A Product Recommendation Method and System Based on High-Efficiency Continuous Negative Sequence Mining in the Uncertain Era
By constructing a sequence information table of probabilistic attributes and expanding continuous terms, combined with pruning strategies, the problems of uncertainty and high computational complexity in negative sequence mining in existing technologies are solved, achieving efficient and accurate negative sequence pattern mining, which is suitable for time series recommendation.
Patent Information
- Application Number
- CN202611354910.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-03
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-09-03
AI Technical Summary
Existing negative sequence mining methods cannot effectively handle data uncertainty, resulting in inaccurate mining results and high computational complexity, making it difficult to meet the needs of real-time product recommendation.
By constructing a sequence information table containing probabilistic attributes, positive candidate patterns are mined and negatively transformed. Continuous term expansion and pruning strategies are adopted to calculate probabilistic utility and its upper limit, and invalid patterns are pruned to ensure temporal continuity and accuracy.
It improves the accuracy of data mining, identifies deep business value, eliminates the paradox of negative term calculation, reduces computational overhead, and ensures temporal continuity, making it suitable for time-series recommendation scenarios.
Smart Images

Figure CN122887571B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of data mining technology, specifically relating to a product recommendation method and system based on uncertain and efficient continuous negative sequence mining. Background Technology
[0002] With the rapid development of e-commerce, user behavior data has experienced explosive growth. How to extract valuable information from massive amounts of data for accurate product recommendations has become a research hotspot. Sequence pattern mining, as an important branch of data mining, can discover the patterns in the occurrence of itemsets over time series.
[0003] Traditional sequence pattern mining primarily focuses on positive sequences, i.e., what users "bought." However, in real-world business scenarios, what users "didn't buy" (i.e., purchases that didn't occur) often implies deeper needs or churn risks. For example, if a user didn't buy product A at a certain point but later bought product B, this negative behavioral pattern containing "no purchases" is a negative sequence pattern. Mining such negative patterns is crucial for analyzing changes in user needs and optimizing recommendation strategies.
[0004] Furthermore, real-world data often exhibits uncertainty. In actual e-commerce operations, the data collection process is frequently affected by factors such as network latency, cache anomalies, proxy ordering, or accidental clicks, leading to noise in the recorded purchase behavior. Therefore, the data itself possesses a probabilistic attribute reflecting its accuracy. Existing negative sequence mining methods are mostly designed for deterministic data and cannot effectively handle data uncertainty, resulting in inaccurate mining results.
[0005] Meanwhile, existing efficient sequence mining algorithms typically face problems of high computational complexity and huge search space. During recursive expansion, without efficient pruning strategies, the algorithm will get bogged down in massive amounts of ineffective computation. In particular, for handling continuity constraints (i.e., emphasizing tight temporal continuity) and negative term expansion, existing technologies often lack a unified optimization framework, and are prone to logical contradictions when calculating the probability and utility of negative terms, leading to low mining efficiency and difficulty in meeting the needs of real-time product recommendation.
[0006] Therefore, there is an urgent need for a mining method and system that can integrate uncertainty, efficiently mine continuous negative sequence patterns, eliminate the paradox of negative term calculation, and effectively reduce computational costs. Summary of the Invention
[0007] To overcome the shortcomings of existing technologies, this invention provides a product recommendation method and system based on uncertain and efficient mining of continuous negative sequences.
[0008] This invention is achieved through the following technical solution: A product recommendation method based on uncertain and efficient continuous negative sequence mining includes the following steps: Step S1: Construct a sequence information table containing probability attributes. The sequence information table includes several user IDs and an uncertain transaction sequence corresponding to each user ID. The uncertain transaction sequence includes several itemsets arranged in chronological order. Each itemset contains at least one item purchased under the same timestamp. Each item includes a product code and the utility, purchase quantity, probability, and remaining utility corresponding to the product code. Step S2: Based on the sequence information table, mine all positive candidate patterns of length 1, sort the positive candidate patterns in descending order of probability utility value and select the first K, perform negative transformation on the first K positive candidate patterns respectively to obtain K negative expansion terms; Step S3: Take the positive candidate pattern as the first generation candidate pattern, establish the position utility probability chain corresponding to the first generation candidate pattern, define the recursive expansion generation r, and initialize r=1; Step S4: Perform recursive expansion, using a single r-th generation candidate pattern as the expansion prefix in sequence, and based on the sequence information table and the position utility probability chain corresponding to the r-th generation candidate pattern, perform continuous term expansion using positive and negative expansion terms respectively to obtain several expanded candidate patterns, and update the position utility probability chain for each expanded candidate pattern; the continuous term expansion is used to ensure the temporal continuity of the expanded pattern in the uncertain transaction sequence. Step S5: Calculate the probabilistic utility pu(candidate) and probabilistic utility bound Bound(α,x,ρ,Q) for each expanded candidate pattern, where α represents a single candidate pattern in the current recursive expansion generation r, x represents the expansion item currently participating in the expansion, ρ represents the itemset position of the expansion item currently participating in the expansion in the user sequence, and Q represents the user ID; define a minimum utility threshold min_util. If pu(candidate) ≥ min_util, then the expanded candidate pattern is determined to be an uncertain high-utility continuous negative sequence pattern. Step S6: Traverse all user IDs, let Q denote a user ID, and determine the maximum value maxBound(Q) of the probabilistic utility upper limit corresponding to each user ID, where the maxBound(Q) is the maximum value among all probabilistic utility upper limits Bound(α,x,ρ,Q) under a single user ID; define a comparison value T as the sum of maxBound(Q) corresponding to all user IDs, if T≥min_util, it is determined to continue recursive expansion, the expanded candidate pattern is taken as the (r+1)-th generation candidate pattern, then let r=r+1, and return to execute step S4; if T<min_util, a pruning operation is performed to prune the current expanded candidate pattern, terminate the recursive expansion, and output all mined uncertain high-utility continuous negative sequence patterns; Step S7: Based on the mined uncertain high-utility continuous negative sequence patterns, the probabilistic utility of the uncertain high-utility continuous negative sequence patterns calculated in step S5 is taken as the actual utility value, the commodity recommendation priority is adjusted according to the rule that a pattern with a larger actual utility value has a higher priority, and commodity recommendation is performed according to the adjusted commodity recommendation priority.
[0009] Further, the construction process of the sequence information table includes: Collect shopping data, where the shopping data includes user IDs, and shopping time, commodity name, commodity price, purchase quantity and probability corresponding to the user IDs; Encode the commodity names to generate unique commodity codes; Calculate the utility and remaining utility of each item, where the utility is the product of the commodity price and the purchase quantity in a single item, and the remaining utility is the sum of the utilities of all items after this item.
[0010] Further, the method for determining the minimum utility threshold min_util is: define the sequence information table as a data set D, calculate the total item utility u(D) of all items in the data set D, define the minimum threshold proportional coefficient , and the minimum utility threshold min_util is calculated by the formula min_util= × u(D).
[0011] Further, the process of establishing the position utility probability chain includes: For each candidate pattern, record its corresponding user ID, the end position of the item set in the sequence data, pattern utility, and pattern probability.
[0012] Further, the specific process of performing continuous item expansion in step S4 is: Determine the user ID and the end item set position of the r-th generation candidate pattern according to the position utility probability chain; The positive expansion items include inner positive expansion items and outer positive expansion items, both of which are determined based on the sequence information table; the inner positive expansion item is the item in the itemset at the end of the itemset position in the sequence information table for the r-th generation candidate pattern; the outer positive expansion item is the item in the itemset next to the end of the itemset position in the sequence information table for the r-th generation candidate pattern. When performing positive item expansion, positive expansion items are added to the r-th generation candidate patterns to generate expanded candidate patterns; the positive item expansion includes positive item set in-set expansion and positive item set out-of-set expansion. The positive item set expansion refers to adding the positive expansion item to the itemset at the end of the candidate pattern in the rth generation, thus forming a positive item set expansion pattern. The positive itemset expansion refers to adding the positive expansion itemset as a new itemset to the end of the r-th generation candidate pattern to form the positive itemset expansion pattern.
[0013] Furthermore, the specific process of expanding consecutive terms in step S4 also includes: When performing negative item expansion, the negative expansion item is added to the r-th generation candidate pattern to generate the expanded candidate pattern; the negative item expansion includes negative item set in-set expansion and negative item set out-of-set expansion; The negative item in-item set expansion refers to directly adding the negative expansion item to the itemset at the end of the candidate pattern in the rth generation; The negative item set expansion refers to adding the negative expansion item as a new itemset to the end of the r-th generation candidate pattern, and determining an accompanying positive expansion item according to the sequence information table, and adding the accompanying positive expansion item as a new itemset after the negative expansion item. The method for determining the accompanying positive expansion item is: taking the item items in the first or second itemset after the end itemset position of the r-th generation candidate pattern in the sequence information table as the accompanying positive expansion item.
[0014] Furthermore, the specific method for synchronously updating the location utility probability chain and calculating the probability utility pu(candidate) is as follows: For each expanded candidate pattern, a position utility probability chain is established. The utility in this positional utility probability chain is the sum of the utility u(α) of the current candidate pattern at the current recursive expansion generation r and the utility of all expansion terms; if the expansion term is a positive expansion term, then the utility u(x) of the positive expansion term is... p ) represents the original utility of the corresponding item in the sequence information table. If the extended item is a negative extended item, the utility u(x) of the negative extended item is... n ) is 0; The probability in this positional utility probability chain is the product of the probability pr(α) of the current candidate mode at the current recursive expansion generation r and the actual probabilities of all expansion terms; if the expansion term is a positive expansion term, then the actual probability pr(x) of the positive expansion term is...p The original probability of the corresponding item in the sequence information table is pr(x). If the extended item is a negative extended item and a corresponding item exists at the corresponding position in the sequence information table, then the actual probability of the negative extended item is pr(x). n The original probability of the corresponding item is p(x). If the extended item is a negative extended item and the corresponding position in the sequence information table does not contain a corresponding item, then the actual probability pr(x) of the negative extended item is p(x). n ) is 1; Calculate the probabilistic utility of the expanded candidate pattern under each user ID, where the probabilistic utility is the maximum value of the product of utility and probability in the positional utility probability chain; sum the probabilistic utilities under all user IDs to obtain the probabilistic utility pu(candidate) of the expanded candidate pattern.
[0015] Furthermore, for both positive itemset in-set expansion and positive itemset out-of-set expansion, the formula for calculating the probability utility upper bound Bound(α,x,ρ,Q) is as follows: Bound(α,x,ρ,Q)=(u(α)+u(x p )+ru(Q\(x p ,ρ)))×(pr(α)×pr(x p )); For positive itemset expansion, the formula for calculating the probability utility upper bound Bound(α,x,ρ,Q) is as follows: Bound(α,x,ρ,Q)=(u(α)+u(x n )+ru(Q\(x n ,ρ)))×(pr(α)×pr(x n )); For negative itemset expansion, the formula for calculating the probability utility upper bound Bound(α,x,ρ,Q) is as follows: Bound(α,x,ρ,Q)=(u(α)+u(x n )+u(x p )+ru(Q\(x p ,ρ)))×(pr(α)×pr(x n )×pr(x p )); in, ru(Q\(x p ,ρ)) represents the residual utility of the positive expansion item at the ρth item set of user Q in the sequence information table; ru(Q\(x n,ρ)) represents the remaining utility of the product item corresponding to the negative extension item at the ρ-th product item set of user Q in the sequence information table. If there is no corresponding product item for the negative extension item at the ρ-th product item set of user Q in the sequence information table, then ru(Q\(x n ,ρ)) represents the remaining utility of the last product item at the ρ-th product item set of user Q in the sequence information table; u(x p ), u(x n ), pr(x p ) and pr(x n ) are determined according to step S5.
[0016] Further, the specific process of step S6 is as follows: For a single user Q, traverse the probabilistic utility upper bound Bound(α,x,ρ,Q) of all extended candidate patterns, and select the maximum value among them as the maximum value maxBound(Q) of the probabilistic utility upper bound corresponding to the user Q; Sum the maxBound(Q) corresponding to all users Q to obtain the comparison value T; If T≥min_util, it is determined to continue the recursive extension; if T<min_util, the pruning operation is performed, and the recursive extension is terminated.
[0017] A product recommendation system based on uncertain high-utility continuous negative sequence mining, comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, wherein the processor implements the method as described above when executing the program.
[0018] The beneficial effects that the present invention can achieve are as follows: 1. Capable of processing data uncertainty and improving mining accuracy: When constructing the sequence information table, the present invention introduces a probability attribute representing "data correct rate" for each product item, and calculates the probabilistic utility on this basis; the probabilistic utility pu(candidate) of the candidate pattern integrates the result of multiplying utility and probability. Through the above technical means, the present invention takes probability as a calculation factor in all links of data mining, so that when facing real e-commerce data with collection noise, it can still accurately quantify the true value of the pattern, and solves the problem of inaccurate mining results in the prior art.
[0019] 2. Effectively Mining Negative Sequence Patterns to Identify Deep Business Value: This invention generates negative extension items by negatively transforming mined positive candidate patterns of length 1. When expanding consecutive items, specific rules are clearly defined for expansion within and outside negative item sets (e.g., adding negative extension items as new itemsets, followed immediately by positive extension items to maintain temporal continuity). Through this expansion mechanism, the system of this invention can accurately construct negative behavioral patterns containing "no purchase" in time series, providing strong data support for analyzing changes in potential user needs and overcoming the limitations of traditional methods that can only mine purchased behaviors.
[0020] 3. Constructing a rigorous probability calculation logic to eliminate the paradox of negative term calculation: This invention employs specific technical means when updating the position utility probability chain: On the one hand, it explicitly states that "if the extension term is a negative extension term, then the utility is recorded as 0," avoiding the unreasonable cancellation of overall utility due to the introduction of negative terms; on the other hand, it innovatively sets classification rules for actual probability values, especially for the absolute negative state of "negative extension term and no corresponding product item in the table," directly setting its actual probability to 1. This technical means completely opens up the mathematical logical loop of negative terms participating in probability multiplication, avoiding program crashes and calculation paradoxes caused by no value found in the table, and ensuring the mathematical self-consistency of the pattern measurement.
[0021] 4. Introducing an efficient pruning strategy to significantly reduce computational overhead: This invention designs a formula for calculating the probability utility upper bound (Bound), constructing an upper bound closure principle using residual utility; and uses the sum of maxBound values for all user IDs as a comparison value for pruning judgment. Simultaneously, it clarifies the temporal logic of "using the expanded candidate pattern as the α+1th generation candidate pattern, and then updating the value of α in the recursive expansion generation," ensuring seamless connection between pruning judgment and recursive loop. Through the above mechanism, this invention can prune branches that are unlikely to become efficient patterns before recursive expansion, greatly compressing the search space and improving the algorithm's running efficiency on large-scale datasets.
[0022] 5. Ensuring temporal continuity and conforming to time-series recommendation scenarios: When expanding, this invention clearly distinguishes between "inner positive expansion items" (items within the itemset at the end of the itemset) and "outer positive expansion items" (items within the next itemset). By adding inner positive expansion items to the current itemset or adding outer positive expansion items as new itemsets, the temporal continuity of the expanded patterns in uncertain transaction sequences is strictly guaranteed. This technique based on continuous item expansion rules ensures that the mined patterns have a close correlation in the time dimension, giving them higher reference value in time-series recommendation scenarios. Attached Figure Description
[0023] Figure 1This is a schematic diagram of the positional utility probability chain of the candidate pattern <{a}> in the embodiment.
[0024] Figure 2 This is a schematic diagram of the positional utility probability chain of the candidate pattern <{ab}> in the embodiment.
[0025] Figure 3 This is a schematic diagram of the positional utility probability chain of the candidate pattern <{a}{b}> in the embodiment.
[0026] Figure 4 In the example, the candidate pattern <{a A schematic diagram of the positional utility probability chain.
[0027] Figure 5 The candidate pattern in the example is <{a}{ A schematic diagram of the positional utility probability chain of}{c}>.
[0028] Figure 6 As an example, and in the UUC-NSPM dataset DS1, the minimum threshold scaling factor is... Set to 5.0×10 -4 ~6.0×10 -4 The test results were compared with and without pruning.
[0029] Figure 7 As an example, and in the UUC-NSPM dataset DS2, the minimum threshold scaling factor is... Set to 6.0×10 -4 ~7.0×10 -4 The test results were compared with and without pruning.
[0030] Figure 8 As an example, and in the UUC-NSPM dataset DS3, the minimum threshold scaling factor is... Set to 7.0×10 -4 ~8.0×10 -4 The test results were compared with those without pruning.
[0031] Figure 9 As an example, and in the UUC-NSPM dataset DS4, the minimum threshold scaling factor is... Set to 7.0×10 -4 ~8.0×10 -4 The test results were compared with and without pruning.
[0032] Figure 10 As an example, and in the UUC-NSPM dataset DS1, the minimum threshold scaling factor is... Set to 5.0×10-4 ~6.0×10 -4 The test results were compared between the number of candidate patterns generated with and without pruning.
[0033] Figure 11 As an example, and in the UUC-NSPM dataset DS2, the minimum threshold scaling factor is... Set to 6.0×10 -4 ~7.0×10 -4 The test results were compared between the number of candidate patterns generated with and without pruning.
[0034] Figure 12 As an example, and in the UUC-NSPM dataset DS3, the minimum threshold scaling factor is... Set to 7.0×10 -4 ~8.0×10 -4 The test results were compared between the number of candidate patterns generated with and without pruning.
[0035] Figure 13 As an example, and in the UUC-NSPM dataset DS4, the minimum threshold scaling factor is... Set to 7.0×10 -4 ~8.0×10 -4 The test results were compared between the number of candidate patterns generated with and without pruning. Detailed Implementation
[0036] To make the objectives, technical solutions, and advantages of this invention clearer, the method of this invention will be further described in detail below with reference to a specific commodity transaction scenario. It should be understood that the specific embodiments described herein are only for explaining this invention and are not intended to limit this invention.
[0037] System operating environment and data preparation phase This embodiment provides a product recommendation system based on uncertain efficient continuous negative sequence mining, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the following steps of the product recommendation method: Step S1: Construct a sequence information table containing probability attributes.
[0038] Specifically, the processor collects historical shopping data from e-commerce platforms, including user IDs and corresponding shopping times, product names, product prices, purchase quantities, and probabilities.
[0039] The product names are coded to generate unique product codes. For example, the product "ballpoint pen" is coded as 'a', the product "notebook" is coded as 'b', and the product "briefcase" is coded as 'c', etc.; the corresponding prices for "ballpoint pen", "notebook", and "briefcase" are 5 yuan, 10 yuan, and 100 yuan, respectively.
[0040] Subsequently, the utility and residual utility of each item are calculated. The utility is the product of the price of the good and the quantity purchased in a single item (for example, if the price of good a is 5 yuan and the quantity purchased is 2, then its utility is 10). The residual utility is the sum of the utilities of all items that follow this item in the same itemset.
[0041] Finally, a sequence information table is constructed, containing several user IDs and the corresponding uncertain transaction sequence for each user ID. The uncertain transaction sequence includes several itemsets arranged in chronological order. Each itemset contains at least one item purchased at the same timestamp (e.g., a single checkout). Each item includes a product code and its corresponding item utility, purchase quantity, probability, and remaining utility. The probability represents the likelihood or confidence level of a user purchasing the product (e.g., due to data uncertainty or privacy anonymization, the probability of a user purchasing a ballpoint pen at a certain moment is 0.8).
[0042] Table 1 Sequence Information Table Step S2: Extract the extended items.
[0043] Based on the constructed sequence information table, the processor first mines all positive candidate patterns of length 1 (i.e., single item items, such as <{a}><{b}><{c}>). Then, it counts the frequency or utility value of these positive candidate patterns in all user sequences, and selects the top K with the highest frequency or utility value as positive expansion items. In this embodiment, K=3.
[0044] Next, the first K positive expansion terms are each negatively transformed (for example, transforming commodity a into non-commodity a, denoted as...). ), resulting in K negative extension terms.
[0045] Step S3: Initialize recursive expansion.
[0046] The processor uses the positive candidate patterns of length 1 as the first-generation candidate patterns and establishes a positional utility probability chain for each candidate pattern (e.g., Figure 1 (As shown). This positional utility probability chain records the user ID corresponding to the candidate pattern, the end position of the itemset in the sequence data, the pattern utility, and the pattern probability.
[0047] Define the recursive extension generation r, and initialize r=1.
[0048] Step S4: Continuous term expansion and positional utility probability chain update.
[0049] The processor enters a recursive expansion process, using a single r-th generation candidate pattern as an expansion prefix in sequence. Based on the sequence information table and the position utility probability chain corresponding to the r-th generation candidate pattern, it performs continuous term expansion using positive and negative expansion terms to obtain several expanded candidate patterns.
[0050] The specific expansion process is as follows: The user ID and the position of the ending itemset of the r-th generation candidate pattern are determined based on the position utility probability chain.
[0051] Positive expansion items include inner positive expansion items and outer positive expansion items, which are determined based on the sequence information table. The inner positive expansion item is the item in the itemset at the end of the itemset position in the sequence information table for the r-th generation candidate pattern; the outer positive expansion item is the item in the itemset next to the end of the itemset position in the sequence information table for the r-th generation candidate pattern.
[0052] When performing positive item expansion, the positive expansion item is added to the r-th generation candidate pattern to generate the expanded candidate pattern; the positive item expansion includes positive item set in-set expansion and positive item set out-of-set expansion.
[0053] The positive itemset expansion refers to adding an internal positive expansion item to the itemset at the end itemet position of the r-th generation candidate pattern to form a positive itemset expansion pattern. For example, in the first generation candidate pattern <{a}>, the end itemet position is 1 in user 2 in Table 1. For itemset expansion, there is a positive expansion item b in this itemset that can be expanded, so the expanded candidate pattern <{ab}> can be generated.
[0054] The positive itemset expansion refers to adding the positive expansion item as a new itemset to the end of the r-th generation candidate pattern, forming a positive itemset expansion pattern. For example, if the next itemset has a positive expansion item c that can be expanded, then the expanded candidate pattern <{a}{c}> can be generated.
[0055] When performing negative item expansion, the negative expansion item is added to the r-th generation candidate pattern to generate the expanded candidate pattern; the negative item expansion includes negative item set in-set expansion and negative item set out-of-set expansion.
[0056] The negative item in-itemset expansion refers to directly adding the negative expanded item to the itemset at the end-itemset position of the r-th generation candidate pattern. For example, in the first-generation candidate pattern <{a}>, in user 4 of Table 1, the end-itemset position is 1. For in-itemset expansion, and Both can be used as negative expansion terms to generate expanded candidate patterns <{a }> and <{a }>, but, <{a In}> There is a corresponding item in the first itemset of user 4, <{a If}> then there is none.
[0057] The negative item set expansion refers to adding the negative expansion item as a new itemset to the end of the r-th generation candidate pattern, and determining an accompanying positive expansion item based on the sequence information table, adding this accompanying positive expansion item as a new itemset after the negative expansion item. The method for determining the accompanying positive expansion item is as follows: the item items in the first or second itemset after the end itemset position of the r-th generation candidate pattern in the sequence information table are taken as the accompanying positive expansion item. For example, in user 4 in Table 1, the first itemset after the end itemset position of the r-th generation candidate pattern in the sequence information table is taken as the accompanying positive expansion item, that is, the accompanying positive expansion item is b. , and Both can be used as negative expansion terms to generate expanded candidate patterns <{a}{ }{b}>、<{a}{ }{b}> and <{a}{ }{b}>, , and There are no corresponding product items; the second itemset after the end itemset position in the sequence information table of the r-th generation candidate pattern is selected as the accompanying positive expansion item, that is, the accompanying positive expansion item is c. , and Both can be used as negative expansion terms to generate expanded candidate patterns <{a}{ }{b}>、<{a}{ }{c}> and <{a}{ }{c}>,<{a}{ In}{b}> There is a corresponding item in the second itemset of user 4, but and Then no.
[0058] Step S5: Calculate probabilistic utility and determine the high-utilization pattern.
[0059] For each expanded candidate pattern, the processor synchronously updates the positional utility probability chain and calculates its probability utility pu(candidate). The specific method is as follows: The utility in this positional utility probability chain is the sum of the utility u(α) of the current candidate pattern at the current recursive expansion generation r and the utility of all expansion terms; if the expansion term is a positive expansion term, then the utility u(x) of the positive expansion term is... pThe original utility of the corresponding item in the sequence information table is denoted as u(x). If the extended item is a negative extended item, the utility of the negative extended item is u(x). n ) is 0; where x p Indicates the positive expansion term, x n This indicates a negative extension term.
[0060] The probability in this positional utility probability chain is the product of the probability pr(α) of the current candidate mode at the current recursive expansion generation r and the actual probabilities of all expansion terms; if the expansion term is a positive expansion term, then the probability pr(x) of the positive expansion term is... p The probability pr(x) is the original probability of the corresponding item in the sequence information table; if the extended item is a negative extended item and a corresponding item exists at the corresponding position in the sequence information table, then the probability pr(x) of the negative extended item is... n ) is the complement of the original probability of the corresponding item (obtained by subtracting 1 from the original probability of the corresponding item); if the extended term is a negative extended term and the corresponding item does not exist at the corresponding position in the sequence information table, then the probability pr(x) of the negative extended term is... n The value is 1.
[0061] For example, using the first-generation candidate pattern <{a}> as a prefix, the following expansion is performed: By expanding the positive itemset, the positional utility probability chain of the expanded candidate pattern <{ab}> is obtained as follows: For user 1, position 1, utility 10+10=20, probability 0.8×0.9=0.72; for user 2, position 1, utility 5+20=25, probability 0.4×0.7=0.28; for user 4, position 1, utility 10+10=20, probability 0.4×0.5=0.2. The corresponding utility probability chains are as follows: Figure 2 As shown.
[0062] By expanding the positive itemset, the positional utility probability chain of the expanded candidate pattern <{a}{b}> is obtained as follows: for user 3, position 2 has a utility of 10+10=20 and a probability of 0.6×0.7=0.42; for user 4, position 2 has a utility of 10+20=30 and a probability of 0.4×0.8=0.32. The corresponding utility probability chains are as follows: Figure 3 As shown.
[0063] By expanding the negative itemset, we obtain the expanded candidate pattern <{a The positional utility probability chain is as follows: For user 1, position 1, utility 10+0=10, probability 0.8×(1-0.9)=0.08; for user 2, position 1, utility 5+0=5, probability 0.7×(1-0.4)=0.42; for user 3, position 1, utility 10+0=10, probability 0.6×1=0.6; for user 4, position 1, utility 10+0=10, probability 0.4×(1-0.5)=0.2; and for position 4, utility 10+0=10, probability 0.8×1=0.8. The corresponding utility probability chain is as follows: Figure 4 As shown.
[0064] By expanding the negative itemset, we obtain the expanded candidate pattern <{a}{ The positional utility probability chain for user 2 is: position 2, utility 5+0+100=105, probability 0.7×1×0.8=0.56; position 3 for user 4, utility 10+0+200=210, probability 0.4×(1-0.8)×0.7=0.056; the corresponding utility probability chain is as follows: Figure 5 As shown.
[0065] Calculate the probabilistic utility of the expanded candidate pattern under each user ID, where the probabilistic utility is the maximum value of the "utility-probability product" in the positional utility probability chain; sum the probabilistic utilities under all user IDs to obtain the global probabilistic utility pu(candidate) of the expanded candidate pattern.
[0066] Simultaneously, the processor calculates the probability utility bound (Bound(α,x,ρ,Q)) for each expanded candidate pattern, where α represents a single candidate pattern in the current recursive expansion generation r, x represents the expansion item currently participating in the expansion, ρ represents the itemset position of the current expansion item in the user sequence, and Q represents the user ID. A minimum utility threshold min_util is defined; if pu(candidate) ≥ min_util, the expanded candidate pattern is determined to be an uncertain high-utility continuous negative sequence pattern and is retained.
[0067] The method for determining the minimum utility threshold min_util is as follows: Define the sequence information table as dataset D, calculate the total item utility u(D) of all items in dataset D, and define the minimum threshold scaling factor. The formula min_util= The minimum utility threshold min_util is calculated using the method of × u(D).
[0068] Step S6: Pruning and Iteration Control.
[0069] To reduce computational complexity, the processor performs pruning. It iterates through all user IDs and determines the maximum probability utility upper bound maxBound(Q) for each user ID. maxBound(Q) is the maximum value among all probability utility upper bounds Bound(α,x,ρ,Q) for a single user ID.
[0070] For both positive itemset in-set expansion and positive itemset out-of-set expansion, the formula for calculating the probability utility upper bound Bound(α,x,ρ,Q) is as follows: Bound(α,x,ρ,Q)=(u(α)+u(x p )+ru(Q\(x p ,ρ)))×(pr(α)×pr(x p )); For positive itemset expansion, the formula for calculating the probability utility upper bound Bound(α,x,ρ,Q) is as follows: Bound(α,x,ρ,Q)=(u(α)+u(x n )+ru(Q\(x n ,ρ)))×(pr(α)×pr(x n )); For negative itemset expansion, the formula for calculating the probability utility upper bound Bound(α,x,ρ,Q) is as follows: Bound(α,x,ρ,Q)=(u(α)+u(x n )+u(x p )+ru(Q\(x p ,ρ)))×(pr(α)×pr(x n )×pr(x p )); in, ru(Q\(x p ,ρ)) represents the residual utility of the positive expansion item at the ρth item set of user Q in the sequence information table; ru(Q\(x n ,ρ)) represents the residual utility of the item corresponding to the negative extension item at the ρth item set of user Q in the sequence information table. If the negative extension item at the ρth item set of user Q in the sequence information table has no corresponding item, then ru(Q\(x n ,ρ)) represents the residual utility of the last item in the ρth item set of user Q in the sequence information table; u(x p ), u(x n ), pr(x p ) and pr(x n Determine according to step S5.
[0071] Taking Table 1 as an example, the first-generation candidate pattern <{a}> can be subjected to positive and negative item extension in user 4: Through intra-positive-item-set extension, the extended candidate pattern <{ab}> is obtained, the extension item is b, and Bound(<{a}>,b,1,4)=(10+10+235)×(0.4×0.5); Through extra-positive-item-set extension, the extended candidate pattern <{a}{b}> is obtained, the extension item is b, and Bound(<{a}>,{b},2,4)=(10+20+215)×(0.4×0.8); Through intra-negative-item-set extension, the extended candidate pattern <{a }>, the extension item is , Bound(<{a}>, ,1,4)=(10+0+235)×(0.4×(1-0.5)); Through extra-negative-item-set extension, the extended candidate pattern <{a}{ }{c}>, the extension item is and c, Bound(<{a}>,{ }{c},3,4)=(10+0+200+15)×(0.4×(1-0.8)×0.7).
[0072] Subsequently, the comparison value T is defined as the sum of maxBound(Q) corresponding to all user IDs. If T≥min_util, it indicates that this extension branch may have a high-utility pattern, it is determined to continue recursive extension, the extended candidate pattern is taken as the (r+1)-th generation candidate pattern, then set r=r+1, and return to execute step S4; if T<min_util, it indicates that this branch cannot generate a high-utility pattern, a pruning operation is performed to prune the current extended candidate pattern, and the recursive extension of this branch is terminated.
[0073] When the recursion of all branches ends, the processor outputs all mined uncertain high-utility sequential negative sequence patterns.
[0074] Step S7: Adjustment and display of commodity recommendation priority.
[0075] Finally, according to the mined uncertain high-utility sequential negative sequence patterns, the processor takes the probability utility pu(candidate) of the patterns calculated in step S5 as the actual utility value, and dynamically adjusts the commodity recommendation priority according to the rule that patterns with larger actual utility values have higher priority. For example, if the mined pattern <{a }> has an extremely high utility value, it indicates that the user group that purchases commodity a and does not purchase commodity b has strong potential value.
[0076] Based on the adjusted product recommendation priority, the processor displays recommended products to the current target user through the platform's front-end interface. The displayed content includes product name, image, unit price, and specifications, thereby achieving personalized, accurate, and realistic product recommendations that conform to complex real-world shopping logic (including negative sequence logic such as "not purchasing certain products").
[0077] The test data for this embodiment are as follows: The dataset for this example comes from the following URL: https: / / www.philippe-fournier-viger.com / spmf / index.php?link=datasets.php The features of the dataset in this embodiment are summarized in Table 2, where, This represents the total number of sequences in the database. The numbers represent different numbers of items, MAX(S) represents the maximum sequence length, Avg(S) represents the average sequence length, and Avg(IS) represents the average number of itemsets in each sequence.
[0078] Table 2 Dataset Description Table To verify the effectiveness of the recommended method (hereinafter referred to as Our) described in this embodiment, an existing method and an ablation experiment comparison method were selected and tested against the method of this embodiment based on the DS1-4 dataset. Among them, the existing method is UUCPM (from the paper "Towards utility-driven contiguous sequential patterns in uncertain multi-sequences"), which only mines uncertain utility-driven contiguous positive sequence patterns and only outputs positive feedback data. The ablation experiment comparison method is called UUC-NSPM in this embodiment. The difference between this method and the method of this embodiment is that the pruning step is deleted.
[0079] The method described in this embodiment, as well as UUCPM in dataset DS1-4, respectively set the minimum threshold ratio coefficient. To avoid redundancy, the maximum candidate pattern expansion length in the pruning experiment was set to 6. A minimum threshold for negative terms was also set; for example, DS15 indicates that in the experiment on dataset DS1, the top 5 positive sequence patterns with the highest probability utility values (length 1) are reversed and used as negative expansion terms of length 1. This setting ensures that the experiment can be completed within an acceptable time, while also guaranteeing the fairness of the comparative experiment and the validity of the conclusions. The main comparison is between the number of generated candidate patterns and the number of patterns in the positive sequence pattern mining, visually demonstrating the mining performance of the method in this embodiment. The test results are shown in Table 3. Table 3 Mining performance test data As shown in Table 3, the method provided in this embodiment can simultaneously mine both positive and negative sequence patterns; the mined positive sequence patterns are completely consistent with the output of the UUCPM algorithm. The experimental results in Table 3 show that the method in this embodiment can effectively mine a rich number of negative sequence candidate patterns and negative sequence patterns, and the results increase with the minimum threshold ratio coefficient. As the number of negative sequence candidate patterns and negative sequence patterns gradually increases, the overall trend shows a gradual decrease.
[0080] Figures 6-13 In order to avoid redundant expansion of the method in the ablation experiment, the maximum sequence expansion length in the experiment was set to 6, and a minimum threshold for negative terms was set. The top 5 positive sequence patterns with the highest probability utility value and a length of 1 were reversed and used as negative expansion terms with a length of 1.
[0081] The method described in this embodiment, along with the UUC-NSPM method, in the DS1 dataset, will use the minimum threshold scaling factor. Set to 5.0×10 -4 ~6.0×10 -4 The test results show that the runtime was compared with and without pruning. Figure 6 As shown; The method described in this embodiment, along with the UUC-NSPM method, in the DS2 dataset, will use the minimum threshold scaling factor. Set to 6.0×10 -4 ~7.0×10 -4 The test results show that the runtime was compared with and without pruning. Figure 7 As shown; The method described in this embodiment, along with the UUC-NSPM method, in the DS3 dataset, will use the minimum threshold scaling factor. Set to 7.0×10 -4 ~8.0×10 -4The test results show that the runtime was compared with and without pruning. Figure 8 As shown; The method described in this embodiment, along with the UUC-NSPM method, in the DS4 dataset, will use the minimum threshold scaling factor. Set to 7.0×10 -4 ~8.0×10 -4 The test results show that the runtime was compared with and without pruning. Figure 9 As shown; The method described in this embodiment, along with the UUC-NSPM method, in the DS1 dataset, will use the minimum threshold scaling factor. Set to 5.0×10 -4 ~6.0×10 -4 The test results show that the number of candidate patterns generated with and without pruning is as follows: Figure 10 As shown; The method described in this embodiment, along with the UUC-NSPM method, in the DS2 dataset, will use the minimum threshold scaling factor. Set to 6.0×10 -4 ~7.0×10 -4 The test results show that the number of candidate patterns generated with and without pruning is as follows: Figure 11 As shown; The method described in this embodiment, along with the UUC-NSPM method, in the DS3 dataset, will use the minimum threshold scaling factor. Set to 7.0×10 -4 ~8.0×10 -4 The test results show that the number of candidate patterns generated with and without pruning is as follows: Figure 12 As shown; The method described in this embodiment, along with the UUC-NSPM method, in the DS4 dataset, will use the minimum threshold scaling factor. Set to 7.0×10 -4 ~8.0×10 -4 The test results show that the number of candidate patterns generated with and without pruning is as follows: Figure 13 As shown; Figures 6-13 In this context, Our represents the method described in this embodiment, and UUC-NSPM represents the UUC-NSPM comparison method for ablation experiments mentioned earlier. This represents the minimum threshold scaling factor. Figures 6-9 The article compares the mining time of the two methods on the DS1-4 dataset under the same minimum threshold scaling factor. Figures 10-13 The paper compares the number of candidate patterns generated by the two methods during the mining process on the DS1-4 dataset under the same minimum threshold ratio coefficient.
[0082] from Figures 6 to 9 It can be seen from this: With the minimum threshold scaling factor Furthermore, the time required for the method described in this embodiment to mine uncertain and efficient continuous negative sequence patterns based on four datasets shows a significant decreasing trend, and the overall running time of the method in this embodiment is significantly reduced compared to the comparative methods, effectively proving that the pruning strategy of the method in this embodiment can greatly improve the overall efficiency of the algorithm.
[0083] from Figure 6 It can be seen that: when the method described in this embodiment is based on the DS1 dataset for mining, the minimum threshold ratio coefficient increases. From 5.0×10 -4 Increase in value: 6.0 × 10 -4 In this embodiment, the running time of the method is reduced from 2.7s to 1.8s, and the running time of the UUC-NSPM method is reduced from 188s to 116s. This embodiment reduces the running time by two orders of magnitude compared to the comparison algorithm.
[0084] from Figure 7 It can be seen that: when the method described in this embodiment is based on the DS2 dataset for mining, the minimum threshold ratio coefficient increases. From 6.0×10 -4 Increase in value: 7.0 × 10 -4 In this embodiment, the running time of the method is reduced from 3.86s to 2.66s, and the running time of the UUC-NSPM method is reduced from 239s to 142s. This embodiment reduces the running time by two orders of magnitude compared to the comparison algorithm.
[0085] from Figure 8 It can be seen that: when the method described in this embodiment is based on the DS3 dataset for mining, the minimum threshold ratio coefficient increases. From 7.0×10 -4 Increase in value: 8.0 × 10 -4 In this embodiment, the running time of the method is reduced from 4.76s to 3.54s, and the running time of the UUC-NSPM method is reduced from 336s to 279s. This embodiment reduces the running time by two orders of magnitude compared to the comparison algorithm.
[0086] from Figure 9 It can be seen that: when the method described in this embodiment is based on the DS4 dataset for mining, the minimum threshold ratio coefficient increases. From 7.0×10 -4 Increase in value: 8.0 × 10 -4 In this embodiment, the running time of the method is reduced from 7.42s to 6.31s, and the running time of the UUC-NSPM method is reduced from 547s to 390s. This embodiment reduces the running time by two orders of magnitude compared to the comparison algorithm.
[0087] from Figures 10 to 13 It can be seen that: With the minimum threshold scaling factor With the increase of the number of data points, the number of candidate patterns for uncertain and efficient continuous negative sequences mined by the method described in this embodiment based on four datasets shows a significant decreasing trend. Moreover, the overall number of candidate patterns mined by the method in this embodiment is significantly reduced compared to the comparative method, which effectively proves that the pruning strategy of the method in this embodiment can remove a large number of hopeless candidate patterns in advance.
[0088] from Figure 10 It can be seen that: when the method described in this embodiment is based on the DS1 dataset for mining, the minimum threshold ratio coefficient increases. From 5.0×10 -4 Increase in value: 6.0 × 10 -4 In this embodiment, the number of uncertain efficient continuous negative sequence pattern candidates mined by the method described above is reduced from 8132 to 7139, and the number of uncertain efficient continuous negative sequence pattern candidates mined by the UUC-NSPM method is reduced from 10000159 to 6864843. This embodiment reduces the number of candidate patterns by about three orders of magnitude compared to the comparison algorithm.
[0089] from Figure 11 It can be seen that: when the method described in this embodiment is based on the DS2 dataset for mining, the minimum threshold ratio coefficient increases. From 6.0×10 -4 Increase in value: 7.0 × 10 -4 In this embodiment, the number of candidate uncertain efficient continuous negative sequence patterns mined by the method is reduced from 6994 to 6605, and the number of candidate uncertain efficient continuous negative sequence patterns mined by the UUC-NSPM method is reduced from 13,517,728 to 7,915,081. This embodiment reduces the number of candidate patterns by about three orders of magnitude compared to the comparison algorithm.
[0090] from Figure 12 It can be seen that: when the method described in this embodiment is based on the DS3 dataset for mining, the minimum threshold ratio coefficient increases. From 7.0×10 -4 Increase in value: 8.0 × 10 -4 In this embodiment, the number of uncertain efficient continuous negative sequence pattern candidates mined by the method described above is reduced from 6636 to 6470, and the number of uncertain efficient continuous negative sequence pattern candidates mined by the UUC-NSPM method is reduced from 17784991 to 6864843. This embodiment reduces the number of candidate patterns by about three orders of magnitude compared to the comparison algorithm.
[0091] from Figure 13 It can be seen that: when the method described in this embodiment is based on the DS4 dataset for mining, the minimum threshold ratio coefficient increases. From 7.0×10 -4 Increase in value: 8.0 × 10 -4 In this embodiment, the number of candidate uncertain efficient continuous negative sequence patterns mined by the method is reduced from 6661 to 6483, and the number of candidate uncertain efficient continuous negative sequence patterns mined by the UUC-NSPM method is reduced from 26018458 to 18674659. This embodiment reduces the number of candidate patterns by about three orders of magnitude compared to the comparison algorithm.
[0092] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.
Claims
1. A commodity recommendation method based on uncertain high-utility continuous sequence mining, characterized in that, comprising the following steps: Step S1: constructing a sequence information table comprising probability attributes, wherein the sequence information table comprises a plurality of user IDs and an uncertain transaction sequence corresponding to each user ID, the uncertain transaction sequence comprises a plurality of item sets arranged in chronological order, each item set comprises at least one item purchased under the same timestamp, and each item comprises a commodity code, and the utility, purchase quantity, probability and remaining utility corresponding to the commodity code; Step S2: mining all positive candidate patterns with a length of 1 based on the sequence information table, sorting the positive candidate patterns in descending order of probability utility values and selecting the top K, respectively performing negative transformation on the top K positive candidate patterns to obtain K negative extension items; Step S3: taking the positive candidate patterns as the 1st generation candidate patterns, establishing a position utility probability chain corresponding to the 1st generation candidate patterns, defining a recursive extension generation number r, and initializing r=1; Step S4: performing recursive extension, taking a single r-th generation candidate pattern as an extension prefix in sequence, performing continuous item extension by using positive extension items and negative extension items respectively based on the sequence information table and the position utility probability chain corresponding to the r-th generation candidate pattern to obtain a plurality of extended candidate patterns, and updating the position utility probability chain for each extended candidate pattern; wherein the continuous item extension is used for ensuring the time continuity of the extended pattern in the uncertain transaction sequence; Step S5: calculating the probability utility pu(candidate) and the probability utility upper bound Bound(α,x,ρ,Q) of each extended candidate pattern, wherein α represents a single candidate pattern under the current recursive extension generation number r, x represents the extension item currently participating in extension, ρ represents the item set position of the extension item currently participating in extension in the user sequence, and Q represents the user ID; defining a minimum utility threshold min_util, and if pu(candidate)≥min_util, determining that the extended candidate pattern is an uncertain high-utility continuous negative sequence pattern; Step S6: traversing all user IDs, letting Q denote the user ID, determining the maximum value maxBound(Q) of the probability utility upper bound corresponding to each user ID, wherein the maxBound(Q) is the maximum value among all probability utility upper bounds Bound(α,x,ρ,Q) under a single user ID; defining a comparison value T as the sum of maxBound(Q) corresponding to all user IDs, if T≥min_util, determining to continue the recursive extension, taking the extended candidate patterns as the (r+1)-th generation candidate patterns, then letting r=r+1, returning to execute step S4; if T<min_util, performing a pruning operation to prune the current extended candidate patterns, terminating the recursive extension, and outputting all mined uncertain high-utility continuous negative sequence patterns; Step S7: Based on the mined uncertain high-utility continuous negative sequence pattern, the probability utility calculated in step S5 for the uncertain high-utility continuous negative sequence pattern is used as the actual utility value. The product recommendation priority is adjusted according to the rule that the pattern with the larger actual utility value has higher priority, and product recommendations are made according to the adjusted product recommendation priority. 2.The commodity recommendation method based on uncertain high-utility continuous negative sequence mining according to claim 1, wherein, The process of constructing the sequence information table includes: Collect shopping data, which includes user ID and the shopping time, product name, product price, purchase quantity, and probability corresponding to the user ID; The product name is encoded to generate a unique product code; Calculate the utility and residual utility of each item, where the utility is the product of the price and quantity of the good in a single item, and the residual utility is the sum of the utilities of all subsequent items.
3. The product recommendation method based on uncertain and efficient mining of continuous negative sequences according to claim 1, characterized in that, The method for determining the minimum utility threshold min_util is: defining a sequence information table as a data set D, calculating the item utility sum u(D) of all items in the data set D, defining a minimum threshold proportion coefficient , and calculating the minimum utility threshold min_util through the formula min_util= × u(D).
4. The product recommendation method based on uncertain and efficient continuous negative sequence mining according to claim 1, characterized in that, The process of establishing the location utility probability chain includes: For each candidate pattern, record its corresponding user ID, the end position of the itemset in the sequence data, the pattern utility, and the pattern probability.
5. The product recommendation method based on uncertain and efficient mining of continuous negative sequences according to claim 1, characterized in that, The specific process of expanding consecutive terms in step S4 is as follows: The user ID and the position of the ending itemset of the r-th generation candidate pattern are determined based on the position utility probability chain. The positive expansion item includes an inner positive expansion item and an outer positive expansion item, both of which are determined based on the sequence information table; the inner positive expansion item is the item in the itemset at the end of the itemset position in the sequence information table for the r-th generation candidate pattern; The external positive expansion term is the term in the next itemset of the r-th generation candidate pattern at the position of the end itemset in the sequence information table; When performing positive item expansion, positive expansion items are added to the r-th generation candidate patterns to generate expanded candidate patterns; the positive item expansion includes positive item set in-set expansion and positive item set out-of-set expansion. The positive item set expansion refers to adding the positive expansion item to the itemset at the end of the candidate pattern in the rth generation, thus forming a positive item set expansion pattern. The positive itemset expansion refers to adding the positive expansion itemset as a new itemset to the end of the r-th generation candidate pattern to form the positive itemset expansion pattern.
6. The product recommendation method based on uncertain and efficient mining of continuous negative sequences according to claim 5, characterized in that, The specific process of expanding consecutive terms in step S4 also includes: When performing negative item expansion, the negative expansion item is added to the r-th generation candidate pattern to generate the expanded candidate pattern; the negative item expansion includes negative item set expansion and negative item set expansion. The negative item in-item set expansion refers to directly adding the negative expansion item to the itemset at the end of the candidate pattern in the rth generation; The negative item set expansion refers to adding the negative expansion item as a new itemset to the end of the r-th generation candidate pattern, and determining an accompanying positive expansion item according to the sequence information table, and adding the accompanying positive expansion item as a new itemset after the negative expansion item. The method for determining the accompanying positive expansion item is: taking the item items in the first or second itemset after the end itemset position of the r-th generation candidate pattern in the sequence information table as the accompanying positive expansion item.
7. The product recommendation method based on uncertain and efficient mining of continuous negative sequences according to claim 6, characterized in that, The specific methods for updating the position utility probability chain and calculating the probability utility pu(candidate) are as follows: For each expanded candidate pattern, a position utility probability chain is established. The utility in the position utility probability chain is the utility u(a) of the current candidate pattern under the current recursive expansion round r plus the utility of all expansion items; if the expansion item is a positive expansion item, the utility u(x p ) of the positive expansion item is the original utility of the corresponding commodity item in the sequence information table, if the expansion item is a negative expansion item, the utility u(x n ) of the negative expansion item is 0; The probability in the position utility probability chain is the product of the probability pr(α) of the current candidate pattern under the current recursive expansion generation r and the actual probabilities of all expansion items; If the extension item is a positive extension item, the actual probability pr(x p ) of the positive extension item is the original probability of the corresponding item in the sequence information table; if the extension item is a negative extension item and the corresponding item exists at the corresponding position in the sequence information table, the actual probability pr(x n ) of the negative extension item is the complement of the original probability of the corresponding item; if the extension item is a negative extension item and the corresponding item does not exist at the corresponding position in the sequence information table, the actual probability pr(x n ) of the negative extension item is 1. Calculate the probability utility of the expanded candidate pattern under each user ID, where the probability utility is the maximum value of the product of utility and probability in the position utility probability chain; sum the probability utilities under all user IDs to obtain the probability utility pu(candidate) of the expanded candidate pattern.
8. The product recommendation method based on uncertain high-utility continuous negative sequence mining according to claim 7, characterized in that, For intra-positive itemset expansion and extra-positive itemset expansion, the calculation formula for the probability utility upper bound Bound(α,x,ρ,Q) is: Bound(a, x, p, Q) = (u(a) + u(x p ) + ru(Q\(x p , p))) x (pr(a) x pr(x p )); For intra-positive itemset expansion, the calculation formula for the probability utility upper bound Bound(α,x,ρ,Q) is: Bound(a, x, p, Q) = (u(a) + u(x n ) + ru(Q\(x n , p))) x (pr(a) x pr(x n )); For intra-negative itemset expansion, the calculation formula for the probability utility upper bound Bound(α,x,ρ,Q) is: Bound(α,x,ρ,Q)=(u(α)+u(x n )+u(x p )+ru(Q\(x p ,ρ)))×(pr(α)×pr(x n )×pr(x p )); in which, ru(Q\(x p ,ρ)) represents the residual utility of the positive expansion item at the ρth item set of user Q in the sequence information table; ru(Q\(x n , p)) represents the remaining utility of the corresponding item of the negative extension at the pth item set of user Q in the sequence information table, if there is no corresponding item of the negative extension at the pth item set of user Q in the sequence information table, ru(Q\(x n , p)) represents the remaining utility of the last item at the pth item set of user Q in the sequence information table. u(x p ), u(x n ), pr(x p ) and pr(x n ) are determined according to claim 7.
9. The product recommendation method based on uncertain and efficient mining of continuous negative sequences according to claim 1, characterized in that, The specific process of step S6 is as follows: For a single user Q, traverse the probability utility upper bounds Bound(α,x,ρ,Q) of all expanded candidate patterns, and select the maximum value among them as the maximum probability utility upper bound maxBound(Q) corresponding to the user Q; Sum the maxBound(Q) corresponding to all users Q to obtain the comparison value T; If T≥min_util, it is determined to continue the recursive expansion; if T<min_util, the pruning operation is performed to end the recursive expansion.
10. A product recommendation system based on uncertain and efficient mining of continuous negative sequences, characterized in that, Comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, wherein the processor implements the method according to any one of claims 1 to 9 when executing the program.
Citation Information
Patent Citations
Probability frequent item set excavating method based on MapReduce
CN103995882A
Commodity recommendation method and system based on target sequence pattern mining
CN121366023A
