An insurance recommendation method and system based on timing similarity
Patent Information
- Application Number
- CN202611134051.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-29
- Publication Date
- 2026-10-20
AI Technical Summary
[0005]本发明的目的是提供一种基于时序相似性的保险推荐方法及系统，解决了现有技术中保险推荐系统在时序行为建模不足和冷启动/稀疏交互场景下难以准确捕捉用户动态意图与真实风险偏好的问题
[0030]本发明的有益效果为：突破传统方案仅使用静态统计数据的局限，通过非均匀重采样构建心理行为时间轴，强化用户高活跃时段的行为特征；结合时序分解、局部异常因子算法提取趋势项、周期项、兴趣突变点多维时序特征，完整还原用户保险需求的演化轨迹，可精准捕捉用户实时意图与需求变化，让推荐更具前瞻性。
Smart Images
Figure FT_1 
Figure FT_2 
Figure FT_3
Abstract
Description
Technical Field
[0001] This invention relates to the fields of financial technology and artificial intelligence, and in particular to an insurance recommendation method and system based on temporal similarity. Background Technology
[0002] With the increasing diversification of the insurance market and the personalization of user needs, insurance recommendation systems have become a key tool for improving user experience and promoting business conversion. Currently, the mainstream recommendation methods mainly fall into three categories: collaborative filtering based on user behavior similarity, content filtering that matches user preferences with product attributes, and machine learning methods that utilize statistical features for modeling. These methods each have their own focus, collectively driving the continuous evolution and application of insurance recommendation technology.
[0003] Currently, existing insurance recommendation systems still face several challenges. First, they fail to adequately utilize time-series data. Commonly used static statistics (such as click counts over the past 30 days) cannot capture deep dynamic patterns in user behavior sequences, such as periodic fluctuations, abrupt changes, and evolving interests, resulting in a lack of sensitivity and foresight regarding changes in user intent. Second, they suffer from cold-start and sparse interaction problems. New or low-frequency users lack sufficient historical data, making it difficult for traditional collaborative filtering to generate effective user vectors. Relying on basic demographic information, however, is too dimensional and coarse-grained to accurately reflect users' true insurance needs and risk preferences. Although some research has introduced a time dimension, these problems have not yet been effectively resolved.
[0004] Therefore, there is an urgent need for a new insurance recommendation technology that can deeply mine temporal behavior patterns and address the challenges of cold starts. Summary of the Invention
[0005] The purpose of this invention is to provide an insurance recommendation method and system based on temporal similarity, which solves the problems of insufficient temporal behavior modeling and difficulty in accurately capturing users' dynamic intentions and true risk preferences in cold start / sparse interaction scenarios in existing insurance recommendation systems.
[0006] To achieve the above-mentioned objectives, the technical solution adopted by the present invention is as follows: an insurance recommendation method based on temporal similarity, comprising the following steps: obtaining temporal data of the insurance interaction history of the target user and the candidate user within a preset time window, the data including various interaction events and corresponding timestamps, product categories, and interaction intensity;
[0007] The data is standardized and converted into psychological and behavioral timeline data through non-uniform resampling; time-series feature vectors containing trend terms, periodic terms, and abrupt change points are extracted.
[0008] The time-decay weighted DTW algorithm is used to calculate the temporal comprehensive similarity between users and to filter out highly similar users.
[0009] Based on the insurance preferences of similar users, a personalized insurance recommendation list is generated for the target user.
[0010] As an improvement, the rule for non-uniform resampling is as follows:
[0011] The original timeline is converted into a psychological behavior timeline, so that the interval between adjacent time points in the psychological behavior timeline is negatively correlated with the intensity of users' real-time behavior, thereby enhancing the accuracy of feature representation during periods of high user activity.
[0012] As an improvement, the time-series feature vector extraction step includes:
[0013] By decomposing the behavioral sequence into trend and periodic components, and relying on the local anomaly factor algorithm to detect the characteristics of user interest mutation points, the processed behavioral sequence is segmented by a sliding window, and combined with temporal embedding encoding and attention mechanism to generate a temporal feature vector representing the evolution trajectory of user interest.
[0014] As an improvement, the calculation method for the temporal comprehensive similarity is as follows:
[0015] Trend similarity, periodic similarity, and mutation similarity are calculated separately. A weighting factor is set using a time decay function to weight recent behaviors. The three types of similarity are then weighted and fused to obtain a time-series comprehensive similarity, which improves the sensitivity of identifying the user's current true intention.
[0016] As an improvement, a candidate user pre-screening step is included before similarity calculation:
[0017] Clustering of all users based on user static attributes and coarse-grained behavioral statistical features preliminarily filters candidate user groups, narrows the similarity calculation range of the DTW algorithm, reduces computational costs, and improves recommendation efficiency.
[0018] As an improvement, for cold start scenarios with sparse new user or behavioral data, the following processing steps are included:
[0019] By integrating auxiliary time-series signals such as user exploratory behavior during the initial registration phase, questionnaire completion pace, and page dwell time trends, and combining them with user basic attributes and risk preference tags, a pre-built group behavior pattern library is matched, and corresponding typical behavior patterns are used as user substitute time-series features to complete similarity calculation.
[0020] As an improvement, the construction steps of the group behavior pattern library are as follows:
[0021] Based on basic attributes, massive amounts of complete historical users are grouped, and the temporal behavioral characteristics of each user group are clustered. The cluster centers containing common trend items, periodic items, and distribution of mutation points of the group are used as the typical behavioral patterns of the corresponding user groups and stored in the group behavior pattern library.
[0022] An insurance recommendation system based on temporal similarity includes: a data acquisition module for acquiring temporal data of the historical insurance interaction behavior of target users and candidate users;
[0023] The resampling preprocessing module is used to complete data standardization and psychological behavior timeline transformation;
[0024] The temporal feature extraction module is used to generate user temporal feature vectors;
[0025] The similarity calculation module is used to calculate the overall temporal similarity based on the weighted DTW algorithm;
[0026] The similar user filtering module is used to filter sets of highly similar users;
[0027] The recommendation generation module is used to output a personalized insurance recommendation list.
[0028] As an improvement, a cold start processing module is also included: when user data is sparse, it integrates auxiliary time-series signals to construct an initial time-series feature vector, matches the group behavior pattern library to obtain alternative time-series features, and provides effective data support for similarity calculation.
[0029] As an improvement, the similarity calculation module integrates a GPU-accelerated parallel DTW computing unit: used to realize parallel temporal similarity retrieval for a large user group, adapting to massive user scenarios and significantly improving the efficiency of similarity calculation and recommendation response.
[0030] The beneficial effects of this invention are as follows: it breaks through the limitations of traditional solutions that only use static statistical data, constructs a psychological behavior timeline through non-uniform resampling, and strengthens the behavioral characteristics of users during periods of high activity; it combines time series decomposition and local anomaly factor algorithms to extract multi-dimensional time series features such as trend items, periodic items, and interest mutation points, and fully restores the evolution trajectory of users' insurance needs, which can accurately capture users' real-time intentions and changes in needs, making recommendations more forward-looking.
[0031] The Time Decay Weighted (DTW) algorithm is used to weight and fuse three types of similarity: trend, period, and mutation. It focuses on assigning weight to recent user behavior, improving the sensitivity of identifying current real insurance intentions, and avoiding interference from historical outdated behaviors in the recommendation results.
[0032] Candidate users are pre-screened by using user static attributes and coarse-grained features, thus narrowing the computational scope of the DTW algorithm. At the same time, a GPU parallel acceleration unit is integrated into the similarity calculation module to support parallel retrieval of massive numbers of users, effectively reducing the overall computational overhead, adapting to large-scale user scenarios in the insurance industry, and ensuring recommendation response speed.
[0033] To address the issue of insufficient data on new users and users with low-frequency interactions, this algorithm integrates auxiliary time-series signals such as registration behavior, page dwell time, and questionnaire completion, and matches alternative features with a pre-built group behavior pattern library. It can complete similarity calculation and recommendation without complete historical behavior data, thus making up for the shortcomings of traditional recommendation algorithms in cold start scenarios. Attached Figure Description
[0034] Figure 1 This is a flowchart of an insurance recommendation method based on temporal similarity according to the present invention.
[0035] Figure 2 This is a block diagram of an insurance recommendation system architecture based on temporal similarity according to the present invention.
[0036] Figure 3 This is a flowchart of the cold start process of the present invention. Detailed Implementation
[0037] To make the content of this invention easier to understand, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Identical components are indicated by the same reference numerals.
[0038] like Figure 1 As shown, an insurance recommendation method based on temporal similarity includes the following steps: acquiring temporal data of the insurance interaction history of target users and candidate users within a preset time window, including various interaction events and corresponding timestamps, product categories, and interaction intensity; standardizing the data and converting it into psychological behavior timeline data through non-uniform resampling; extracting temporal feature vectors containing trend terms, periodic terms, and abrupt change points; calculating the comprehensive temporal similarity between users using the time decay weighted DTW algorithm, and filtering highly similar users; and generating a personalized insurance recommendation list for the target user based on the insurance preferences of similar users. The non-uniform resampling rule is to convert the original timeline into a psychological behavior timeline, making the interval between adjacent time points in the psychological behavior timeline negatively correlated with the user's real-time behavior intensity, thereby enhancing the feature representation accuracy of user's highly active behavior periods. The temporal feature vector extraction step includes: separating the trend terms and periodic terms of the behavior sequence through time series decomposition, detecting user interest abrupt change point features based on the local anomaly factor algorithm; performing sliding window segmentation on the processed behavior sequence, and generating a temporal feature vector representing the user's interest evolution trajectory by combining time embedding encoding and attention mechanisms.
[0039] like Figure 1As shown, the calculation method for the temporal comprehensive similarity is as follows: trend similarity, periodic similarity, and mutation similarity are calculated separately. A weighting factor is set using a time decay function to weight recent behavior. The three types of similarity are then weighted and fused to obtain the temporal comprehensive similarity, improving the sensitivity to identifying the user's current true intent. Before similarity calculation, a candidate user pre-screening step is included: clustering all users based on user static attributes and coarse-grained behavioral statistical features to initially screen candidate user groups, narrowing the similarity calculation range of the DTW algorithm, reducing computational costs, and improving recommendation efficiency.
[0040] The time decay function adopts an exponential decay form: Where T is the current time and t is the time when the action occurred. As a preset decay rate parameter, this weighting factor is applied to sample points or segments in the three types of similarity calculation processes: trend, periodic, and mutation, so that recent behavior contributes significantly more to the overall similarity than long-term behavior.
[0041] like Figure 1 As shown, for cold start scenarios with new users or sparse behavioral data, the following processing steps are included: Integrating auxiliary time-series signals such as user exploratory behavior during initial registration, questionnaire completion rhythm, and page dwell time trends; combining user basic attributes and risk preference tags; matching with a pre-built group behavior pattern library; and using corresponding typical behavior patterns as user substitute time-series features to complete similarity calculation. The construction steps of the group behavior pattern library are as follows: Grouping massive amounts of complete historical users based on basic attributes; clustering the time-series behavioral features of each user group; and storing the cluster centers containing common trend items, periodic items, and mutation point distributions as typical behavior patterns for the corresponding user group in the group behavior pattern library.
[0042] like Figure 2 As shown, an insurance recommendation system based on temporal similarity includes: a data acquisition module for acquiring temporal data of the historical insurance interaction behaviors of target users and candidate users; a resampling preprocessing module for completing data standardization and psychological behavior timeline transformation; a temporal feature extraction module for generating user temporal feature vectors; a similarity calculation module for calculating temporal comprehensive similarity based on the weighted DTW algorithm; a similar user filtering module for filtering highly similar user sets; and a recommendation generation module for outputting a personalized insurance recommendation list.
[0043] like Figure 3As shown, it also includes a cold start processing module: when user data is sparse, it fuses auxiliary time-series signals to construct an initial time-series feature vector, matches it with a group behavior pattern library to obtain alternative time-series features, and provides effective data support for similarity calculation. The similarity calculation module integrates a GPU-accelerated parallel DTW computing unit: it is used to realize parallel time-series similarity retrieval for large-scale user groups, adapts to massive user scenarios, and significantly improves the efficiency of similarity calculation and recommendation response.
[0044] The cold start questionnaire completion rate can be measured by the number of questions completed per unit time; the trend of page dwell time can be linearly fitted to its first-order difference sequence, using the slope as a feature. The GPU-accelerated parallel DTW computing unit is implemented using a CUDA architecture, specifically including: storing user temporal feature vectors in matrix form in GPU global memory; each thread block is responsible for calculating the DTW distance between a target user and a group of candidate users; each thread separately processes the dynamic programming matrix calculation for a pair of users, using a block optimization strategy to reduce GPU memory access latency; for temporal sequences of different lengths, padding alignment is used to ensure matrix dimension uniformity, and shared memory is used to cache the intermediate cumulative distance matrix. This parallel computing unit achieves a speedup of more than 200 times compared to a single-core CPU implementation at a user scale of millions.
[0045] In practice, the system first acquires time-series data on the insurance interaction history of target users and candidate users within a preset time window, including interaction events, timestamps, product categories, and interaction intensity. Second, the data is standardized and converted into a psychological behavior timeline through non-uniform resampling to enhance behavioral characteristics during high-activity periods. Next, time-series feature vectors containing trend terms, periodic terms, and abrupt change points are extracted to characterize the evolution trajectory of user interests. Then, a time-decay weighted DTW algorithm is used to calculate the comprehensive time-series similarity between users, and highly similar users are selected. Finally, based on the insurance preferences of these similar users, a personalized insurance recommendation list for the target user is generated. Furthermore, in cold-start or data-sparse scenarios, the system can also integrate auxiliary time-series signals and match them with a group behavior pattern library to complete effective similarity calculation and recommendation.
[0046] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An insurance recommendation method based on temporal similarity, characterized in that, Includes the following steps: Acquire time-series data of insurance interaction history between target users and candidate users within a preset time window. The data includes various interaction events and their corresponding timestamps, product categories, and interaction intensity. The data is standardized and converted into psychological and behavioral timeline data through non-uniform resampling; time-series feature vectors containing trend terms, periodic terms, and abrupt change points are extracted. The time-decay weighted DTW algorithm is used to calculate the temporal comprehensive similarity between users and to filter out highly similar users. Based on the insurance preferences of similar users, a personalized insurance recommendation list is generated for the target user.
2. The insurance recommendation method based on temporal similarity according to claim 1, characterized in that, The rule for non-uniform resampling is as follows: The original timeline is converted into a psychological behavior timeline, so that the interval between adjacent time points in the psychological behavior timeline is negatively correlated with the intensity of users' real-time behavior, thereby enhancing the accuracy of feature representation during periods of high user activity.
3. The insurance recommendation method based on temporal similarity according to claim 1, characterized in that, The time-series feature vector extraction step includes: By decomposing the behavioral sequence into trend and periodic components, and relying on the local anomaly factor algorithm to detect the characteristics of user interest mutation points, the processed behavioral sequence is segmented by a sliding window, and combined with temporal embedding encoding and attention mechanism to generate a temporal feature vector representing the evolution trajectory of user interest.
4. The insurance recommendation method based on temporal similarity according to claim 1, characterized in that, The calculation method for the temporal comprehensive similarity is as follows: Trend similarity, periodic similarity, and mutation similarity are calculated separately. A weighting factor is set using a time decay function to weight recent behaviors. The three types of similarity are then weighted and fused to obtain a time-series comprehensive similarity, which improves the sensitivity of identifying the user's current true intention.
5. The insurance recommendation method based on temporal similarity according to claim 1, characterized in that, Before similarity calculation, a candidate user pre-screening step is included: Clustering of all users based on user static attributes and coarse-grained behavioral statistical features preliminarily filters candidate user groups, narrows the similarity calculation range of the DTW algorithm, reduces computational costs, and improves recommendation efficiency.
6. The insurance recommendation method based on temporal similarity according to claim 1, characterized in that, For cold start scenarios with new users or sparse behavioral data, the following processing steps are included: By integrating auxiliary time-series signals such as user exploratory behavior during the initial registration phase, questionnaire completion pace, and page dwell time trends, and combining them with user basic attributes and risk preference tags, a pre-built group behavior pattern library is matched, and corresponding typical behavior patterns are used as user substitute time-series features to complete similarity calculation.
7. The insurance recommendation method based on temporal similarity according to claim 1, characterized in that, The steps for constructing the group behavior pattern library are as follows: Based on basic attributes, massive amounts of complete historical users are grouped, and the temporal behavioral characteristics of each user group are clustered. The cluster centers containing common trend items, periodic items, and distribution of mutation points of the group are used as the typical behavioral patterns of the corresponding user groups and stored in the group behavior pattern library.
8. An insurance recommendation system based on temporal similarity, characterized in that, include: The data acquisition module is used to obtain time-series data on the historical insurance interaction behavior of target users and candidate users; The resampling preprocessing module is used to complete data standardization and psychological behavior timeline transformation; The temporal feature extraction module is used to generate user temporal feature vectors; The similarity calculation module is used to calculate the overall temporal similarity based on the weighted DTW algorithm; The similar user filtering module is used to filter sets of highly similar users; The recommendation generation module is used to output a personalized insurance recommendation list.
9. An insurance recommendation system based on temporal similarity according to claim 8, characterized in that, It also includes a cold start processing module: when user data is sparse, it integrates auxiliary time series signals to construct an initial time series feature vector, matches the group behavior pattern library to obtain alternative time series features, and provides effective data support for similarity calculation.
10. An insurance recommendation system based on temporal similarity according to claim 9, characterized in that, The similarity calculation module integrates a GPU-accelerated parallel DTW computing unit: it is used to realize parallel time-series similarity retrieval for a large user group, adapt to massive user scenarios, and greatly improve the efficiency of similarity calculation and recommendation response.