is completed. Otherwise, a pop-up window will immediately appear at the bottom of the page. This pop-up window occupies 1 / 3 of the screen height and does not obscure the core operation buttons. The pop-up window has two options, "Do not fill in for now" and "Close", in the upper right corner. If the user selects "Do not fill in for now", the preference will be recorded and stored in the user's local cache for 7 days. In the same scenario, the pop-up window will be delayed by 24 hours in the future. The system collects questionnaire completion status data in real time, including whether the questionnaire has been submitted, the completion time, and the skip rate. 0 indicates that the questionnaire has not been submitted and 1 indicates that it has been submitted. The completion time is the time from when the pop-up window appears until the questionnaire is submitted or closed. The skip rate is the ratio of the number of unanswered questions to the total number of questions. This data is encrypted using the AES-256 encryption algorithm and then transmitted to S5. S5 receives the encrypted questionnaire feedback data from S4 and the behavioral analysis results from S3, restores the data through a decryption algorithm, and statistically analyzes two types of core indicators: one is the average satisfaction rate, which is the arithmetic mean of all user ratings, and the other is the option distribution ratio, which is the ratio of the number of people who selected a certain option to the total number of submissions. The correlation between behavioral data and questionnaire feedback scores is calculated using the Pearson correlation coefficient. When the absolute value of r is ≥0.6, it is considered a strong correlation. The XGBoost prediction model is introduced. This model has a learning rate of 0.1, a number of decision trees of 100, and a maximum tree depth of 6. After inputting behavioral features and feedback data, it outputs the predicted probability of potential user needs. Two types of personalized reports are generated: the operational version report includes scenario matching rate, user feedback hotspots, and conversion optimization suggestions; the product version report includes behavior-feedback correlation analysis and feature iteration direction. The federated learning framework uses the FedAvg algorithm to aggregate 5 business line servers (with a cycle of 24 hours) and updates the global model with anonymized matching strategies, such as semantic extraction weights and dynamic weight rules. This, in turn, optimizes the LoRA fine-tuning parameters of S1 and the initial values ​​of dynamic weights of S3, forming a self-learning closed loop.

[0021] Therefore, the present invention adopts the above-mentioned AI behavior recognition-based survey questionnaire matching method. This method uses semantic demand description, real-time page behavior collection and AI large model intent analysis technology to achieve simple rule setting, eliminate the cost of preliminary profiling, accurately match the user's current situation, and greatly improve the real-time performance and accuracy of questionnaire matching.

[0022] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0023] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A survey questionnaire matching method based on AI behavior recognition, characterized in that, Includes the following steps: S1. Users input their contextual semantic requirements and mark their priorities. The semantic rules are then transformed into system technical rules through a rule generation algorithm. After word segmentation and stop word removal, the semantic feature vectors are extracted and semantic feature confidence is calculated using the LoRA lightweight fine-tuned BERT-base model. The results are then transmitted synchronously after being associated with priority tags. S2. Pre-configure collection rules, capture user page operation data in real time through event listening and DOM structure parsing technology, and generate structured behavioral data after data cropping, outlier removal, standardization, time series feature extraction and differential privacy protection processing, and transmit it in encrypted form. S3. Obtain the semantic feature vector of S1 and the structured behavioral data of S2. Combine historical event data and the current DOM structure, and obtain the weighted matching degree through feature vector mapping, AI large model intent recognition, dynamic weight optimization and cosine similarity calculation. This will assist in the analysis of the questionnaire title. When the intent confidence is lower than the threshold, the user follow-up questioning mechanism will be activated to select the best matching questionnaire. S4. Calculate dynamic thresholds based on business load, user activity, and time period information. When the weighted matching degree meets the standard, trigger pop-up rules based on user operation status, collect questionnaire filling status data in real time, and transmit it in encrypted form. S5. Decipher the questionnaire feedback data and behavior analysis results, statistically analyze the core indicators and the degree of correlation between behavior and feedback, predict users' potential needs through the XGBoost model, generate personalized reports, and optimize model parameters in reverse with the help of the federated learning framework to form a self-learning closed loop.

2. The survey questionnaire matching method based on AI behavior recognition according to claim 1, characterized in that, In S1, the contextual semantic requirements support single intent description or multi-intent combination description; The BERT-base model has a hidden layer dimension of 768, 10 fine-tuning rounds, and a learning rate of [missing information]. The formula for calculating the confidence score of semantic features is: ； In the formula, Confidence of semantic features; The number of successfully matched intent slots is the number of key elements extracted from the user's semantic needs that were successfully identified and matched by the system; the total number of intent slots is the total number of all key elements obtained after the user's semantic needs were extracted; the model prediction probability is the confidence probability of the AI ​​model that processes semantic needs for the current semantic needs recognition result. The priority weight calculation formula is as follows: ； In the formula, The priority weight is assigned to the requirement; the priority set by the user is the priority of the requirement labeled by the user; the historical matching success rate is the proportion of the semantic requirement that was successfully matched with the questionnaire in past scenarios.

3. The survey questionnaire matching method based on AI behavior recognition according to claim 2, characterized in that, In S2, pre-configured data collection rules are used to capture user page operation data in real time through event listening and DOM structure parsing technology. The collection dimensions include basic behavioral data, operation sequence data, and environmental feature data, with a collection frequency of 100ms / time. An isolated forest algorithm is used to remove outliers, constructing 50 isolated trees with a depth limit of 10. 200 samples are randomly sampled from each tree. An outlier is considered an outlier when the outlier score s ≥ 0.

6. The outlier score is calculated using the following formula: ； ； In the formula, The target data sample to be detected; The number of samples randomly sampled when constructing the isolated tree; For the sample A path in an isolated forest; For the sample Average path length across all isolated trees; It is the harmonic number; This represents the expected path length when the number of samples is n. These are abnormal scores; Data dimensions are standardized using Z-score, and time series features are extracted using a sliding window method with a window size of 5s. Laplace noise is used for differential privacy protection, and finally 128-dimensional structured behavioral data is generated and transmitted via TLS 1.3 protocol.

4. The survey questionnaire matching method based on AI behavior recognition according to claim 3, characterized in that, The Laplace noise formula is: ； In the formula, This represents the original number of clicks; Privacy-preserving data obtained after adding noise; For privacy budget, ; The random noise value follows a Laplace distribution with a location parameter of 0 and a scale parameter of 1.

5. The survey questionnaire matching method based on AI behavior recognition according to claim 4, characterized in that, In S3, structured behavioral data is mapped into 768-dimensional behavioral feature vectors through a fully connected layer. These vectors are then input into a large model to identify user intent and output confidence scores. The dynamic weight optimization strategy updates the behavioral feature weights every 500 samples. The weight update formula is as follows: ； In the formula, For the first The cross-entropy loss gradient of class features is used; if the gradient is positive, the weights are adjusted in the positive direction. For the first The updated weights of class-behavioral features; For the first The current weight of the class behavior feature before the update; For the Sigmoid function; The user follow-up questioning mechanism is triggered when the confidence level of intent is less than 0.

7.

6. The survey questionnaire matching method based on AI behavior recognition according to claim 5, characterized in that, In S3, the matching degree between semantic feature vectors and behavioral feature vectors is calculated using cosine similarity. The weighted matching degree calculation formula is as follows: ； In the formula, Weighted matching degree; Priority weights for requirements; The number of behavioral feature categories involved in the calculation; For the first Class behavioral characteristics in the current round Dynamic weights; For the first The matching degree corresponding to the class behavior features.

7. The survey questionnaire matching method based on AI behavior recognition according to claim 6, characterized in that, In S4, the dynamic threshold The calculation rules are as follows: When the query rate per second is >100,000 or the user activity is ≥5 times / month, =0.6; Between 00:00 and 06:00 in the morning or in sensitive business scenarios, =0.75; In other scenarios =0.7; The pop-up window occupies 1 / 3 of the screen height and does not obscure the core operation buttons. It provides "Do not fill in for now" and "Close" options. The "Do not fill in for now" preference is valid for 7 days and will pop up 24 hours later in the same scenario. The questionnaire completion status data is encrypted using the AES-256 encryption algorithm.

8. The survey questionnaire matching method based on AI behavior recognition according to claim 7, characterized in that, The user follow-up questioning mechanism is as follows: After collecting additional intent information from the user by displaying a pop-up radio button on the page, the matching degree is recalculated.

9. A computer device, characterized in that, include: A processor configured to be coupled to a memory, read and execute instructions and / or program code in the memory to perform the method as described in any one of claims 1-8.

10. A computer-readable medium, characterized in that, The computer-readable medium stores computer program code that, when executed on a computer, causes the computer to perform the method as described in any one of claims 1-8.

Images