AI intention shunting recommendation method and system based on user re-purchase behavior recognition

Abstract

The invention discloses an AI intention shunting recommendation method and system based on user re-purchase behavior recognition, and is applied to a take-out service platform. According to the method, a re-purchase frequency N is established by tracking historical order placing times of a user to a service provider, and a page type currently accessed by the user is combined as a technical signal to judge whether the user is in an efficiency scene or an exploration scene. In an efficiency scene, the system generates a service list arranged in a descending order according to a re-purchase frequency, and provides a quick entry of'one-order-again '; in an exploration scene, the system generates an initial candidate set based on multi-dimensional behavior preference, and applies negative weight adjustment to a high-frequency re-purchase service provider according to a re-purchase frequency N or a time-weighted re-purchase frequency N ', a weight reduction coefficient alpha = max (0, 1-k. (N-1)) or alpha = max (0, 1-k. (N'-1)), and when N is greater than or equal to S, direct shielding is performed. And after adjustment, performing diversity filtering, and setting a commercial cooperation weight factor to zero to generate a pure exploration recommendation list. According to the invention, through a set of quantitative AI-driven shunting and counterbalance mechanism, two opposite demands of exploration and re-purchase are met. An A / B test verifies that the exposure click rate of a new service provider in an exploration scene is increased by about 25%, and the average operation path of a user in a re-purchase scene is shortened to about two times of click.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and recommendation systems, specifically to an intelligent recommendation method and system applied to a food delivery service platform. This invention aims to solve the problem that recommendation systems cannot distinguish and efficiently respond to the two opposing intentions of users—"exploration" and "repurchase"—by establishing an intent recognition mechanism based on a quantified repeat purchase frequency decay rule. Background Technology

[0002] With the rapid development of the local life services (especially food delivery) industry, personalized recommendation systems have become a core technical component for improving user experience, promoting merchant conversion, and optimizing the platform ecosystem. However, the recommendation algorithms of existing mainstream food delivery platforms generally suffer from the following technical deficiencies: Relying on user-configured preferences: Most systems require users to pre-fill information such as taste, price range, and category preferences. If users do not set these, the recommendation results are often based on global popular trends or simple historical clicks, which are difficult to reflect real needs and result in low matching accuracy.

[0003] The recommendation logic is too simplistic and easily leads to "information cocoons": Existing algorithms mostly use collaborative filtering or content-based similarity recommendations, which over-focus on users' historical order categories or browsing history, resulting in long-term recommendations of similar service providers and inhibiting consumption diversity; at the same time, new service providers and non-leading service providers lack exposure opportunities, which undermines the fairness of the platform ecosystem.

[0004] The weighting calculation dimension is one-sided: Traditional solutions usually only consider limited factors such as category, price or rating, ignoring the basic needs dimensions (such as delivery distance and timeliness) and extended evaluation dimensions (such as packaging quality and on-time rate) that users are highly sensitive to in ordering decisions, resulting in a disconnect between the recommendation results and actual selection behavior.

[0005] The existing system fails to build an intent-based traffic segmentation mechanism centered on repurchase frequency: it fails to deeply utilize the key behavioral signal of "the number of times a user repurchases from a service provider." The core contradiction lies in the fact that whether a user wants to quickly find their favorite restaurant or discover new options, the system uses the same complex sorting logic that mixes historical preferences, similarity, and commercial weights. This results in: when users want to repurchase, they have to painstakingly search through a multitude of recommendations for familiar favorites; when users want to explore, they are bombarded with familiar faces and advertisements. The system lacks the intelligence to proactively "segment" and "check" based on repurchase frequency, failing to achieve precise "what you want is what you get" service. Therefore, there is an urgent need for a recommendation method and system that can intelligently determine user intent based on repurchase behavior and provide both an extremely efficient repurchase channel and a clean and novel exploration channel. Summary of the Invention

[0006] I. Technical problems to be solved This invention aims to solve the following key technical problems existing in current food delivery recommendation systems: Exploring the issues of information cocoons and commercial interference in various scenarios: When users browse without a clear goal, the system cannot effectively suppress the excessive influence of their own high-frequency repurchase history and commercial promotional content on the recommendation results.

[0007] The issues of search efficiency and memory burden in repeat purchase scenarios: When users have a clear intention to repeat purchase, the system lacks a direct and efficient path, and users cannot quickly locate their target based on the popularity of repeat purchases.

[0008] The system lacks quantifiable and reproducible control rules: there is no clear mathematical model or threshold logic for how to "reduce weight" or "block" based on repurchase frequency, resulting in unstable technical effects and inaccurate evaluation.

[0009] Technical solution To address the aforementioned issues, this invention proposes an "AI intent recognition and dual-path traffic splitting and balancing mechanism based on repurchase frequency." Its core idea is to elevate "user-service provider repurchase frequency" to the core metadata and AI control hub of the system, allowing it to play distinctly different but complementary roles in different scenarios, thereby achieving intelligent traffic allocation and simultaneously tackling the two major pain points.

[0010] Methods and solutions The intelligent recommendation method includes the following core steps: 1. Basic steps for repeat purchase frequency tracking and intent segmentation determination: Repeat purchase frequency tracking: The system continuously records users' historical order behavior for each service provider. When the cumulative number of orders N ≥ 1, the service provider is marked as an associated item for the user, and its subsequent behavior will be subject to contextualized control.

[0011] Intent-based scenario segmentation: The system classifies user visits to the order history page or dedicated repeat purchase channel as "efficiency scenarios," and visits to the application homepage or regular recommendation feed as "exploration scenarios." Page visits become a reliable technical signal for distinguishing users' instantaneous intent.

[0012] Efficiency Scenario Response: Positive Sequence List of Repeat Purchase Popularity and Minimalist Operation When an efficiency-critical scenario is identified, the system activates its highly efficient response logic: it extracts the user's related items from the repeat purchase frequency management pool, generates a list sorted in descending order of repeat purchase frequency N, and provides a "Order Again" shortcut for each item in the list. This allows users to complete the entire process from entering the page to successfully placing an order within 2-3 clicks, directly solving the pain point of "wanting to eat something you like but not being able to find it."

[0013] Exploring Scenario Response: AI-Driven Negative Balance Logic When the scenario is determined to be an exploration scenario, the system initiates a pure exploration logic centered on "quantitatively suppressing historical inertia": Step A: Generate an initial candidate set. An initial recommendation candidate set is generated based on the user's recent multi-dimensional behavioral preferences. These multi-dimensional preferences include at least: the frequency of interaction with different service providers within a preset sliding time window; category preference weights calculated using a time decay factor; distance sensitivity calculated based on historical order distance distribution; and delivery timeliness preferences calculated based on historical order delivery timeliness feedback.

[0014] Step B: Quantifying and Adjusting Negative Weights Based on Repeat Purchase Frequency (Core). The system queries the cumulative number of orders (N) placed by each service provider for the current user in the initial candidate set, and adjusts its original recommendation weight according to the following quantification rules: (i) If N=1, the weights are not reduced; (ii) If N≥2, calculate the weight reduction coefficient α, with the formula: α = max(0, 1 - k·(N - 1)), or optionally calculate the time-weighted repurchase frequency N' = Σ(order number_i · time decay coefficient_i), and calculate α = max(0, 1 - k·(N' - 1)) based on N'; the adjusted weight = original weight × α; (iii) If N ≥ the masking threshold S, then the weight is directly reset to zero, which is equivalent to completely masking it from the current exploration list.

[0015] Step C: Generate the final exploration list. After weight adjustment, further rules such as category diversity and service provider type balance are executed, and the fixed weight factor of commercial cooperation is set to zero during the sorting process, finally generating the exploration recommendation list.

[0016] System Solution Corresponding to the above method, the intelligent recommendation system of the present invention achieves a clear logical and physical separation in its architecture, including: Repeat purchase frequency tracking module: used to record and maintain the cumulative number of orders N between users and service providers; Intent Scenario Determination Module: Used to determine whether the current scenario is an efficiency scenario or an exploration scenario based on the type of page the user visits; Efficiency Scenario Response Engine: Connected to the repeat purchase frequency tracking module, it is used to generate and output a service list sorted in descending order by the cumulative number of orders N and with the "order again" function in efficiency scenarios; Explore the scenario response engine: connected to the repurchase frequency tracking module, integrating the following units: Preference modeling unit: used to generate the initial recommendation candidate set; Negative weight adjustment unit: used to execute AI control strategies based on repurchase frequency; Filtering and Optimization Unit: Used to perform diversity filtering and set the business weight factor to zero.

[0017] Beneficial effects Compared with the prior art, the present invention has the following significant advantages: Achieve accurate differentiation and efficient response to user intent: Through page access signals and quantitative repurchase control rules, the system can clearly distinguish and optimize the exploration and repurchase paths respectively.

[0018] The technical effects are stable and measurable: By introducing quantitative parameters such as the weight reduction coefficient α = max(0, 1 - k·(N - 1)) and the masking threshold S, the control logic is clear, calculable, and optimizable. Platform A / B testing verified that with a configuration of k=0.2 and S=5, the click-through rate of new service providers in the exploration scenario increased by approximately 25%, and the average user operation path in the repeat purchase scenario was shortened to approximately 2 clicks.

[0019] Building an endogenous ecological balance mechanism within the platform: Quantitative control rules automatically suppress excessive exposure of a single top service provider on the user side, providing stable exposure opportunities for new service providers.

[0020] Non-obviousness of parameter selection and peak technical effect: Large-scale parameter scanning experiments revealed a significant non-linear impact of the attenuation coefficient k and the masking threshold S. When k < 0.15, the improvement effect on the exposure of new service providers is insufficient (<15%); when k > 0.25, the user complaint rate regarding the relevance of the recommendation results increases significantly (+37%). Similarly, when S < 4, high-frequency repeat customers feel a disconnect in the experience due to the overly aggressive exploration flow, leading to increased churn; when S > 6, the novelty of the exploration flow does not significantly improve, and marginal benefits diminish. Comparative experiments show that when the attenuation coefficient k = 0.2 and the masking threshold S = 5, the product of the click-through rate of new service providers and user satisfaction in the exploration scenario reaches its peak. This configuration is 19% better than the suboptimal solution (k = 0.15 / S = 3) and 23% better than another alternative solution (k = 0.25 / S = 6). This proves that the specific parameter combination (k=0.2, S=5) proposed in this invention is not a conventional or obvious choice in the field, but rather an optimal balance point that can maximize the technical effect, as verified by experiments. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of the present invention clearer, a detailed description is provided below in conjunction with specific embodiments.

[0022] Example 1: Demonstration of Diversion and Quantification Check and Balance Mechanism User A is an active user on the platform. They placed 12 orders (N=12) from Sichuan cuisine provider X, 8 orders (N=8) from milk tea provider Y, and 3 orders (N=3) from light meal provider Z. The system configuration is k=0.2, S=5.

[0023] Scenario 1: Efficiency Scenario Response User A enters the "Order History Page". The system determines it to be an efficiency scenario and displays the following list: ① Sichuan cuisine service provider X (12 times) [Order again]; ② Milk tea service provider Y (8 times) [Order again]; ③ Light meal service provider Z (3 times) [Order again].

[0024] Scenario 2: Exploring Scenario Response User A opens the app's homepage. The system identifies it as an exploration scenario.

[0025] Generate an initial candidate set (including the aforementioned service providers).

[0026] Negative weight adjustment: "Sichuan cuisine service provider X": N=12 ≥ S=5 → Weight reset to zero (masked).

[0027] “Milk tea service provider Y”: N=8 ≥ S=5 → Weight reset to zero (masked).

[0028] “Light Meal Service Provider Z”: N=3, α = max(0, 1 - 0.2·(3-1)) = 0.6 → weight reduction of 40%.

[0029] Implement diversity filtering and zero-commercial-rights reset.

[0030] Final presentation: In the homepage recommendation list, high-frequency repurchase service providers are blocked, "light food service provider Z" is ranked lower, and new service providers are prominently displayed.

[0031] Example 2: Integrated Application of Quantitative Adjustment and Multi-Dimensional Preferences In user C's personalized weight vector, "free delivery" has a weight of 40%, while "distance ≤ 2 km" has a weight of 32%, which is conflicting. Their repurchase frequency with "service provider M" is N=3.

[0032] In the exploration scenario, the system first generates an initial candidate set containing "service provider M".

[0033] Negative weight adjustment: N=3, α = 0.6, reduce the weight of "service provider M".

[0034] The system handles weight conflicts by prioritizing service providers that meet the "free delivery" requirement. Service provider M meets the "free delivery" requirement but is 2.5 kilometers away.

[0035] Comprehensive ranking: The system combines the reduced weight of "service provider M" with the distance factor to calculate its final ranking position.

[0036] Example 3: The effect of time decay factor on weighted repurchase frequency User D placed a total of 3 orders with "Service Provider P", with the following time distribution: 2 within the last 7 days (time decay coefficient 1.2) and 1 earlier (coefficient 1.0).

[0037] In step B of the scenario exploration, the system can calculate the time-weighted repurchase frequency N' = 2·1.2 + 1·1.0 = 3.4.

[0038] Substituting N' into the adjustment formula: α = max(0, 1 - 0.2·(3.4 - 1)) = 0.52. The adjustment based on N' is more stringent than the adjustment based on N=3 (α=0.6), reflecting a stronger suppression of recent concentrated repurchase behavior.

[0039] Example 4: Comprehensive Application of Time Weighting Scheme in Exploration Scenarios User E's repurchase behavior with "Service Provider Q" includes: 5 orders within 30 days (time decay coefficient 1.0), 3 orders within 15 days (coefficient 1.1), and 2 orders within 7 days (coefficient 1.2). The system calculates the time-weighted repurchase frequency N' = 5·1.0 + 3·1.1 + 2·1.2 = 10.7.

[0040] In the exploration scenario, if N'=10.7 ≥ S=5, the system directly resets the service provider's weight to zero and blocks it from the exploration recommendation list. Compared to using only the original repurchase frequency N=10, this solution can more accurately identify recently active high-frequency repurchase behaviors and effectively prevent users from being disturbed by overly familiar service providers in the exploration scenario.

[0041] Example 5: Calculation process of multi-dimensional preference weights User F's interaction behavior across different product categories in the past 30 days is as follows: Chinese meals: 12 times (8 times in the last 7 days, 3 times in days 8-15, and 1 time in days 16-30). Western meals: 5 times (3 times in the last 7 days, once in days 8-15, and once in days 16-30). Japanese food: 3 times (2 times in the last 7 days, 1 time in the last 8-15 days) The system sets the time decay coefficient as follows: 1.2 for the last 7 days, 1.1 for 8-15 days, and 1.0 for 16-30 days.

[0042] Category preference weight calculation: Chinese cuisine: (8·1.2 + 3·1.1 + 1·1.0) = 13.9 Western cuisine: (3·1.2 + 1·1.1 + 1·1.0) = 5.7 Japanese cuisine: (2·1.2 + 1·1.1) = 3.5 After normalization: Chinese food: 13.9 / (13.9 + 5.7 + 3.5) = 60.4% Western food: 24.8% Japanese cuisine: 15.2% The category preference weight will be used to generate the initial recommendation candidate set.

[0043] Example 6: Specific application of diversity filtering rules During the final list generation phase of the exploration scenarios, the system executes the following diversity rules: If the proportion of the same category is ≤30%, then if 35% of the candidates are Chinese food service providers, then some Chinese food service providers will be randomly removed until the proportion of Chinese food service providers drops to 30%.

[0044] New / Non-leading service providers ≥ 40%: If new and non-leading service providers account for only 30% of the candidate pool, such service providers will be added from the pool until the total reaches 40%.

[0045] Commercial rights reset to zero: The commercial cooperation weight factor of all candidate service providers is set to 0 to ensure that the recommendation results are not affected by commercial promotion.

[0046] After the above processing, a clean list of exploration recommendations is generated, which avoids information cocoons and ensures the diversity of the platform ecosystem. Attached Figure Description

[0047] Figure 1 The flowchart of the AI ​​intent-based recommendation method based on user repeat purchase behavior recognition provided in this embodiment of the invention illustrates the dual-path recommendation execution logic for efficiency scenarios and exploration scenarios.

[0048] Figure 2 The structural block diagram of the intelligent recommendation system provided in the embodiments of the present invention shows the connection relationship between the repurchase frequency tracking module, the intent scene determination module, the efficiency scene response engine, the exploration scene response engine, and their internal units.

[0049] Detailed implementation method (continued) The above embodiments are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this invention, and these modifications or substitutions should all be covered within the scope of protection of this invention. For example: The specific value of the time decay coefficient can be adjusted according to business needs. For example, the most recent 3 days can be set to 1.3, 4–7 days to 1.2, 8–14 days to 1.1, and 15–30 days to 1.0. The percentage thresholds (30%, 40%) in the diversity filtering rules can be dynamically adjusted according to the platform's development stage and business objectives; The blocking threshold S can be customized according to user behavior characteristics, such as setting S=3 for new users and S=5 for old users. The weight reduction coefficient α can be calculated using a nonlinear function, such as α = max(0, 1 - k·(N - 1)^2), to achieve more precise control.

Claims

1. An AI intent-based recommendation method for user repeat purchase behavior recognition, characterized in that, Includes the following steps: User repurchase frequency tracking steps: Record users' historical order behavior with service providers and maintain the cumulative number of orders N for each service provider; User intent scenario determination steps: Based on the type of page the user is currently visiting, determine whether their intent scenario is an efficiency scenario or an exploration scenario; Efficiency scenario response steps: When the scenario is determined to be an efficiency scenario, the user is shown a list of service providers containing the number of orders N≥1 and sorted in descending order of repurchase frequency N. The list items are associated with the "Order Again" quick operation function. Exploration Scene Response Steps: When a scene is identified as an exploration scene, perform the following operations to generate an exploration recommendation list.

2. A system for implementing the recommended method of claim 1, characterized in that, include: The repeat purchase frequency tracking module is used to record and maintain the cumulative number of orders N between users and service providers; The intent scenario determination module is used to determine whether the current scenario is an efficiency scenario or an exploration scenario based on the type of page the user visits; An efficiency scenario response engine, connected to the repeat purchase frequency tracking module, is used to generate and output a service list sorted in descending order by the cumulative number of orders N and with the "order again" function in efficiency scenarios. An exploration scenario response engine, connected to the repurchase frequency tracking module, includes: a preference modeling unit, used to generate an initial recommendation candidate set based on multi-dimensional user behavior data, the multi-dimensional behavior data including at least: the frequency of interaction behavior with different category service providers within a preset sliding time window, the category preference weight calculated by combining time decay factor, the distance sensitivity calculated based on historical order distance distribution, and the delivery time preference calculated based on historical order delivery time feedback; a negative weight adjustment unit, used to perform weight reduction based on the cumulative number of orders N, wherein the weight reduction method includes any of the following: (i) calculating the weight reduction coefficient α according to the formula α = max(0, 1 - k·(N - 1)); (ii) calculating the time-weighted repurchase frequency N' = Σ(number of orders_i · time decay coefficient_i), and calculating the weight reduction coefficient α according to the formula α = max(0, 1 - k·(N' - 1)); when N ≥ S, the weight is reset to zero, where k is the decay coefficient and S is the masking threshold; The filtering and optimization unit is used to perform diversity filtering on the weighted results and set the business cooperation weight factor to zero.

3. The method or system according to claim 1 or 2, characterized in that, In step (2), the weight reduction coefficient α = max(0, 1 - k·(N - 1)) or α = max(0, 1 - k·(N' - 1)), where k is a preset attenuation coefficient.

4. The method or system according to claim 3, characterized in that, The attenuation coefficient k is between 0.1 and 0.

25.

5. The method or system according to claim 1 or 2, characterized in that, The shielding threshold S is between 2 and 5.

6. The method or system according to claim 1 or 2, characterized in that, When a user visits the order history page or a dedicated repeat purchase aggregation page, it is considered an efficiency scenario; when a user visits the application's main page or a regular recommendation feed page, it is considered an exploration scenario.

7. The method according to claim 1, characterized in that, In step (1), the category preference weight is calculated by combining the frequency of user interaction with different category service providers within a preset sliding time window with a time decay factor.

8. The method according to claim 1, characterized in that, The diversity filtering rules include: limiting the proportion of service providers of the same category to no more than 30%, and / or ensuring that the combined proportion of newly joined service providers and non-leading service providers is no less than 40%.

9. The method according to claim 1, characterized in that, The diversity filtering rules include: limiting the proportion of service providers of the same category to no more than 30%, and / or ensuring that the combined proportion of newly joined service providers and non-leading service providers is no less than 40%.

Images