New alternative item selection prediction modeling method and system based on latent consideration mechanism

Abstract

The application provides a novel alternative item selection prediction modeling method and system based on a latent consideration mechanism, comprising: obtaining willingness survey selection task data and arranging the data into panel data, recording a selection task presentation set and determining a target novel alternative item and a non-target alternative item for this modeling; calculating a latent consideration probability of the target novel alternative item predicting entering an effective consideration set and a complement set thereof based on a respondent level variable, a task level variable and individual random effects; calculating a first conditional selection probability and a second conditional selection probability respectively, constructing a two-state observation likelihood according to an observed selection result, and building a prediction model through simulated integration to generate a prediction result associated with the probability of selection. The application can distinguish between not considering and not selecting after considering, reduce the bias caused by the full consideration hypothesis, and improve the reliability of selection prediction.

Description

Technical Field

[0001] This invention relates to the field of selection prediction modeling technology, and in particular to a novel alternative selection prediction modeling method and system based on a potential consideration mechanism. Background Technology

[0002] In evaluating travel modes, products / services, and policy solutions, surveys or stated preference experiments are often used to collect respondents' choices among multiple options. Discrete choice models, such as multinomial logit, nested logit, mixed logit, and latent category models, are then used to estimate the impact of option attributes on the probability of selection. US Patent 8682709B2 discloses a choice behavior modeling system that uses a parametric discrete choice model to assess the probability of decision-makers choosing different options and uses attribute parameters to reflect attribute importance. US Patent 8799186B2 discloses a choice modeling system for survey data, used for market research or preference prediction. While these techniques can handle similarities or competition among options, they are primarily based on the "full consideration assumption," meaning they assume all presented options participate in the probability calculation and do not establish a separate latent consideration stage for situations where novel options may be excluded from the effective consideration set.

[0003] Regarding research on attribute focus and information processing strategies, Hensheer et al. pointed out that respondents may ignore certain attributes, and failure to consider this behavior can lead to biases in the estimation of willingness to pay or attribute coefficients. However, such methods focus on the ignoring of attributes at the "attribute level" rather than the overall disregard for alternatives at the "alternative level," making it difficult to explain the phenomenon of novel alternatives being excluded entirely due to a lack of experience.

[0004] Existing consideration sets and two-stage choice models decompose decision-making into two stages: consideration set formation and final selection. For example, Assele et al. explored a scheme in a stated preference survey to first collect consideration options and then collect selection results, comparing latent consideration set models with mixed Logit models. These methods rely on explicit consideration data or general screening rules and have not yet developed a nested testable modeling system for the problem of "single or a few novel options being asymmetrically excluded due to a lack of behavioral experience."

[0005] In research on new modes of transportation (such as urban air mobility and electric air taxis), demand forecasting often relies on intention surveys and discrete choice models. Existing studies have used latent categories, mixed logit methods, or a combination of both to identify heterogeneity in observed and unobserved preferences. However, these studies typically treat new modes of transportation in experiments as alternatives already evaluated by respondents, failing to separate their inclusion in the effective consideration set from conditional choices. If some respondents do not consider new modes of transportation as feasible options, their observed data will have limited ability to identify the marginal utility of specific attributes. Directly incorporating such observed data into the full consideration model will distort the constant term and attribute coefficients of new modes of transportation.

[0006] In summary, the existing technologies have the following main drawbacks: (1) The full consideration model cannot distinguish between "not considering the target new alternative" and "not choosing the new alternative after considering it"; (2) When respondents are unfamiliar with the new alternative, the low selection rate is absorbed into the constant term, resulting in a systematic downward shift of the constant term; (3) For attributes that only change within the new alternative (cost, waiting, reliability, safety perception, etc.), if the non-considered state is not modeled, the identifiability between attribute changes and selection probabilities will be weakened; (4) The attribute-ignoring model mainly deals with attribute-level neglect and cannot solve the overall non-consideration at the alternative level, especially the asymmetric non-consideration of new alternatives; (5) Existing two-stage models rely on explicit consideration data or general screening mechanisms and do not make full use of repeated selection tasks, psychological variables, and task context variables to jointly identify the potential consideration probability of new alternatives; (6) New transportation preference models such as UAM / eVTOL directly interpret the low selection rate of new modes as low utility, affecting demand forecasting, service design, and policy evaluation. Summary of the Invention

[0007] The purpose of this invention is to solve the following technical problem: In a choice task of intention survey, when new options are presented, but respondents may not include them in the effective consideration set due to lack of experience or cognitive framework, how to accurately separate the "probability of consideration" and the "probability of conditional selection after consideration" through computer modeling, so as to reduce the systematic bias caused by the full consideration assumption.

[0008] To achieve the above objectives, this invention provides a novel alternative selection prediction modeling method based on a latent consideration mechanism, the method comprising: Obtain the task selection data from the intention survey, standardize and organize the data into panel data, record the task selection presentation set for each task, determine a target novel alternative from the task selection presentation set, and take the alternatives other than the target novel alternative as the non-target alternatives in this modeling. Based on respondents' level variables, task level variables, and individual random effects, the potential consideration probability of the target novel alternative being predicted to enter the effective consideration set is calculated, and the probability of the target novel alternative not being predicted to enter the effective consideration set is the complement of the potential consideration probability. When the target novel alternative is included in the valid consideration set, the first conditional selection probability is calculated based on the complete set containing the target novel alternative; when the target novel alternative is not included in the valid consideration set, the second conditional selection probability under the exclusion state is calculated based on the set after removing the target novel alternative. Based on the observation selection results, a two-state observation likelihood is constructed: when the selection result is a new target alternative, the likelihood contribution is the product of the potential consideration probability and the first conditional selection probability; when the selection result is a non-target alternative, the observation is decomposed into a first branch that has considered the new target alternative but selects a non-target alternative, the probability of the first branch being the potential consideration probability multiplied by the first conditional selection probability; and a second branch that selects a non-target alternative without considering the new target alternative, the probability of the second branch being the probability of not entering the effective consideration set multiplied by the second conditional selection probability, and the sum of the probabilities of the two branches is used as the observation likelihood contribution. By simulating and integrating the individual random effects and preference random coefficients on which the probabilities depend under the aforementioned different conditions, and constructing a prediction model, the prediction model is used to generate prediction results related to the probability of selection for the target novel alternative.

[0009] As a further improvement to the above technical solution, the intention survey task selection data includes respondent identifiers, task selection numbers, task selection presentation sets, actual selection results, alternative attributes, respondent attributes, psychological scales, historical experience, and task context variables. The data is standardized and organized into panel data, including: merging multiple selection tasks of the same respondent into panel selection data, forming a three-level index of respondents, selection tasks, and alternatives, and unifying the dimensions of the alternative attributes.

[0010] As a further improvement to the above technical solution, the step of determining the target novel alternative includes: acquiring at least one of the following information from the selected task presentation set: historical usage rate, commercial operation time, respondents' self-reported familiarity, whether they have directly experienced the alternative, whether they have stable evaluation standards, and whether they have formed a regular usage habit; and summarizing the respondents' self-reported familiarity and whether they have directly experienced the alternative for each candidate alternative; when a candidate alternative meets at least one of the following conditions, the candidate alternative is identified as a candidate novel alternative: the historical usage rate is lower than a preset historical usage rate threshold; the commercial operation time is shorter than a preset operation time threshold; the sum of respondents' self-reported familiarity is lower than a preset familiarity threshold; indicating that the number of respondents who have not directly experienced the candidate alternative reaches a certain threshold. The following criteria are applied: a preset quantity threshold or its proportion reaches a preset ratio threshold; whether there are records with stable evaluation standards indicates that there are no stable evaluation standards; whether there are records of established routine usage habits indicates that routine usage habits have not been established; or expert rules or platform rules determine that the candidate alternative does not belong to the category of normalized behavior; when there is only one candidate novel alternative, the candidate novel alternative is determined as the target novel alternative for this modeling; when there are multiple candidate novel alternatives, one is selected as the target novel alternative for this modeling based on at least one of the preset target identifier, business focus object, priority rules, or user-specified rules; alternatives in the selected task presentation set other than the target novel alternative are used as non-target alternatives for this modeling.

[0011] As a further improvement to the above technical solution, the calculation steps of the potential consideration probability include: setting whether the target novel alternative enters the effective consideration set as a potential state; constructing a consideration stage variable based on the respondent-level variable and the task-level variable, and using the individual random effect as the random component of the consideration stage variable; mapping the combination result of the consideration stage variable and the individual random effect to a potential consideration probability with a value range of 0 to 1 through a probability link function. The respondent-level variables are input variables that characterize the characteristics of the respondents, the task-level variables are input variables that characterize the conditions for selecting the task, and the probability link function is one of the probit link function and the logit link function.

[0012] As a further improvement to the above technical solution, the calculation steps of the conditional selection probability include: constructing the utility value of each alternative based on the constant term of the alternative, the attribute of the alternative, and the individual preference coefficient containing random components, and calculating the conditional selection probability through a hybrid Logit model or a random utility model; the conditional selection probability is used for the complete set when the target novel alternative enters the effective consideration set, as the first conditional selection probability; and is used for the exclusion set when the target novel alternative does not enter the effective consideration set, as the second conditional selection probability.

[0013] As a further improvement to the above technical solution, a posterior consideration state probability recovery step is also included: after constructing the prediction model, the posterior consideration state probability corresponding to the observation is recovered based on the likelihood of the two-state observation; when the observation selection result is the target novel alternative, the target novel alternative consideration state corresponding to the observation is determined as considered; when the observation selection result is not the target alternative, based on the relative contribution of the probability of the first branch and the probability of the second branch in the sum of the probabilities of the two branches, the posterior probability of the observation belonging to the target novel alternative that has been considered but not selected, and the posterior probability of belonging to the target novel alternative that has not been considered are determined respectively.

[0014] As a further improvement to the above technical solution, the construction steps of the prediction model include: generating multiple simulated sampling points for individual random effects and preference random coefficients; calculating the potential consideration probability, the first conditional selection probability, the second conditional selection probability, and the observed likelihood contribution for each simulated sampling point; forming a panel likelihood based on the observed likelihood contributions of multiple selection tasks for the same respondent, and forming a simulated panel log-likelihood based on the panel likelihoods corresponding to multiple simulated sampling points; and constructing the prediction model by maximizing the simulated panel log-likelihood.

[0015] As a further improvement to the above technical solution, the multiple simulated sampling points are generated using Halton sequences, Sobol sequences, Latin hypercube sampling, or pseudo-random normal sampling; the individual random effects are set to a normal distribution, and the preference random coefficients are set to a normal distribution, log-normal distribution, triangular distribution, or a sign-constrained transformation distribution; the parameter update of the simulated panel log-likelihood is achieved using BFGS, L-BFGS, Newton's method, quasi-Newton's method, or automatic differential gradient optimization.

[0016] As a further improvement to the above technical solution, the prediction result associated with the probability of selection includes: the potential consideration probability, the first conditional selection probability, the second conditional selection probability, and at least one of the following: the potential consideration probability, the complement of the potential consideration probability, the first conditional selection probability, and the second conditional selection probability. The method further includes at least one of a model verification step and a strategy simulation step; the model verification step includes: comparing the prediction model with a fully considered hybrid Logit model, a task-invariant consideration model, a conservative exclusion constraint model, or a relaxed exclusion constraint model, and generating a model diagnostic result; the strategy simulation step includes: after considering the strategy adjustment parameters of a novel alternative to the target, recalculating the potential consideration probability, the first conditional selection probability, the second conditional selection probability, and the comprehensive selection probability, and generating a strategy simulation result associated with the selection probability; The model diagnostic results include at least one of the following: simulated log-likelihood, AIC, BIC, retained sample prediction accuracy, likelihood ratio statistic, correction magnitude of the constant term of the novel alternative, and correction magnitude of the slope of the novel alternative specific attribute.

[0017] This invention also provides a novel alternative selection prediction modeling system based on a latent consideration mechanism, comprising: The data processing module is used to acquire the intention survey selection task data, standardize and organize the data into panel data, record the selection task presentation set for each selection task, determine a target new alternative from the selection task presentation set, and take the alternatives other than the target new alternative as non-target alternatives in this modeling. The potential probability calculation module is used to calculate the potential consideration probability of the target novel alternative predicted to enter the effective consideration set based on the respondent-level variables, task-level variables and individual random effects, and to determine the probability that the target novel alternative predicted to not enter the effective consideration set as the complement of the potential consideration probability. The conditional selection probability calculation module is used to calculate the first conditional selection probability based on the complete set containing the target novel alternative when the target novel alternative enters the valid consideration set, and to calculate the second conditional selection probability under the exclusion state based on the set after removing the target novel alternative when the target novel alternative does not enter the valid consideration set. An observation likelihood construction module is used to construct a two-state observation likelihood based on the observation selection result. When the selection result is a novel alternative to the target, the likelihood contribution is the product of the potential consideration probability and the first conditional selection probability. When the selection result is a non-target alternative, the observation is decomposed into a first branch that considers the novel alternative to the target but selects a non-target alternative, and a second branch that does not consider the novel alternative to the target but selects a non-target alternative. The sum of the probabilities of the two branches is used as the observation likelihood contribution. The probability of the first branch is the potential consideration probability multiplied by the first conditional selection probability, and the probability of the second branch is the probability of not entering the effective consideration set multiplied by the second conditional selection probability. The prediction model building module is used to construct a prediction model by simulating and integrating the individual random effects and preference random coefficients on which the probabilities depend under the aforementioned different conditions. The prediction result generation module is used to generate prediction results related to the probability of selection for the target novel alternative using the prediction model.

[0018] The beneficial effects of the novel alternative selection prediction modeling method and system provided by this invention are as follows: By separating the potential consideration probability of a target novel alternative entering the effective consideration set from the conditional selection probability after entering the effective consideration set, the selection of the target novel alternative is no longer equivalent to its actual consideration. By distinguishing between the two branches of considered but not selected and unconsidered by using two-state observation likelihood, the parameter bias caused by the full consideration assumption directly interpreting low selection rate as low utility is reduced. A prediction model is constructed using panel data and simulated integration, outputting prediction results such as potential consideration probability, conditional selection probability, and comprehensive selection probability. It also supports the execution of at least one step in the model verification step and the strategy simulation step, thereby improving the accuracy and stability of the prediction of novel alternative selection. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the novel alternative selection prediction modeling method in an embodiment of the present invention; Figure 2 This is a schematic diagram of the modules of the novel alternative selection prediction modeling system according to an embodiment of the present invention; Figure 3 This is a detailed system structure block diagram in an embodiment of the present invention; Figure 4 This is a detailed flowchart of the method operation in an embodiment of the present invention. Detailed Implementation

[0020] The technical solution of the present invention will be further described below with reference to the accompanying drawings and specific embodiments. The following embodiments are used to illustrate the present invention and are not intended to limit the scope of protection of the present invention.

[0021] In this invention, "alternative" has the same meaning as "options" in intention surveys or choice experiments, referring to products, services, or modes of transportation that respondents can compare or potentially compare when choosing a task. Novel alternatives refer to products, services, or modes of transportation for which stable usage experience, evaluation standards, or behavioral habits have not yet been established; traditional alternatives refer to products, services, or modes of transportation for which respondents have relatively sufficient knowledge, usage experience, or stable evaluation standards. "Non-target alternatives" refer to alternatives other than the target novel alternative in this modeling; this modeling only sets a potential consideration state for the target novel alternative, while non-target alternatives participate in the calculation of the probability of conditional selection after excluding the target novel alternative.

[0022] Example 1 like Figure 1 and Figure 4 As shown, this embodiment provides a novel alternative selection prediction modeling method based on a potential consideration mechanism. This method is executed by a computer system, which acquires intention survey selection task data, standardizes and organizes the data into panel data, records the selection task presentation set for each selection task, determines a target novel alternative from the selection task presentation set, and uses alternatives other than the target novel alternative as non-target alternatives in this modeling.

[0023] Based on the panel data described above, the system calculates the potential consideration probability of the target novel alternative being included in the effective consideration set, using respondents-level variables, task-level variables, and individual random effects. The probability that the target novel alternative is not included in the effective consideration set is defined as the complement of the potential consideration probabilities. Here, the effective consideration set refers to the set of alternatives that respondents actually include in the comparison and evaluation process during the actual decision-making process, while the potential consideration state refers to the unobservable state of whether the target novel alternative is actually included in the effective consideration set by the respondents.

[0024] When the target novel alternative is included in the valid consideration set, the system calculates the first conditional selection probability based on the complete set containing the target novel alternative. When the target novel alternative is not included in the valid consideration set, the system calculates the second conditional selection probability under the exclusion state based on the set after removing the target novel alternative, and sets the second conditional selection probability of the target novel alternative under the exclusion state to 0. Based on the observed selection results, the system constructs a two-state observation likelihood: when the selection result is the target novel alternative, the product of the potential consideration probability and the first conditional selection probability is used as the likelihood contribution; when the selection result is a non-target alternative, the observation is decomposed into a first branch that considers the target novel alternative but selects a non-target alternative, and a second branch that does not consider the target novel alternative but selects a non-target alternative, and the sum of the probabilities of the two branches is used as the observation likelihood contribution.

[0025] Furthermore, the system constructs a prediction model by simulating and integrating the individual random effects and preference random coefficients upon which the probabilities depend under the aforementioned different conditions. This prediction model is then used to generate prediction results that correlate the probability of selecting the target novel alternative with the chosen probability. Thus, the selection of the target novel alternative is no longer equated with its actual consideration; the system can differentiate between "not included in the consideration set" and "having low conditional utility after being included in the consideration set."

[0026] Example 2 like Figure 2 and Figure 3As shown in the figure, this embodiment also provides a novel alternative selection prediction modeling system based on a latent consideration mechanism. The system includes a data processing module, a latent probability calculation module, a conditional selection probability calculation module, an observation likelihood construction module, a prediction model construction module, and a prediction result generation module.

[0027] The data processing module is used to acquire the intention survey selection task data, standardize and organize the data into panel data, record the selection task presentation set for each selection task, determine a target novel alternative from the selection task presentation set, and take the alternatives other than the target novel alternative as non-target alternatives in this modeling.

[0028] The potential probability calculation module is used to calculate the potential consideration probability of the target novel alternative predicted to enter the effective consideration set based on the respondent-level variables, task-level variables, and individual random effects, and to determine the probability that the target novel alternative predicted to not enter the effective consideration set as the complement of the potential consideration probability.

[0029] The conditional selection probability calculation module is used to calculate the first conditional selection probability based on the complete set containing the target novel alternative when the target novel alternative enters the valid consideration set, and to calculate the second conditional selection probability under the exclusion state based on the set after removing the target novel alternative when the target novel alternative does not enter the valid consideration set, and to determine the second conditional selection probability of the target novel alternative under the exclusion state as 0.

[0030] The observation likelihood construction module is used to construct two-state observation likelihoods based on the observation selection results. When the actual selection is the target novel alternative, it is automatically determined that it has been considered. When the actual selection is a non-target alternative, the observation is decomposed into two probability branches: "the target novel alternative is considered but not selected" and "the target novel alternative is not considered".

[0031] The prediction model building module is used to construct a prediction model by simulating and integrating the individual random effects and preference random coefficients on which the probabilities depend under the aforementioned different conditions; the prediction result generation module is used to generate prediction results related to the probability of selection for the target novel alternative using the prediction model.

[0032] In one implementation, the system recovers the posterior state probability for each observation after estimation. If the observation selection result is a novel alternative to the target, then... If the observed selection result is a non-target option, the system compares the relative contributions of the two likelihood branches, "considered but not selected" and "not considered," and calculates... Let this be the posterior probability that the observation belongs to "the new target alternative has been considered but not selected", and let This represents the posterior probability of "no new target alternatives have been considered".

[0033] When the model contains individual random effects and random preference coefficients, the system can first calculate the aforementioned posterior weights of the branches at each simulated sampling point, and then perform a normalized weighted average based on the panel likelihood contribution of the respondents corresponding to each simulated sampling point, outputting the respondent-task level results. , The individual average consideration probability and the probability of not considering are used to distinguish between "not considered" and "considered but not selected". The weight of each simulated sampling point is determined by the proportion of the panel likelihood contribution of the corresponding respondent at that sampling point to the total panel likelihood contribution of all simulated sampling points.

[0034] The aforementioned posterior consideration state probabilities are used to recover respondents' probabilities of considering novel alternatives, conditional choice preferences, and individual random effects under different tasks after estimation is completed, and to distinguish between "not considered" and "considered but not chosen".

[0035] The demand forecasting and strategy simulation module can predict the consideration rate, conditional selection rate, overall selection rate, and target user group after given the price, time, reliability, availability, or security information strategy of a new service.

[0036] Example 3 like Figure 4 As shown, this embodiment provides a specific operation method for predictive modeling of novel alternative selection, and the process of this method specifically includes the following steps.

[0037] S1. Collect data on the choice tasks in the intention survey. The system receives the alternative attributes and actual selection results of respondents in multiple choice tasks. The data on the choice tasks in the intention survey includes the respondent identifier, choice task number, choice task presentation set, actual selection result, alternative attribute, respondent attribute, psychological scale, historical experience, and task context variables.

[0038] S2, Unifying Respondent and Task Indexes. The system merges multiple task selections by the same respondent into panel selection data, forming a three-level index of respondent, selected task, and alternatives. It also records the set of selected tasks for each task and unifies the dimensions of alternative attribute properties. Alternative attribute properties may include cost, time, reliability, security, waiting time, and connectivity. Panel selection data refers to repeated selection observation data formed by the same respondent in multiple selection tasks.

[0039] S3, Determine the target novel alternative. The system determines a target novel alternative from the selected task presentation set, and treats all alternatives other than the target novel alternative as non-target alternatives in this modeling. The target novel alternative can be an alternative that has not yet formed a normalized usage behavior or that respondents lack direct experience with; non-target alternatives are alternatives other than the target novel alternative in this modeling, and may include alternatives with sufficient usage experience, stable evaluation criteria, and regular behavioral habits.

[0040] In one implementation, the identification of candidate novel alternatives adopts an executable decision rule: the system obtains at least one of the following information from the selection task presentation set: historical usage rate, commercial operation time, respondents' self-reported familiarity, whether there is a direct experience record, whether there is a stable evaluation standard, and whether a regular usage habit has been formed for each candidate alternative, and generates a candidate novel alternative tag.

[0041] In one implementation, the system first identifies candidate novel alternatives from the set of task presentations based on historical usage rates, commercial operation time, respondents' self-reported familiarity, whether there are direct experience records, whether there are stable evaluation standards, whether regular usage habits have been formed, expert rules, or platform rules. Specifically, if the alternative... If at least one of the following conditions is met, it is identified as a candidate novel alternative: historical usage rate The following criteria were considered: 1. The total familiarity score reported by respondents was lower than the preset familiarity threshold; 2. The number of respondents indicating they had never directly experienced the alternative reached the preset quantity threshold or its proportion reached the preset percentage threshold; 3. Whether there were records of stable evaluation standards (indicating no stable evaluation standards); 4. Whether there were records of established regular usage habits (indicating no established regular usage habits); 5. Commercial operation time. Alternatively, expert rules or platform rules may determine that it does not belong to the category of normalized behavior. When there is only one candidate novel alternative, that candidate novel alternative is determined as the target novel alternative for this modeling. When there are multiple candidate novel alternatives, the system selects one from the multiple candidate novel alternatives as the target novel alternative for this modeling based on at least one of the following: preset target identifier, business focus object, priority rule, or user-specified rule. For other candidate novel alternatives that are not selected as the target novel alternative for this modeling, the system can repeat this method as the target novel alternative in other modeling rounds. Non-target alternatives may include alternatives with sufficient usage experience, stable evaluation criteria, and regular behavioral habits. and These are the historical usage rate threshold and the operating time threshold, respectively. Indicates alternative item Historical usage rate Indicates alternative item The commercial operation period. The summary of respondents' self-reported familiarity and direct experience records, as well as the threshold determination, will be carried out in accordance with the method described in the next paragraph.

[0042] For respondents' self-reported familiarity and whether they have directly experienced the alternative, the system summarizes the records of multiple respondents by candidate alternative, forming a summary value of respondents' self-reported familiarity at the alternative level, as well as the number of respondents or their proportion indicating that they have not directly experienced the candidate alternative. When the summary value of respondents' self-reported familiarity is lower than a preset familiarity threshold, or when the number of respondents reaches a preset quantity threshold or their proportion reaches a preset proportion threshold, the corresponding familiarity condition or direct experience condition is determined to be met. The aforementioned information at the respondent level is not used to determine different target novel alternatives for different respondents.

[0043] S4. Constructing Consideration Stage Variables. The system sets whether the target novel alternative enters the valid consideration set as a potential state; it constructs consideration stage variables based on respondent-level variables and task-level variables, and uses individual random effects as the random components of the consideration stage variables. Respondent-level variables are input variables characterizing the characteristics of the respondents, and may include familiarity, trust in technology, risk attitude, income, occupation, existing experience, psychological scales, and knowledge level of new services; task-level variables are input variables characterizing the conditions for choosing a task, and may include travel distance, travel purpose, time pressure, price context, safety information, reliability, service availability, and suitability for the current scenario.

[0044] S5, Calculate the potential consideration probability of the target novel alternative. The system maps the combined result of the consideration stage variables and individual random effects to a potential consideration probability ranging from 0 to 1 using a probability link function. The probability link function is either a probit link function or a logit link function. The potential consideration probability refers to the probability that the target novel alternative is predicted to enter the respondent's valid consideration set.

[0045] In the probit implementation, the system calculates... ,in In the implementation of logit, it can be Replace with .in, , and To take stage parameters into account, For respondents, For task-level variables, For individual random effects, This represents the variance parameter of the individual random effect.

[0046] S6, Calculate the first conditional selection probability. When the target novel alternative enters the effective consideration set, the system calculates the first conditional selection probability based on the complete set containing the target novel alternative. Specifically, the system constructs the utility value of each alternative based on the alternative constant term, the alternative attribute, and the individual preference coefficient containing a random component, and calculates the conditional selection probability using a mixed Logit model or a random utility model. The first conditional selection probability refers to the conditional selection probability calculated in the complete set containing the target novel alternative when it enters the effective consideration set.

[0047] S7, Calculate the second conditional selection probability. When the target novel alternative is not included in the valid consideration set, the system calculates the second conditional selection probability under the exclusion state based on the set after removing the target novel alternative. This second conditional selection probability is used to represent the probability that the respondent selects the corresponding non-target alternative from the set of non-target alternatives when the target novel alternative is not included in the actual comparison range; when the corresponding alternative is the target novel alternative, its second conditional selection probability under the exclusion state is determined to be 0.

[0048] S8, forming a two-state observation likelihood at the observation level. If the observation selection result is a target novel alternative, then the target novel alternative consideration state corresponding to the observation is determined as considered, and the product of the potential consideration probability and the first conditional selection probability is used as the likelihood contribution; if the observation selection result is a non-target alternative, then the observation is decomposed into a first branch and a second branch, where the first branch is where the target novel alternative has been considered but a non-target alternative is selected, and its probability is the potential consideration probability multiplied by the first conditional selection probability; the second branch is where a non-target alternative is selected after not considering the target novel alternative, and its probability is the probability of not entering the effective consideration set multiplied by the second conditional selection probability, and the sum of the probabilities of the two branches is used as the observation likelihood contribution.

[0049] The constraint of the observation results on the potential state is: if Then the target new alternative must have entered the effective consideration set, that is... Observational likelihood is ;like If so, two branches are retained: those that have been considered but are not the target alternatives, and those that have not considered the target novel alternatives but are selected from the set of non-target alternatives.

[0050] The above rules can be uniformly written as The indicator ensures that the branch is automatically zero when selecting the target novel alternative; at the same time, it is agreed that the second condition selection probability of the target novel alternative in the exclusion state is 0.

[0051] Two-state observation likelihood refers to the probability expression constructed after selecting the same observation result and distinguishing between two potential states: the target novel alternative has been considered and the target novel alternative has not been considered.

[0052] S9 generates multiple simulated sampling points for individual random effects and preference random coefficients. At each simulated sampling point, the system calculates the potential consideration probability, the first conditional choice probability, the second conditional choice probability, and the observed likelihood contribution, and forms a panel likelihood based on the observed likelihood contributions from multiple choice tasks for the same respondent.

[0053] Simulated integration can employ Halton sequences, Sobol sequences, Latin hypercube sampling, or pseudo-random normal sampling. The system generates [data / information] for each respondent. Group sampling When considering random effects, the distribution is usually assumed to be normal. The random coefficients of the preference can be set to normal, log-normal, triangular, or sign-constrained transformation distribution according to the meaning of the attribute.

[0054] The simulation panel seems to be clickable. Calculation. The number of samples can be set according to the sample size and the dimension of the random item, for example... to Furthermore, the number of sampling points can be increased in the final estimation stage to test stability. Here, R represents the number of simulated sampling points used in the simulated integration, and θ represents the parameter to be estimated in the prediction model. This indicates the number of tasks selected by respondent i.

[0055] S10, Construct the prediction model. The system generates a simulated panel log-likelihood based on the panel likelihoods corresponding to multiple simulated sampling points, and constructs a prediction model by maximizing the simulated panel log-likelihood.

[0056] In engineering implementation, the system maximizes the simulated log-likelihood. Parameter updates can employ BFGS, L-BFGS, Newton's method, quasi-Newton's method, or automatic differential gradient optimization; probability and likelihood calculations use the log-sum-exp technique to avoid numerical underflow.

[0057] Convergence conditions can be set as follows: the log-likelihood increment is below a threshold, the gradient norm is below a threshold, or the change in parameters during continuous iterations is below a threshold. For example... to The system uses the standard deviation parameter. The system performs a positive transformation and calculates the corresponding variance parameter from the standard deviation parameter. The system outputs the parameter estimate, standard error, covariance matrix, random effects distribution parameter, individual posterior probability, model diagnostic index, and prediction result table.

[0058] Simulated maximum likelihood estimation is a method for estimating model parameters by sampling approximate integration and maximizing the likelihood function when the model contains random effects or random preference coefficients and the integral cannot be directly calculated analytically.

[0059] After the prediction model is built, the system can execute the posterior consideration state probability recovery step. The system recovers the posterior consideration state probability corresponding to the observation based on the two-state observation likelihood: when the observation selects the target novel alternative, the target novel alternative consideration state corresponding to the observation is determined as considered; when the observation selects a non-target alternative, based on the relative contribution of the probability of the first branch and the probability of the second branch in the sum of the probabilities of the two branches, the posterior probability of the observation belonging to the target novel alternative that has been considered but not selected, and the posterior probability of belonging to the target novel alternative that has not been considered, are determined respectively.

[0060] In one implementation, the system recovers the posterior state probability for each observation after estimation. If the observation selection result is a novel alternative to the target, then... If the observed selection result is a non-target option, the system compares the relative contributions of the two likelihood branches, "considered but not selected" and "not considered," and calculates... Let this be the posterior probability that the observation belongs to "the new target alternative has been considered but not selected", and let This represents the posterior probability of "no new target alternatives have been considered".

[0061] When the model contains individual random effects and random preference coefficients, the system can first calculate the aforementioned posterior weights of the branches at each simulated sampling point, and then perform a normalized weighted average based on the panel likelihood contribution of the respondents corresponding to each simulated sampling point, outputting the respondent-task level results. , The individual average consideration probability and the probability of not considering are used to distinguish between "not considered" and "considered but not selected". The weight of each simulated sampling point is determined by the proportion of the panel likelihood contribution of the corresponding respondent at that sampling point to the total panel likelihood contribution of all simulated sampling points.

[0062] In this embodiment, the system performs a model verification step and a strategy simulation step; in other implementations, the system performs at least one of these steps. S11, Perform model verification. The system compares the predicted model with a fully considered hybrid Logit model, a task-invariant consideration model, a conservative exclusion constraint model, or a relaxed exclusion constraint model, and generates a model diagnostic result.

[0063] The full-consideration deviation diagnostic output includes: the target novel alternative constant term correction magnitude. The slope correction range of the target new alternative exclusive attribute Or correct the ratio , where LC represents the potential consideration model and FC represents the full consideration benchmark model.

[0064] Model comparison metrics include simulated log-likelihood, AIC, BIC, retained sample prediction accuracy, and likelihood ratio statistic. In the eVTOL application example, after separating the unconsidered states, the target novel alternative constant term is derived from the fully considered model. Upgraded to the potential consideration model Furthermore, considering the entire baseline, the likelihood ratio test significantly rejected the hypothesis. This can serve as an example of the system output format.

[0065] Model diagnostic results may also include at least one of the following: simulated log-likelihood, AIC, BIC, retained sample prediction accuracy, likelihood ratio statistic, correction magnitude of the constant term of the novel alternative, and correction magnitude of the slope of the specific attribute of the novel alternative.

[0066] S12, output the prediction result and perform strategy simulation. The prediction result associated with the selection probability includes at least one of the following: potential consideration probability, first conditional selection probability, second conditional selection probability, and comprehensive selection probability formed based on the potential consideration probability, the complement of the potential consideration probability, the first conditional selection probability, and the second conditional selection probability.

[0067] Inputs to the strategy simulation can include new service pricing adjustments, wait time improvements, reliability enhancements, safety alerts, availability expansion, familiarization promotions, trial ride experiences, subsidy rules, or changes in service coverage. The system recalculates the probability of consideration for each of these changes. Complete consideration of selection rate Exclusion status selection rate And overall selection rate.

[0068] For the target novel alternatives, the overall selection rate was [missing information]. For non-target alternatives The overall selection rate was The output can be broken down into the contribution of increased consideration rate, the contribution of increased condition selection rate, and changes in total demand, which can be used to evaluate different promotion or service design strategies.

[0069] Given the strategy adjustment parameters for the target novel alternative, the system can recalculate the potential consideration probability, the first condition selection probability, the second condition selection probability, and the comprehensive selection probability, generating strategy simulation results associated with the selection probabilities. Strategy adjustment parameters may include price adjustments for the new service, improvements to waiting time, enhanced reliability, safety reminders, expanded availability, promotion of familiarity, test drive experiences, changes in subsidy rules, or changes to service coverage.

[0070] Example 4 This embodiment is a mathematical model description of the latent consideration variables, effective consideration set, conditional choice probability, two-state observation likelihood, and simulated panel log likelihood in Embodiment 3.

[0071] make The respondents indicated that... Respondents The task selection This indicates the selection of a task set to display. This indicates the target novel alternative. This indicates the result of the observation selection.

[0072] This invention sets potential consideration indicator variables. : in , Respondents In the mission China has already included the target new alternatives Included in the effective consideration set; This indicates that although the novel alternative to the target was presented in the task, it was not included in the actual comparison. This state is a latent variable and can change with both the respondent and the task.

[0073]

[0074] Effective consideration of set transitions based on potential states: .when At that time, the probability of a complete selection is given in the set of task presentations. Above calculation; when At that time, the system first eliminates the target new alternative. And then The probability of selecting the exclusion state for non-target alternatives is calculated.

[0075]

[0076] Considering the stage probability as:

[0077] in, Respondent-level variables that characterize the traits of respondents. To characterize the task-level variables for selecting task conditions. For individual random effects, It is the standard normal distribution function.

[0078] Given a valid set of considerations At that time, alternatives Its utility is:

[0079] in, For the constant term in the options, For alternative attributes, For individual preference coefficients that may contain random components, This is a Type I extreme error.

[0080] When the target novel alternative is included in the consideration set, the probability of selecting the complete condition is:

[0081] When the target novel alternative is excluded, the probability of conditional selection in the excluded state is:

[0082] At the same time, it was agreed that when New alternatives for the target When the condition is excluded, the probability of choosing the condition is 0.

[0083] The likelihood contribution at the observation level is:

[0084] when When the indicator function is zero, the likelihood degenerates to This indicates that selecting a new alternative to the target necessarily requires its consideration. When At that time, the likelihood simultaneously includes two branches: those considered but not selected and those not considered.

[0085] Given a random item, the respondents The condition panel likelihood is:

[0086] Integrating over the random term yields the unconditional likelihood:

[0087] use For each simulated sampling point, the simulated panel likelihood is:

[0088] The simulated log-likelihood is:

[0089] When system settings At this point, the observed likelihood degenerates into a standard full-consideration mixed Logit. Therefore, the model of this invention has a nested relationship with the full-consideration baseline model, and the significance of the underlying consideration mechanism can be determined through the likelihood ratio test. Where θ represents the model parameters to be estimated, R represents the number of simulated sampling points, and I represents the total number of respondents. This indicates the number of tasks selected by respondent i; the full consideration assumption refers to the assumption in traditional discrete choice models that all presented alternatives are actually considered and evaluated by the respondent.

[0090] Example 5 This embodiment explains the technical basis for parameter deviation under the full consideration assumption, and explains why it is necessary to set the potential consideration probability and the conditional selection probability under the exclusion state.

[0091] In the process of generating real data, if the probability of considering the target novel alternative is... The probability of conditional selection is Then the probability of observation selection is:

[0092] If the researcher incorrectly applies the assumption of full consideration and estimates... Then the pseudo-true parameters need to match the average observation share of the target novel alternatives:

[0093] because exist If the share is lower than the conditional selection share, the full-consideration model will absorb this difference by reducing the constant term of the target novel alternative. The constant term offset under the representative share is:

[0094] For attributes that change only in the target new alternative Mistakes in fully considering the conditions can also lead to local attenuation of the slope:

[0095] This result indicates that when some respondents do not consider the novel alternative to the target, their non-selection does not fully reflect low attribute utility. If the consideration phase is not modeled, the attractiveness and specific attribute sensitivity of the novel alternative to the target will be systematically underestimated. This represents the overall average conditional choice share of the target novel alternative n. This represents the overall average probability. This indicates an approximate shift in the representative share of the target novel alternative constant term relative to the traditional reference term. This represents the true slope parameter of attribute m. This indicates the pseudo-true slope parameter of attribute m under full consideration of misdefined conditions. This represents the slope decay factor under representative conditions. The above constant term offset and slope decay conclusions are based on the representative share approximation and local linear approximation conditions, and are affected by the higher-order remainder terms.

[0096] Based on the above bias analysis, the novel alternative selection prediction modeling method and system provided in this embodiment can distinguish between "not considered" and "not selected after consideration", avoiding interpreting all observations of not selecting the target novel alternative as having low attractiveness of the target novel alternative, thereby improving the accuracy of novel alternative preference estimation; at the same time, it reduces the problems of underestimation of the novel alternative constant term and weakening of the attribute coefficient.

[0097] Example 6 This embodiment describes the identification constraints, limiting models, and related technical parameters for potential consideration probabilities and conditional selection probabilities.

[0098] This invention identifies potential consideration probabilities and conditional selection probabilities through four types of information.

[0099] (1) Structural differences in the results of observation selection. When At that time, the state was revealed as ;when At that time, the observed likelihood is a mixture of two branches.

[0100] (2) Panel repeated selection of tasks. The selection sequence of the same respondent across tasks can distinguish between a state of continuous non-consideration and a state of consideration without selection. Respondents in the state of continuous non-consideration are less responsive to changes in the attributes of new alternatives; respondents in the state of consideration without selection may still show some sensitivity to attributes.

[0101] (3) Exclusion of Restrictions. Some respondents' level variables are included in the consideration equation but not in the conditional utility equation, such as cognitive familiarity, technological trust, and understanding of new services, etc., which are used to make... The changes occur without directly altering the conditional choice utility.

[0102] (4) Task-level variables. Variables such as distance, purpose, time pressure, availability, safety tips, and price context are included in the consideration equation, which makes the probability of consideration by the same respondent vary in different tasks.

[0103] For ease of verification, the present invention can also establish the following constraint models: relaxed exclusion constraint model, conservative exclusion constraint model, fully considered hybrid Logit model, and task-invariant consideration model. Fully considered model settings. Considering model settings while keeping the task unchanged By comparing the above-mentioned constrained model with the model of this invention, the likelihood ratio test results and information criteria can be output.

[0104] In this embodiment, the target novel alternative label indicates a target novel alternative presented in the selected task but which may not yet have formed a normalized usage behavior; the potential consideration variable is a binary latent variable used to indicate whether the target novel alternative is included in the effective consideration set; the consideration probability ranges from 0 to 1 and is jointly determined by the respondent-level variable, the task-level variable, and the random effect.

[0105] Respondent-level variables may include familiarity, risk attitude, trust in technology, income, occupation, and existing experience; task-level variables may include travel distance, purpose, time pressure, price scenario, reliability, safety information, and service availability; individual consideration of random effects is used to characterize the heterogeneity of unobserved considerations and can usually be set to a normal distribution.

[0106] Preference coefficients can include fixed and random coefficients to represent the marginal utility of attributes such as cost, time, waiting, reliability, and security; the number of samplings is the number of simulated sampling points used for simulated integration, which can be set according to computing resources and accuracy requirements, for example, from 200 to 2000; model diagnostic indicators can include simulated log-likelihood, AIC, BIC, likelihood ratio statistic, consideration rate distribution, ASC correction magnitude, and slope correction magnitude of the target novel alternative specific attribute.

[0107] For emerging mobility solutions such as eVTOL, autonomous driving, and MaaS, when respondents have not yet considered them, the system can distinguish between the unconsidered state and the considered-but-not-chosen state by using the potential consideration probability and the posterior consideration state probability. If respondents have not considered emerging alternatives such as eVTOL, changes in the price, waiting time, safety, and other attributes of these alternatives will not elicit a choice response from the respondents. This embodiment reduces the possibility of misjudging "no choice response" as "attribute insensitivity" by using the two-state observation likelihood and the second conditional choice probability under the excluded state.

[0108] The prediction model in this embodiment can output the potential consideration probability, the first conditional selection probability, the second conditional selection probability, and the comprehensive selection probability, respectively, and can execute at least one of the model verification step and the strategy simulation step to generate the model diagnosis result or strategy simulation result corresponding to the executed step.

[0109] Example 7 This embodiment provides an electronic device, including a processor, a memory, a communication interface, and a communication bus. The processor, the memory, and the communication interface are connected through the communication bus. The memory stores a computer program, as well as intention survey selection task data, model parameters, and prediction results required to execute the computer program. The processor executes the computer program to implement the novel alternative selection prediction modeling method based on the potential consideration mechanism described in any of the above method embodiments.

[0110] In one implementation, the computer program includes data processing program code, latent probability calculation program code, conditional selection probability calculation program code, observation likelihood construction program code, prediction model construction program code, and prediction result generation program code. When executed by the processor, each program code respectively implements the functions of the data processing module, latent probability calculation module, conditional selection probability calculation module, observation likelihood construction module, prediction model construction module, and prediction result generation module. The electronic device may be a server, workstation, desktop computer, portable computer, cloud computing node, or a computing device composed of multiple computing nodes.

[0111] This embodiment also provides a computer-readable storage medium on which a computer program is stored; when the computer program is executed by a processor, it implements the steps of the method described in any of the above method embodiments. The computer-readable storage medium may be a non-transitory computer-readable storage medium, including a read-only memory, random access memory, solid-state memory, magnetic disk, optical disk, or other storage medium capable of storing computer program code.

[0112] This embodiment also provides a computer program, which includes computer program code; when the computer program code is executed by a processor, it causes an electronic device to implement the steps of the method described in any of the above method embodiments. The computer program may be provided separately, or it may be stored as part of a computer program product in a computer-readable storage medium or transmitted via a network.

[0113] The above are merely exemplary embodiments of the present invention. Those skilled in the art can make various modifications, variations, or equivalent substitutions without departing from the spirit and scope defined by the claims. All modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A novel alternative selection prediction modeling method based on a latent consideration mechanism, characterized in that, include: Obtain the task selection data from the intention survey, standardize and organize the data into panel data, record the task selection presentation set for each task, determine a target novel alternative from the task selection presentation set, and take the alternatives other than the target novel alternative as the non-target alternatives in this modeling. Based on respondents' level variables, task level variables, and individual random effects, the potential consideration probability of the target novel alternative being predicted to enter the effective consideration set is calculated, and the probability of the target novel alternative not being predicted to enter the effective consideration set is the complement of the potential consideration probability. When the target novel alternative is included in the valid consideration set, the probability of the first conditional selection is calculated based on the complete set containing the target novel alternative. When the target novel alternative is not included in the effective consideration set, the second conditional selection probability under the exclusion state is calculated based on the set after removing the target novel alternative; Based on the observation selection results, a two-state observation likelihood is constructed: when the selection result is a new target alternative, the product of the potential consideration probability and the first conditional selection probability is used as the likelihood contribution; when the selection result is a non-target alternative, the observation is decomposed into a first branch that has considered the new target alternative but selected a non-target alternative, and the probability of the first branch is the potential consideration probability multiplied by the first conditional selection probability. And a second branch that selects a non-target alternative without considering the target novel alternative, the probability of the second branch being the probability of not entering the effective consideration set multiplied by the probability of the second conditional selection, and the sum of the probabilities of the two branches as the observation likelihood contribution; By simulating and integrating the individual random effects and preference random coefficients on which the probabilities depend under the aforementioned different conditions, and constructing a prediction model, the prediction model is used to generate prediction results related to the probability of selection for the target novel alternative.

2. The novel alternative selection prediction modeling method based on a latent consideration mechanism according to claim 1, characterized in that, The data on the willingness survey for selecting tasks includes respondent identifiers, task numbers, task presentation sets, actual selection results, alternative attributes, respondent attributes, psychological scales, historical experience, and task context variables. The data is standardized and organized into panel data, including: merging multiple selection tasks of the same respondent into panel selection data, forming a three-level index of respondents, selection tasks, and alternatives, and unifying the dimensions of the alternative attributes.

3. The novel alternative selection prediction modeling method based on a potential consideration mechanism according to claim 1, characterized in that, The steps for determining the target novel alternative include: acquiring at least one of the following information from the selected task presentation set: historical usage rate, commercial operation time, respondents' self-reported familiarity, whether they have directly experienced the alternative, whether they have stable evaluation standards, and whether they have formed a regular usage habit; and summarizing the respondents' self-reported familiarity and whether they have directly experienced the alternative for each candidate alternative; when a candidate alternative meets at least one of the following conditions, it is identified as a candidate novel alternative: historical usage rate is lower than a preset historical usage rate threshold; commercial operation time is shorter than a preset operation time threshold; the sum of respondents' self-reported familiarity is lower than a preset familiarity threshold; the number of respondents who have not directly experienced the candidate alternative reaches a preset number threshold or... The proportion reaches a preset threshold; whether there is a stable evaluation standard indicates that there is no stable evaluation standard; whether there is a record of regular usage habits indicates that regular usage habits have not been formed; or expert rules or platform rules determine that the candidate alternative does not belong to the category of normalized behavior; when there is only one candidate novel alternative, the candidate novel alternative is determined as the target novel alternative for this modeling; when there are multiple candidate novel alternatives, one is selected as the target novel alternative for this modeling based on at least one of the preset target identifier, business focus object, priority rules or user-specified rules; alternatives in the selected task presentation set other than the target novel alternative are used as non-target alternatives for this modeling.

4. The novel alternative selection prediction modeling method based on a potential consideration mechanism according to claim 1, characterized in that, The calculation steps for the potential consideration probability include: setting whether the target novel alternative enters the effective consideration set as a potential state; constructing a consideration stage variable based on respondent-level variables and task-level variables, and using individual random effects as random components of the consideration stage variable; and mapping the combination result of the consideration stage variable and the individual random effects to a potential consideration probability with a value range of 0 to 1 through a probability link function. The respondent-level variables are input variables that characterize the characteristics of the respondents, the task-level variables are input variables that characterize the conditions for selecting the task, and the probability link function is one of the probit link function and the logit link function.

5. The novel alternative selection prediction modeling method based on a latent consideration mechanism according to claim 4, characterized in that, The calculation steps for the conditional selection probability include: constructing the utility value of each alternative based on the constant term of the alternative, the attribute of the alternative, and the individual preference coefficient containing a random component, and calculating the conditional selection probability through a mixed Logit model or a random utility model; the conditional selection probability is used for the complete set when the target novel alternative enters the effective consideration set, as the first conditional selection probability; and is used for the exclusion set when the target novel alternative does not enter the effective consideration set, as the second conditional selection probability.

6. The novel alternative selection prediction modeling method based on a latent consideration mechanism according to claim 1, characterized in that, It also includes a posterior consideration state probability recovery step: after constructing the prediction model, the posterior consideration state probability corresponding to the observation is recovered based on the likelihood of the two-state observation; when the observation selection result is the target novel alternative, the consideration state of the target novel alternative corresponding to the observation is determined as considered; When the observation selection result is a non-target alternative, the posterior probability of the observation belonging to the target novel alternative that has been considered but not selected, and the posterior probability of the observation belonging to the target novel alternative that has not been considered, are determined according to the relative contributions of the probabilities of the first branch and the second branch to the sum of the probabilities of the two branches.

7. The novel alternative selection prediction modeling method based on a latent consideration mechanism according to claim 1, characterized in that, The steps for constructing the prediction model include: generating multiple simulated sampling points for individual random effects and preference random coefficients; calculating the potential consideration probability, the first conditional choice probability, the second conditional choice probability, and the observed likelihood contribution for each simulated sampling point; forming a panel likelihood based on the observed likelihood contributions of multiple selection tasks for the same respondent, and forming a simulated panel log-likelihood based on the panel likelihoods corresponding to multiple simulated sampling points; and constructing the prediction model by maximizing the simulated panel log-likelihood.

8. The novel alternative selection prediction modeling method based on a latent consideration mechanism according to claim 7, characterized in that, The multiple simulated sampling points are generated using Halton sequences, Sobol sequences, Latin hypercube sampling, or pseudo-random normal sampling; the individual random effects are set to a normal distribution, and the preference random coefficients are set to a normal distribution, log-normal distribution, triangular distribution, or a sign-constrained transformation distribution; the parameter updates of the simulated panel log-likelihood are implemented using BFGS, L-BFGS, Newton's method, quasi-Newton's method, or automatic differential gradient optimization.

9. The novel alternative selection prediction modeling method based on a latent consideration mechanism according to claim 1, characterized in that, The prediction results associated with the probability of selection include: the potential consideration probability, the first conditional selection probability, the second conditional selection probability, and at least one of the following: the potential consideration probability, the complement of the potential consideration probability, the first conditional selection probability, and the second conditional selection probability. The method further includes at least one of a model verification step and a strategy simulation step; the model verification step includes: comparing the prediction model with a fully considered hybrid Logit model, a task-invariant consideration model, a conservative exclusion constraint model, or a relaxed exclusion constraint model, and generating a model diagnostic result; the strategy simulation step includes: after considering the strategy adjustment parameters of a novel alternative to the target, recalculating the potential consideration probability, the first conditional selection probability, the second conditional selection probability, and the comprehensive selection probability, and generating a strategy simulation result associated with the selection probability; The model diagnostic results include at least one of the following: simulated log-likelihood, AIC, BIC, retained sample prediction accuracy, likelihood ratio statistic, correction magnitude of the constant term of the novel alternative, and correction magnitude of the slope of the novel alternative specific attribute.

10. A novel alternative selection prediction modeling system based on a latent consideration mechanism, characterized in that, include: The data processing module is used to acquire the intention survey selection task data, standardize and organize the data into panel data, record the selection task presentation set for each selection task, determine a target new alternative from the selection task presentation set, and take the alternatives other than the target new alternative as non-target alternatives in this modeling. The potential probability calculation module is used to calculate the potential consideration probability of the target novel alternative predicted to enter the effective consideration set based on the respondent-level variables, task-level variables and individual random effects, and to determine the probability that the target novel alternative predicted not to enter the effective consideration set as the complement of the potential consideration probability. The conditional selection probability calculation module is used to calculate the first conditional selection probability based on the complete set containing the target novel alternative when the target novel alternative enters the valid consideration set, and to calculate the second conditional selection probability under the exclusion state based on the set after removing the target novel alternative when the target novel alternative does not enter the valid consideration set. An observation likelihood construction module is used to construct a two-state observation likelihood based on the observation selection result. When the selection result is a novel alternative to the target, the likelihood contribution is the product of the potential consideration probability and the first conditional selection probability. When the selection result is a non-target alternative, the observation is decomposed into a first branch that considers the novel alternative to the target but selects a non-target alternative, and a second branch that does not consider the novel alternative to the target but selects a non-target alternative. The sum of the probabilities of the two branches is used as the observation likelihood contribution. The probability of the first branch is the potential consideration probability multiplied by the first conditional selection probability, and the probability of the second branch is the probability of not entering the effective consideration set multiplied by the second conditional selection probability. The prediction model building module is used to construct a prediction model by simulating and integrating the individual random effects and preference random coefficients on which the probabilities depend under the aforementioned different conditions. The prediction result generation module is used to generate prediction results related to the probability of selection for the target novel alternative using the prediction model.

Images