A commodity recommendation method, device and medium

Abstract

The present application relates to the technical field of data processing, and particularly relates to a commodity recommendation method, device and medium, which comprises the following steps: constructing commodity co-occurrence mapping information based on normalized transaction information; generating vectorized representation result information based on the normalized transaction information; generating an associated candidate commodity set based on the commodity co-occurrence mapping information; generating a similar candidate commodity set based on the vectorized representation result information; selecting a main parameter combination from a preset parameter candidate set; fusing and sorting the associated candidate commodity set and the similar candidate commodity set based on the main parameter combination to obtain main recommendation result information; selecting a shadow parameter combination from the preset parameter candidate set; and fusing and sorting the associated candidate commodity set and the similar candidate commodity set based on the shadow parameter combination to obtain shadow recommendation result information. The present application has the effect of solving the problem that the existing commodity recommendation scheme is difficult to adaptively adjust the recommendation process in combination with behavior feedback.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a product recommendation method, apparatus, and medium. Background Technology

[0002] In e-commerce, membership retail and other business scenarios, the number of products and the frequency of updates are constantly increasing. The scale of historical transaction information and behavioral feedback information generated by members in browsing, purchasing and other processes is also growing. Platforms usually need to generate personalized product recommendation results for target members based on the above data in order to meet the differentiated needs of different members at different times.

[0003] Existing product recommendation solutions typically rely on historical transaction information to mine relationships between products, or on member-product interaction data to generate member and product representations, and then sort and output candidate product sets accordingly. However, in practical applications, the effectiveness of product relationships and interaction representations varies among different members, product categories, and time periods. If a fixed fusion method or fixed weights are used for candidate fusion, the recommendation results may become insensitive to changes in behavioral feedback, making it difficult to achieve continuous adaptive optimization. Summary of the Invention

[0004] To address the problem that existing product recommendation schemes struggle to adaptively adjust the recommendation process by incorporating behavioral feedback, this invention provides a product recommendation method, apparatus, and medium.

[0005] The above-mentioned objective of this application is achieved through the following technical solution:

[0006] A product recommendation method, the method comprising:

[0007] Historical transaction information and behavioral feedback information are obtained, and the historical transaction information and behavioral feedback information are standardized to obtain standardized transaction information and standardized feedback information.

[0008] Commodity co-occurrence mapping information is constructed based on standardized transaction information, and vectorized representation results are generated based on standardized transaction information;

[0009] After obtaining the recommendation request from the target member, a set of associated candidate products is generated based on the product co-occurrence mapping information, and a set of similar candidate products is generated based on the vectorized representation results.

[0010] Select a main parameter combination from the preset parameter candidate set, and merge and sort the associated candidate product set and the similar candidate product set based on the main parameter combination to obtain the main recommendation result information;

[0011] While outputting the main recommendation results, shadow parameter combinations are selected from the preset parameter candidate set. Based on the shadow parameter combinations, the related candidate product set and the similar candidate product set are fused and sorted to obtain the shadow recommendation results.

[0012] By adopting the above technical solutions, and by acquiring historical transaction information and behavioral feedback information and standardizing them respectively to obtain standardized transaction information and standardized feedback information, it is possible to maintain consistency in field names, field formats, and value ranges between historical transaction information and behavioral feedback information, thereby reducing the risk of field mismatch. By constructing product co-occurrence mapping information based on standardized transaction information and generating vectorized representation results, it is possible to simultaneously generate product co-occurrence mapping information and vectorized representation results under the same data caliber, thereby reducing the risk of data caliber discrepancies. After obtaining the recommendation request from the target member, by generating a set of associated candidate products based on product co-occurrence mapping information and a set of similar candidate products based on vectorized representation results, it is possible to form two categories of products from different sources after the recommendation request is triggered. The system generates a candidate set and reduces the risk of insufficient candidate coverage. It selects a main parameter combination from the preset parameter candidate set and merges and sorts the associated candidate product set and the similar candidate product set based on the main parameter combination to obtain the main recommendation result information. It can introduce configurable association recall weight, similarity recall weight and preset segmentation rule information in the fusion and sorting stage and reduce the risk of insufficient sorting adaptation. At the same time as outputting the main recommendation result information, it selects a shadow parameter combination from the preset parameter candidate set and merges and sorts the associated candidate product set and the similar candidate product set based on the shadow parameter combination to obtain shadow recommendation result information. It can form a comparison sorting result corresponding to the main recommendation result information under the same recommendation request and reduce the risk of verification difficulties caused by the lack of comparison basis for parameter adjustment.

[0013] In a preferred embodiment, this application can be further configured as follows: the construction of commodity co-occurrence mapping information based on normalized transaction information includes:

[0014] Extract incremental transaction records from standardized transaction information;

[0015] The incremental transaction records are summarized by sliding according to a preset sliding time window to obtain window statistics;

[0016] The window statistics are updated by weighting them using a preset time decay factor to obtain weighted statistics.

[0017] The total number of transactions in the sliding window is determined based on window statistics, and the support count threshold is determined based on the total number of transactions in the sliding window.

[0018] Low-frequency filtering is performed on the weighted statistics based on the support count threshold information to obtain weighted statistics for co-occurrence screening;

[0019] Based on the weighted statistical information used for co-occurrence screening, a set of mapping items is generated between the source product identification information combination and the target product identification information combination, and the set of mapping items is summarized to obtain the product co-occurrence mapping information.

[0020] By adopting the above technical solutions, it is possible to limit the statistical range based on incremental transaction records and make the window statistical information consistent with the preset sliding time window; to perform weighted updates on the window statistical information based on the preset time decay factor and make the contribution of different time windows to the weighted statistical information different; to determine the support count threshold information based on the total number of transactions in the sliding window and keep the threshold value related to the transaction scale; to complete low-frequency filtering based on the support count threshold information and retain statistical items that meet the counting requirements using the weighted statistical information used for co-occurrence screening; and to summarize the mapping item set corresponding to the source product identification information combination and the target product identification information combination to form product co-occurrence mapping information and reduce duplicate counting and duplicate summarization in the process of generating co-occurrence mapping information.

[0021] In a preferred embodiment, this application can be further configured such that: the generation of vectorized representation result information based on normalized transaction information includes:

[0022] Member identification information and product identification information are extracted from standardized transaction information, and member-product interaction relationship information is formed based on the correspondence between member identification information and product identification information;

[0023] Based on the member-product interaction relationship information, determine the identifier jump set information, and generate sequence information along the identifier jump set information according to the preset sequence generation rule information;

[0024] The training sample set information is constructed based on sequence information, and the parameters of the training sample set information are iteratively updated according to the preset vector training rules to obtain the vectorized representation result information.

[0025] By adopting the above technical solution, member identification information and product identification information can be extracted from standardized transaction information to form member-product interaction relationship information to maintain the consistency of the interaction correspondence structure. Based on the member-product interaction relationship information, the identification jump set information can be determined and the sequence information can be generated according to the preset sequence generation rule information to form a trainable identification sequence expression. Based on the sequence information, the training sample set information can be constructed and the parameters can be iteratively updated according to the preset vector training rule information to obtain the vectorized representation result information and keep the generation process of member representation vector and product representation vector consistent with the standardization of transaction information.

[0026] In a preferred embodiment, this application can be further configured such that both the main parameter combination and the shadow parameter combination are selected from a preset parameter candidate set, and the main parameter combination and the shadow parameter combination include at least the association recall weight, the similarity recall weight, and the preset segmentation rule information.

[0027] By adopting the above technical solution, the main parameter combination and the shadow parameter combination can be uniformly limited to the parameter space of the preset parameter candidate set and the parameter source boundary can be clearly defined. The composition structure of the main parameter combination and the shadow parameter combination can be constrained by the associated recall weight, similar recall weight and preset segmentation rule information, and the weight and rule reference relationship in the fusion ranking stage can be clearly defined. This can reduce the risk of inconsistent ranking calculation caused by parameter missing or rule missing in the fusion ranking stage of the main parameter combination and the shadow parameter combination.

[0028] In a preferred embodiment, this application can be further configured as follows: while outputting the main recommendation result information, selecting shadow parameter combinations from a preset parameter candidate set, and merging and sorting the associated candidate product set and the similar candidate product set based on the shadow parameter combinations to obtain shadow recommendation result information, including:

[0029] Based on the preset perturbation rule information, perturbation generation is performed on the associated recall weight and similar recall weight, and perturbation generation is performed on the preset segmentation rule information according to the preset perturbation rule information to obtain the shadow parameter combination;

[0030] The sets of related and similar candidate products are deduplicated and merged to obtain a merged candidate product set.

[0031] Iterate through the candidate products in the merged candidate product set, generate the association score and similarity score for each candidate product, and calculate the score difference information for each candidate product.

[0032] Candidate weight combinations are selected based on score difference information and preset segmentation rules. Based on the candidate weight combinations, the associated scores and similar scores are weighted and fused to obtain the fused scores corresponding to the candidate products.

[0033] The candidate product set is sorted according to the fusion score to obtain the shadow recommendation results.

[0034] By adopting the above technical solution, it is possible to generate shadow parameter combinations that differ from the main parameter combinations based on preset perturbation rule information, and form a reference parameter basis for shadow recommendation result information. It is possible to perform deduplication and merging of related candidate product sets and similar candidate product sets based on shadow parameter combinations to form a merged candidate product set. It is possible to select candidate weight combinations and generate fusion scores based on association scores, similarity scores, and score difference information. It is possible to sort the merged candidate product set according to the fusion scores and form shadow recommendation result information, and keep the shadow recommendation result information and the main recommendation result information comparable under the same candidate set and the same sorting process.

[0035] In a preferred embodiment, this application can be further configured as follows: [method / method], also including:

[0036] The main parameter combination, shadow parameter combination, main recommendation result information and shadow recommendation result information are associated and written into the sample information;

[0037] The feedback result information corresponding to the sample information is read from the normalized feedback information. The main recommendation result information and the shadow recommendation result information are compared and evaluated based on the feedback result information to obtain the reward value information. The fusion decision strategy is updated based on the reward value information to obtain the updated fusion decision strategy.

[0038] By adopting the above technical solution, it is possible to associate the main parameter combination, shadow parameter combination, main recommendation result information and shadow recommendation result information into the sample information and form a traceable recommendation request record. It is possible to read the feedback result information corresponding to the sample information from the normalized feedback information and establish a comparison relationship between the feedback result information and the main recommendation result information and the shadow recommendation result information. It is possible to generate reward value information based on the feedback result information and use the reward value information to update the fusion decision strategy and make the selection of the main parameter combination of subsequent recommendation requests form a closed loop association with the sample information.

[0039] In a preferred embodiment, this application can be further configured to: evaluate the main recommendation result information and the shadow recommendation result information based on the feedback result information to obtain reward value information, including;

[0040] Feedback scores are generated based on the feedback results, and main chain evaluation values ​​and shadow chain evaluation values ​​are generated separately. Comparison difference scores are generated based on the main chain evaluation values ​​and shadow chain evaluation values, and reward value information is generated based on the comparison difference scores.

[0041] By adopting the above technical solution, it is possible to generate feedback scores based on feedback results and make the feedback results participate in the calculation process of the main chain evaluation value and the shadow chain evaluation value. It is also possible to generate a comparison difference score based on the main chain evaluation value and the shadow chain evaluation value and make the comparison difference score reflect the difference between the main recommendation result information and the shadow recommendation result information under the feedback result information. Furthermore, it is possible to generate reward value information based on the comparison difference score and make the reward value information form a numerical expression consistent with the feedback result information.

[0042] In a preferred example, this application can be further configured to: update the fusion decision policy based on reward value information to obtain an updated fusion decision policy, including:

[0043] Based on the current value function value of the main chain and the current value function value of the shadow chain in the fusion decision strategy, and read the subsequent value function value of the main chain and the subsequent value function value of the shadow chain;

[0044] Based on the difference score, temperature parameters were read and then... Calculate the gating coefficient g(q);

[0045] Obtain the update coefficient and discount coefficient, and calculate the updated main chain value function value and the updated shadow chain value function value based on the reward value information;

[0046] The updated main chain value function value is written into the fusion decision strategy to update the main chain value function information, and the updated shadow chain value function value is written into the fusion decision strategy to update the shadow chain value function information, so as to obtain the updated fusion decision strategy.

[0047] By adopting the above technical solution, it is possible to read the current value function value of the main chain, the current value function value of the shadow chain, and the subsequent value function values ​​of the main chain and the shadow chain based on the fusion decision strategy, and to make the value function values ​​participate in the update calculation. It is possible to read the temperature parameter based on the comparison difference score and calculate the gate coefficient g(q) to form the update amplitude adjustment factor. It is possible to obtain the update coefficient and the discount coefficient and combine them with the reward value information to calculate the updated value function value of the main chain and the updated value function value of the shadow chain. It is possible to write the updated value function value of the main chain and the updated value function value of the shadow chain into the fusion decision strategy to update the value function information of the main chain and the value function information of the shadow chain and obtain the updated fusion decision strategy.

[0048] The second objective of this invention is achieved through the following technical solution:

[0049] A product recommendation device, characterized in that the product recommendation device comprises:

[0050] The standardization module is used to acquire historical transaction information and behavioral feedback information, and to standardize the historical transaction information and behavioral feedback information to obtain standardized transaction information and standardized feedback information.

[0051] The mapping and representation module is used to construct commodity co-occurrence mapping information based on normalized transaction information and generate vectorized representation results based on normalized transaction information.

[0052] The candidate generation module is used to generate a set of associated candidate products based on product co-occurrence mapping information after obtaining the recommendation request from the target member, and to generate a set of similar candidate products based on the vectorized representation result information.

[0053] The main sorting module is used to select main parameter combinations from the preset parameter candidate set, and to merge and sort the related candidate product set and the similar candidate product set based on the main parameter combinations to obtain the main recommendation result information;

[0054] The shadow ranking module is used to select shadow parameter combinations from a preset parameter candidate set while outputting the main recommendation result information. Based on the shadow parameter combinations, the related candidate product set and the similar candidate product set are merged and ranked to obtain the shadow recommendation result information.

[0055] By adopting the above technical solutions, and by acquiring historical transaction information and behavioral feedback information and standardizing them respectively to obtain standardized transaction information and standardized feedback information, it is possible to maintain consistency in field names, field formats, and value ranges between historical transaction information and behavioral feedback information, thereby reducing the risk of field mismatch. By constructing product co-occurrence mapping information based on standardized transaction information and generating vectorized representation results, it is possible to simultaneously generate product co-occurrence mapping information and vectorized representation results under the same data caliber, thereby reducing the risk of data caliber discrepancies. After obtaining the recommendation request from the target member, by generating a set of associated candidate products based on product co-occurrence mapping information and a set of similar candidate products based on vectorized representation results, it is possible to form two categories of products from different sources after the recommendation request is triggered. The system generates a candidate set and reduces the risk of insufficient candidate coverage. It selects a main parameter combination from the preset parameter candidate set and merges and sorts the associated candidate product set and the similar candidate product set based on the main parameter combination to obtain the main recommendation result information. It can introduce configurable association recall weight, similarity recall weight and preset segmentation rule information in the fusion and sorting stage and reduce the risk of insufficient sorting adaptation. At the same time as outputting the main recommendation result information, it selects a shadow parameter combination from the preset parameter candidate set and merges and sorts the associated candidate product set and the similar candidate product set based on the shadow parameter combination to obtain shadow recommendation result information. It can form a comparison sorting result corresponding to the main recommendation result information under the same recommendation request and reduce the risk of verification difficulties caused by the lack of comparison basis for parameter adjustment.

[0056] The above-mentioned objective three of this application is achieved through the following technical solution:

[0057] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the aforementioned product recommendation method.

[0058] In summary, this application includes at least one of the following beneficial technical effects:

[0059] By acquiring and standardizing historical transaction information and behavioral feedback information, we can obtain standardized transaction information and standardized feedback information. This ensures consistency in field names, formats, and value ranges between the two information, reducing the risk of field mismatch. By constructing a product co-occurrence mapping based on the standardized transaction information and generating a vectorized representation, we can simultaneously generate both product co-occurrence mapping and vectorized representation information under the same data caliber, reducing the risk of data caliber discrepancies. Furthermore, after receiving a target member's recommendation request, we can generate a set of associated candidate products based on the product co-occurrence mapping information and a set of similar candidate products based on the vectorized representation information. This allows us to form two candidate sets from different sources after a recommendation request is triggered. To mitigate the risk of insufficient candidate coverage, the system selects a main parameter combination from a preset parameter candidate set and then merges and sorts the associated candidate product set and the similar candidate product set based on the main parameter combination to obtain the main recommendation result information. This allows for the introduction of configurable association recall weights, similarity recall weights, and preset segmentation rules during the fusion and sorting stage, thereby reducing the risk of insufficient sorting adaptation. Simultaneously, by selecting a shadow parameter combination from the preset parameter candidate set while outputting the main recommendation result information, and then merging and sorting the associated candidate product set and the similar candidate product set based on the shadow parameter combination to obtain shadow recommendation result information, the system can generate a comparative sorting result corresponding to the main recommendation result information under the same recommendation request, thereby reducing the risk of verification difficulties caused by a lack of comparative basis for parameter adjustment. Attached Figure Description

[0060] Figure 1 This is a flowchart of a product recommendation method according to an embodiment of this application.

[0061] Figure 2 This is a schematic block diagram of a product recommendation device according to one embodiment of this application. Detailed Implementation

[0062] The present application will be further described in detail below with reference to the accompanying drawings.

[0063] In one embodiment, such as Figure 1 As shown, this application discloses a product recommendation method, which specifically includes the following steps:

[0064] S10: Obtain historical transaction information and behavioral feedback information, and standardize the historical transaction information and behavioral feedback information respectively to obtain standardized transaction information and standardized feedback information.

[0065] In this embodiment, historical transaction information refers to the record information of transaction behavior between member identification information and product identification information within a preset time range. The record information includes at least the transaction occurrence time information and transaction quantity information. Behavioral feedback information refers to the feedback record information formed when members browse, click, favorite, add to cart, purchase, etc. The feedback record information includes at least the feedback occurrence time information and feedback type identification information. Standardized transaction information refers to the transaction record information after the field names, field formats and value ranges in historical transaction information are unified. Standardized feedback information refers to the feedback record information after the field names, field formats and value ranges in behavioral feedback information are unified.

[0066] Specifically, when acquiring historical transaction information and behavioral feedback information, the system first reads transaction record entries containing member identification information and product identification information within a preset time range to form a historical transaction information set, and reads feedback record entries containing member identification information and feedback type identification information to form a behavioral feedback information set. Then, the historical transaction information set is aligned and formatted, converting transaction occurrence time information to a unified time format, transaction quantity information to a unified numerical format, and member identification information and product identification information to a unified encoding format. Missing field entries are filled using elimination or completion rules to maintain field integrity and consistency. Next, the behavioral feedback information set is aligned and formatted, converting feedback occurrence time information to a unified time format, mapping feedback type identification information to a preset feedback category code, and deduplicating and merging duplicate feedback record entries to avoid duplicate entries within the same time segment. Finally, the historical transaction information set after field alignment, format unification, and encoding unification is determined as standardized transaction information, and the behavioral feedback information set after field alignment, format unification, and encoding mapping is determined as standardized feedback information.

[0067] S20: Construct commodity co-occurrence mapping information based on normalized transaction information, and generate vectorized representation result information based on normalized transaction information.

[0068] In this embodiment, the product co-occurrence mapping information refers to the mapping record information formed based on the co-occurrence relationship of product identification information in the normalized transaction information within the same preset sliding time window. The mapping record information includes at least the source product identification information combination, the target product identification information combination, and co-occurrence statistical value information. The vectorized representation result information refers to the vector set information generated based on the interaction correspondence between member identification information and product identification information in the normalized transaction information. The vector set information includes at least the member representation vector and the product representation vector.

[0069] Specifically, when constructing product co-occurrence mapping information based on standardized transaction information, the transaction occurrence time information, member identification information, and product identification information are first read from the standardized transaction information. The transaction occurrence time information is then divided into multiple window intervals according to a preset sliding time window. Within each window interval, the product identification information that has appeared is summarized to form a window product set information. Then, co-occurrence counting is performed on the combinations of product identification information within the window product set information to form co-occurrence statistical value information. This co-occurrence statistical value information is then bound to the corresponding source product identification information combination and target product identification information combination to form mapping record entries. Finally, cross-window summarization is performed on the mapping record entries to form a product co-occurrence mapping information set. When generating vectorized representation result information based on standardized transaction information, member identification information is first read from the standardized transaction information. A set of interaction correspondences between member identification information and product identification information is formed by combining product identification information with the interaction correspondence set. Then, an interaction sample set is constructed based on the interaction correspondence set, and vector parameters are initialized. Member identification information and product identification information are mapped to a vector parameter space of a preset dimension. Then, the vector parameters are iteratively updated along the traversal order of the interaction sample set information, so that the vector parameter change direction of product identification information that is matched multiple times under the same member identification information tends to be consistent, and the vector parameter change direction of member identification information that is matched multiple times under the same product identification information tends to be consistent. After the traversal is completed, the converged vector parameters corresponding to member identification information are determined as member representation vectors, and the converged vector parameters corresponding to product identification information are determined as product representation vectors. The member representation vector and the product representation vector are summarized to determine the vectorized representation result information.

[0070] S30: After obtaining the recommendation request from the target member, generate a set of associated candidate products based on the product co-occurrence mapping information, and generate a set of similar candidate products based on the vectorized representation result information.

[0071] In this embodiment, the recommendation request refers to the request information initiated by the target member and carrying the member identification information corresponding to the target member. The associated candidate product set refers to the candidate set information formed by summarizing the target product identification information combination that has a mapping relationship with the target product identification information combination in the product co-occurrence mapping information. The similar candidate product set refers to the candidate set information obtained by filtering based on the similarity calculation results between the member representation vector and multiple product representation vectors in the vectorized representation result information.

[0072] Specifically, after obtaining the recommendation request from the target member, the member identification information is first read from the recommendation request and identified as the member identification information corresponding to the target member. Then, based on the member identification information, the set of product identification information corresponding to the most recent transaction time of the target member is retrieved from the normalized transaction information, and the target product identification information combination is determined from the product identification information set. Then, in the product co-occurrence mapping information, the target product identification information combination is used as the source product identification information combination to perform mapping retrieval. The target product identification information combination that matches the source product identification information combination is read and deduplicated to form a set of associated candidate products. At the same time, the member representation vector corresponding to the member identification information is read from the vectorized representation result information, and the product representation vectors corresponding to multiple product identification information outside the set of associated candidate products are read. The similarity is calculated based on the member representation vector and each product representation vector to obtain similarity score information. The product identification information is sorted according to the similarity score information, and the top-ranked product identification information is selected to form a set of similar candidate products.

[0073] S40: Select a main parameter combination from the preset parameter candidate set, and merge and sort the associated candidate product set and the similar candidate product set based on the main parameter combination to obtain the main recommendation result information.

[0074] In this embodiment, the preset parameter candidate set refers to the set of parameter combinations that are pre-configured and used for fusion ranking. Each item in the parameter combination set includes association recall weight, similarity recall weight, and preset segmentation rule information. The main parameter combination refers to the parameter combination information selected from the preset parameter candidate set and used to generate the main recommendation result information. The association recall weight refers to the weight parameter information used to weight the association score. The similarity recall weight refers to the weight parameter information used to weight the similarity score. The preset segmentation rule information refers to the rule information used to determine the candidate weight combination based on the score difference information. The main recommendation result information refers to the ranking result information formed after performing fusion ranking on the associated candidate product set and the similar candidate product set.

[0075] Specifically, when selecting the main parameter combination from the preset parameter candidate set, firstly, each parameter combination item in the preset parameter candidate set is read and a candidate parameter combination list is formed. Then, the value function information corresponding to the fusion decision strategy is read, and the value function values ​​corresponding to the candidate parameter combination list are extracted to form the sorting basis. Then, the candidate parameter combination list is sorted according to the value function values, and the first candidate parameter combination of the sorting result is read to determine the main parameter combination. When fusion sorting the associated candidate product set and the similar candidate product set based on the main parameter combination, the associated candidate product set and the similar candidate product set are first deduplicated and merged. The process involves forming a merged candidate product set, traversing the candidate products in the merged candidate product set, generating association scores and similarity scores for each candidate product, calculating the score difference information for each candidate product, reading the preset segmentation rule information in the main parameter combination, selecting candidate weight combinations based on the score difference information and the preset segmentation rule information, reading the association recall weight and similarity recall weight in the main parameter combination, performing weighted fusion on the association score and similarity score based on the candidate weight combination to obtain the fusion score for each candidate product, sorting the merged candidate product set based on the fusion score, and reading the sorting results to obtain the main recommendation result information.

[0076] S50: While outputting the main recommendation result information, select shadow parameter combinations from the preset parameter candidate set, and merge and sort the associated candidate product set and the similar candidate product set based on the shadow parameter combinations to obtain shadow recommendation result information.

[0077] In this embodiment, the shadow parameter combination refers to the parameter combination information selected from the preset parameter candidate set and used to generate the shadow recommendation result information. The shadow recommendation result information refers to the ranking result information formed after performing a fusion ranking on the associated candidate product set and the similar candidate product set based on the shadow parameter combination.

[0078] Specifically, when selecting shadow parameter combinations from a preset parameter candidate set while outputting the main recommendation result information, the system reads the candidate parameter combination list from the preset parameter candidate set and the value function information corresponding to the fusion decision strategy. Based on the value function information, it determines the value function value for each candidate parameter combination and sorts the candidate parameter combination list to obtain the main parameter combination. It then reads the association recall weight and similarity recall weight from the main parameter combination and combines them to form the sovereign weight vector. For each candidate parameter combination, the associated recall weight and similarity recall weight are read and combined to form a candidate weight vector. The similarity score ci between the sovereign weight vector and the candidate weight vector is calculated according to the following formula, where Read the value function value corresponding to the candidate parameter combination and denote it as Qi, and read the value function value corresponding to the main parameter combination and denote it as QM. Calculate the shadow selection score Ji according to the following formula, where... In the formula and Ii is a preset coefficient, representing a value of 1 when the preset segmentation rule information in the candidate parameter combination is consistent with the preset segmentation rule information in the main parameter combination, and a value of 0 when the preset segmentation rule information is inconsistent. The first candidate parameter combination is determined as the shadow parameter combination by sorting the candidate parameter combination list according to the shadow selection score Ji and reading the sorting result. When merging and sorting the associated candidate product set and the similar candidate product set based on the shadow parameter combination, the associated candidate product set and the similar candidate product set are deduplicated and merged to form a merged candidate product set. The candidate products in the merged candidate product set are traversed, and the association score and similarity score corresponding to the candidate products are generated respectively, and the score difference information corresponding to the candidate products is calculated. The preset segmentation rule information in the shadow parameter combination is read, and the candidate weight combination is selected according to the score difference information and the preset segmentation rule information. The association recall weight and similarity recall weight in the shadow parameter combination are read, and the association score and similarity score are weighted and merged according to the candidate weight combination to obtain the fusion score corresponding to the candidate product. The merged candidate product set is sorted according to the fusion score, and the sorting result is read to obtain the shadow recommendation result information.

[0079] In one embodiment, step S20, namely constructing commodity co-occurrence mapping information based on normalized transaction information, includes:

[0080] S201: Extract incremental transaction records from standardized transaction information.

[0081] In this embodiment, incremental transaction records refer to transaction record entries newly formed within a preset time range relative to preset benchmark time information. Preset benchmark time information refers to time stamp information used to distinguish processed transaction records from pending transaction records.

[0082] Specifically, when extracting incremental transaction records from standardized transaction information, the transaction occurrence time information corresponding to each transaction record in the standardized transaction information is first read, and the preset benchmark time information is also read. Then, the transaction occurrence time information is compared with the preset benchmark time information to determine the time sequence. Transaction record entries whose transaction occurrence time information is later than the preset benchmark time information are marked as entries to be extracted. At the same time, the marked entries to be extracted are sorted in ascending order according to the transaction occurrence time information to maintain the consistency of the time sequence. Finally, the sorted entries to be extracted are summarized to form incremental transaction records.

[0083] S202: Slide and summarize incremental transaction records according to the preset sliding time window to obtain window statistics.

[0084] In this embodiment, the preset sliding time window refers to the time interval parameter information used to segment and summarize the transaction occurrence time information. The preset sliding time window includes window length information and sliding step information. The window statistics information refers to the statistical record information formed after summarizing the incremental transaction records in each preset sliding time window. The statistical record information includes at least the window start time information, window end time information, number of transaction records in the window, and count information of the appearance of product identification information in the window.

[0085] Specifically, when summarizing incremental transaction records according to a preset sliding time window, the window length and sliding step information are first read from the preset sliding time window. The transaction occurrence time information in the incremental transaction records is then read and sorted in ascending order to form a time series. The earliest transaction occurrence time in the time series is used as the window start time information, and the window end time information is determined based on the window length information. Transaction record entries whose occurrence time information falls between the window start time information and the window end time information are selected from the incremental transaction records to form the set of transaction records within the window, and the number of transaction records within the window is counted. The system iterates through each item in the transaction record set within the window and accumulates the occurrence count to form the item identification information occurrence count. Then, it binds the window start time, window end time, number of transaction records within the window, and item identification information occurrence count to form the first window statistics. Subsequently, it shifts the window start time backward based on the sliding step and updates the window end time synchronously. The process of filtering the transaction record set within the window and accumulating the item identification information occurrence count is repeated until the end of the time series. Finally, the window statistics corresponding to each window are summarized to obtain the window statistics.

[0086] S203: Use a preset time decay factor to weight and update the window statistics to obtain weighted statistics.

[0087] In this embodiment, the preset time decay factor refers to the parameter information used to assign different time weights to different window statistical information. The preset time decay factor is used to characterize the weight change law corresponding to the time interval between the window start time information and the reference time information. The reference time information refers to the time base information used to uniformly calculate the time interval. The weighted statistical information refers to the statistical record information formed by introducing time weights into the number of transaction records in the window and the count of the appearance of product identification information in the window in the window statistical information.

[0088] Specifically, when updating window statistics using a preset time decay factor, the reference time information is first read, followed by the window start time and end time information from the window statistics. The time interval information is then calculated based on the window end time and the reference time information. The time weight w is calculated based on the preset time decay factor, and the time weight w is calculated according to... Calculation, where k represents the preset time decay factor, This indicates the time interval information, then reads the number of transaction records within the window from the window statistics information and follows the... Calculate the weighted transaction quantity information Nw, where N represents the number of transaction records within the window. Then, iterate through the window statistics to count the occurrences of product identifiers within the window, and for each product identifier, calculate the occurrence count C according to... Calculate the weighted occurrence count value Cw, and bind the product identification information with the weighted occurrence count value Cw to form a weighted count item set. Then bind the window start time information, window end time information, and weighted transaction quantity information Nw with the weighted count item set to form a weighted window record entry. Repeatedly perform time weight calculation and weighted update on the window statistics information corresponding to each window, and summarize the weighted window record entries to obtain the weighted statistics information.

[0089] S204: Determine the total number of transactions in the sliding window based on window statistics, and determine the support count threshold information based on the total number of transactions in the sliding window.

[0090] In this embodiment, the total number of transactions in the sliding window refers to the sum of the number of transaction records within multiple preset sliding time windows corresponding to the window statistics information, and the support count threshold information refers to the threshold information used to determine whether the combination of product identification information in the window statistics information meets the preset occurrence frequency requirement.

[0091] Specifically, when determining the total number of transactions in the sliding window based on window statistics, the number of transaction records within each preset sliding time window in the window statistics is first read and then summed to obtain the total number of transactions in the sliding window. Total number of transactions in the sliding window according to The calculation is performed, where m represents the number of windows included in the window statistics, and Ni represents the number of transaction records within the window corresponding to the i-th window. When determining the support counting threshold based on the total number of transactions in the sliding window, a preset support ratio parameter is read. And in accordance with Calculate support count threshold information In the formula This indicates the rounding up operation. This indicates the preset support ratio parameter. This represents the total number of transactions in the sliding window.

[0092] S205: Based on the support count threshold information, the weighted statistics are filtered for low frequency to obtain the weighted statistics used for co-occurrence screening.

[0093] In this embodiment, the weighted statistical information used for co-occurrence screening refers to the weighted statistical record information formed by removing product identification information or combinations of product identification information whose occurrence frequency is lower than the support count threshold information from the weighted statistical information. Low-frequency filtering refers to the screening process of retaining or removing the weighted occurrence count value in the weighted statistical information based on the support count threshold information.

[0094] Specifically, when performing low-frequency filtering on weighted statistics based on support count threshold information, the support count threshold information is read first. It then reads the set of weighted count items from the weighted statistics, iterates through the set of weighted count items for each product identifier, and compares the weighted occurrence count value Cw with the support count threshold information. Compare, in the comparison result indication The corresponding product identification information and weighted occurrence count value Cw are retained at the time of comparison result indication. The corresponding product identification information and weighted occurrence count value Cw are removed at each step. After traversal, the remaining product identification information and weighted occurrence count values ​​are summarized to form a set of filter count items. Then, the window start time information, window end time information, weighted transaction quantity information and the set of filter count items are bound to form the filtered weighted window record entries. The weighted count item set filtering is repeated for each weighted window record entry, and the filtered weighted window record entries are summarized to obtain the weighted statistical information used for co-occurrence filtering.

[0095] S206: Generate a set of mapping items between the source product identification information combination and the target product identification information combination based on the weighted statistical information used for co-occurrence screening, and summarize the set of mapping items to obtain product co-occurrence mapping information.

[0096] In this embodiment, the source product identification information combination refers to the product identification information combination that serves as the starting point of the co-occurrence relationship. The target product identification information combination refers to the product identification information combination that co-occurs and forms a co-occurrence relationship with the source product identification information combination within the same preset sliding time window. The mapping item set refers to the set of information composed of multiple mapping items. Each mapping item is used to characterize the mapping relationship from a source product identification information combination to a target product identification information combination. The product co-occurrence mapping information refers to the mapping information set formed after summarizing the mapping item set.

[0097] Specifically, when generating the mapping item set of source product identifier information combination and target product identifier information combination based on the weighted statistical information used for co-occurrence screening, the filter count item set corresponding to each filtered weighted window record entry in the weighted statistical information used for co-occurrence screening is first read. All product identifier information is then read from the filter count item set to form the window candidate product set information. Next, product identifier information is paired and combined in the window candidate product set information to form a product identifier information pair set information. The first product identifier information in each product identifier information pair is determined as the source product identifier information combination, and the second product identifier information in each pair is determined as the target product identifier information combination. Then, the weighted statistical information is read... The system records the weighted occurrence count of the corresponding window entries and accumulates the co-occurrence counts of the source product identifier information combination and the target product identifier information combination to form a mapping count value. Then, it binds the source product identifier information combination, the target product identifier information combination, and the mapping count value information to form a mapping item. The system repeats the process of constructing the window candidate product set information and generating the product identifier information set information for each filtered weighted window record entry. Each mapping item is accumulated and merged according to the source product identifier information combination and the target product identifier information combination to form a mapping item set. Finally, the mapping item set is summarized and the mapping items of the source product identifier information combination and the target product identifier information combination are written into a unified set structure to obtain the product co-occurrence mapping information.

[0098] In one embodiment, step S20, namely generating vectorized representation result information based on normalized transaction information, includes:

[0099] S207: Extract member identification information and product identification information from standardized transaction information, and form member-product interaction relationship information based on the correspondence between member identification information and product identification information.

[0100] In this embodiment, member-product interaction relationship information refers to the set of relationship information consisting of the interaction correspondence between member identification information and product identification information. The interaction correspondence refers to the pairing relationship information between member identification information and product identification information in the same standardized transaction information record.

[0101] Specifically, when extracting member identification information and product identification information from standardized transaction information to form member-product interaction relationship information, the process first iterates through the transaction record entries in the standardized transaction information and reads the member identification information and product identification information from each transaction record entry. Then, the member identification information and product identification information are written into the interaction correspondence set according to the pairing relationship of the same transaction record entry to form interaction correspondence entries. At the same time, the transaction occurrence time information is read and the transaction occurrence time information is bound to the interaction correspondence entries to maintain the consistency of the interaction occurrence order. Then, the pairing relationships of member identification information and product identification information that occur repeatedly in the interaction correspondence set are merged and summarized to generate correspondence count information. Finally, the member identification information, product identification information, and correspondence count information are written together into the member-product interaction relationship information to form the member-product interaction relationship information set.

[0102] S208: Determine the identifier jump set information based on the member-product interaction relationship information, and generate sequence information along the identifier jump set information according to the preset sequence generation rule information.

[0103] In this embodiment, the identifier jump set information refers to the jumpable corresponding set information constructed based on the correspondence between member identifier information and product identifier information in the member-product interaction relationship information. The preset sequence generation rule information refers to the rule information used to determine the jump start selection method, jump termination condition and jump path recording method. The sequence information refers to the identifier sequence set information formed by performing multiple jumps along the identifier jump set information and recording them in the jump order.

[0104] Specifically, when determining the identifier jump set information based on member-product interaction relationship information, the interaction correspondence entries between member identifier information and product identifier information in the member-product interaction relationship information are first read, and the member identifier information and product identifier information are written as jumpable identifier nodes into the node set information. Then, the correspondence between member identifier information and product identifier information and the correspondence between product identifier information and member identifier information in each interaction correspondence entry are written into the jump relationship set to form the identifier jump set information. When generating sequence information along the identifier jump set information according to the preset sequence generation rule information, the starting point selection rule information and termination condition rule information in the preset sequence generation rule information are first read, and the starting point identifier information is determined in the node set information according to the starting point selection rule information. Then, the jumpable identifier information set corresponding to the starting point identifier information is read from the identifier jump set information, and the next jump identifier information is determined according to the jump selection rule information in the preset sequence generation rule information. The starting point identifier information and the next jump identifier information are written into the sequence record information in the jump order. Then, the reading of the jumpable identifier information set and the determination of the next jump identifier information are repeated until the termination condition rule information is satisfied. The resulting sequence record information is written into the sequence information set to obtain the sequence information.

[0105] S209: Construct training sample set information based on sequence information, and perform parameter iterative updates on the training sample set information according to preset vector training rules to obtain vectorized representation results.

[0106] In this embodiment, the training sample set information refers to the sample set information formed by the pairing relationship between adjacent identifier information or interval identifier information in the sequence information; the preset vector training rule information refers to the rule information used to determine the vector parameter initialization method, sample pairing method, parameter update order and parameter update termination condition; and the parameter iterative update refers to the process information of performing multiple rounds of updates on the vector parameters according to the preset vector training rule information.

[0107] Specifically, when constructing the training sample set information based on sequence information, the sequence record information is first read and expanded into an identifier information sequence according to the record order. Then, the sample pairing method in the preset vector training rule information is read, and target identifier information and context identifier information are selected from the identifier information sequence according to the sample pairing method to form paired sample entries. These paired sample entries are then written into the training sample set information to form the training sample set information set. When performing parameter iterative updates on the training sample set information according to the preset vector training rule information, the vector parameter initialization method in the preset vector training rule information is first read, and vector parameters are initialized for member identifier information and product identifier information respectively. Then, the parameter update order in the preset vector training rule information is read, and... The paired sample entries in the training sample set are traversed along the parameter update order. For each paired sample entry, the target identifier information and context identifier information are read, and the vector parameter increment is calculated according to the parameter update rules in the preset vector training rule information. The vector parameters corresponding to the target identifier information and the vector parameters corresponding to the context identifier information are updated. Then, the parameter update termination condition in the preset vector training rule information is read, and the parameter update iteration ends when the parameter update termination condition is met. Finally, the vector parameters corresponding to the member identifier information after the parameter iteration update are determined as the member representation vector, and the vector parameters corresponding to the product identifier information after the parameter iteration update are determined as the product representation vector. The member representation vector and the product representation vector are summarized to obtain the vectorized representation result information.

[0108] In one embodiment, in step S30, both the main parameter combination and the shadow parameter combination are selected from a preset parameter candidate set, and the main parameter combination and the shadow parameter combination include at least the association recall weight, the similarity recall weight and the preset segmentation rule information.

[0109] In this embodiment, the preset parameter candidate set refers to the set of parameter combinations that are pre-configured and used for parameter selection. Each parameter combination entry in the preset parameter candidate set includes association recall weight, similarity recall weight, and preset segmentation rule information. The main parameter combination refers to the parameter combination entry information selected from the preset parameter candidate set and used to generate the main recommendation result information. The shadow parameter combination refers to the parameter combination entry information selected from the preset parameter candidate set and used to generate the shadow recommendation result information.

[0110] Specifically, when both the main parameter combination and the shadow parameter combination are selected from the preset parameter candidate set, the process first reads each parameter combination entry in the preset parameter candidate set and reads the associated recall weight, similarity recall weight, and preset segmentation rule information corresponding to each parameter combination entry. Then, the value range of the associated recall weight and similarity recall weight is checked to ensure that the values ​​of the associated recall weight and similarity recall weight meet the preset range constraints. The rule integrity check of the preset segmentation rule information is performed to ensure that the preset segmentation rule information contains the segment boundary values ​​and the corresponding weight mapping relationship of the segment. Then, the parameter combination entries that pass the check are summarized to form an optional parameter combination list. Subsequently, the selection is performed in the optional parameter combination list to obtain the main parameter combination and the selection is performed in the optional parameter combination list to obtain the shadow parameter combination. After the selection is completed, the associated recall weight, similarity recall weight, and preset segmentation rule information corresponding to the main parameter combination and the shadow parameter combination are read again to complete the confirmation of the parameter combination content.

[0111] In one embodiment, in step S50, while outputting the main recommendation result information, a shadow parameter combination is selected from a preset parameter candidate set, and the associated candidate product set and the similar candidate product set are fused and sorted based on the shadow parameter combination to obtain shadow recommendation result information, including:

[0112] S501: Based on the preset perturbation rule information, perform perturbation generation on the associated recall weight and similar recall weight, and perform perturbation generation on the preset segmentation rule information according to the preset perturbation rule information to obtain the shadow parameter combination.

[0113] In this embodiment, the preset perturbation rule information refers to the rule information used to generate the shadow parameter combination. The preset perturbation rule information includes at least the weight perturbation amplitude information, the weight perturbation direction information, the segment boundary perturbation amplitude information, and the segment boundary constraint information. Perturbation generation refers to the generation process information of performing offset calculation on the parameter values ​​and forming new parameter values ​​under the premise of maintaining the preset range constraint and the preset rule integrity constraint.

[0114] Specifically, when generating perturbations for the associated recall weights and similar recall weights based on preset perturbation rule information, the associated recall weights and similar recall weights in the main parameter combination are first read, along with the weight perturbation amplitude and direction information from the preset perturbation rule information. Then, the offset direction of the associated recall weights and similar recall weights is determined based on the weight perturbation direction information. Finally, the weight offset of the associated recall weights and the weight offset of the similar recall weights are calculated based on the weight perturbation amplitude information. The perturbation value of the associated recall weights is determined according to... The perturbation value of the similarity recall weight is calculated according to... Calculate, where This represents the associated recall weight in the main parameter combination. This represents the similarity recall weight in the main parameter combination. This represents the weight offset of the associated recall weight. This represents the weight offset of the similarity recall weight, and then... and Each comparison is performed with the preset range constraints. When the comparison result indicates that the result exceeds the preset range constraints, truncation correction is performed according to the preset range constraints to obtain the perturbation results of the associated recall weight and similar recall weight that satisfy the preset range constraints. When generating perturbation based on the preset perturbation rule information, the preset segmentation rule information in the main parameter combination is read first, and the segment boundary perturbation amplitude information and segment boundary constraint information in the preset perturbation rule information are read. Then, the boundary offset is calculated for each segment boundary value in the preset segmentation rule information, and the perturbed segment boundary value is generated based on the boundary offset. The perturbed segment boundary value and the segment boundary constraint information are subjected to order verification and range verification to ensure that the segment boundary value order is consistent and satisfies the segment boundary constraint information. Then, the segment boundary value that satisfies the verification is written back to the preset segmentation rule information to form the perturbed preset segmentation rule information. Finally, the associated recall weight perturbation result, the similar recall weight perturbation result and the perturbed preset segmentation rule information are bound together to obtain the shadow parameter combination.

[0115] S502: Deduplicatively merge the associated candidate product set and the similar candidate product set to obtain a merged candidate product set.

[0116] In this embodiment, deduplication and merging refers to the process of removing duplicate items from the product identification information in the associated candidate product set and the similar candidate product set and forming a unified set. The merged candidate product set refers to the candidate product set information obtained after deduplication and merging. The elements in the merged candidate product set use the product identification information as a unique identification key.

[0117] Specifically, when deduplicating and merging the associated candidate product set and the similar candidate product set, the product identification information in the associated candidate product set is first read and written into the merge buffer set in sequence, and a written identification set is established for each product identification information. Then, the product identification information in the similar candidate product set is read and a consistency comparison is performed with the written identification set one by one. When the consistency comparison result indicates that the product identification information does not exist in the written identification set, the product identification information is written into the merge buffer set and simultaneously written into the written identification set. When the consistency comparison result indicates that the product identification information already exists in the written identification set, the writing is skipped. Then, an integrity check is performed on the merge buffer set to confirm that each product identification information in the merge buffer set meets the uniqueness constraint. Finally, the merge buffer set that has completed deduplication and merging is determined as the merged candidate product set.

[0118] S503: Traverse the candidate products in the merged candidate product set, generate the association score and similarity score corresponding to the candidate products respectively, and calculate the score difference information corresponding to the candidate products.

[0119] In this embodiment, the association score refers to the score information formed after normalizing the mapping count value information of the candidate product's corresponding product identifier information in the product co-occurrence mapping information; the similarity score refers to the similarity calculation result information between the target member's corresponding member representation vector and the candidate product's corresponding product representation vector; and the score difference information refers to the difference calculation result information between the association score and the similarity score.

[0120] Specifically, when traversing the candidate products in the merged candidate product set and generating association scores and similarity scores, and calculating score difference information, the product identification information in the merged candidate product set is first read and traversed in the order of product identification information. When a single product identification information is encountered, the mapping count value information C corresponding to the product identification information in the product co-occurrence mapping information is read, and the total number of transactions in the sliding window corresponding to the weighted statistics information used for co-occurrence filtering is read. }},in accordance with Calculate the association score sa, where If the parameter is positive, then read the member representation vector vu from the vectorized representation result information and read the product representation vector vi corresponding to the product identifier information, based on... Calculate similarity score s s Then based on Calculate score difference information And will include product identification information, association score sa, and similarity score s s Information on score difference Binding is performed to form candidate product score record entries. The mapping count value information reading, member representation vector reading, product representation vector reading, association score calculation, similarity score calculation and score difference calculation are repeatedly performed on all product identification information in the merged candidate product set. The candidate product score record entries are then summarized to complete the traversal calculation.

[0121] S504: Select candidate weight combinations based on score difference information and preset segmentation rule information, and perform weighted fusion of related scores and similar scores based on candidate weight combinations to obtain the fusion score corresponding to the candidate product.

[0122] In this embodiment, the candidate weight combination refers to the weight value combination information composed of the association recall weight and the similarity recall weight, and the fusion score refers to the score information obtained after weighted fusion of the association score and the similarity score based on the candidate weight combination.

[0123] Specifically, when selecting candidate weight combinations and generating a fusion score based on score difference information and preset segmentation rule information, the process first reads the mapping relationship between the segment boundary value set and the corresponding weight in the preset segmentation rule information, and then reads the score difference information corresponding to the candidate product. Based on the set of values ​​for the segment boundary Execute interval determination to determine The system identifies the relevant segment interval and then, based on the segment-corresponding weight mapping relationship, reads the associated recall weight value and similarity recall weight value corresponding to the segment interval. These two weight values ​​are then bound together to form a candidate weight combination [w_a, w_s]. Subsequently, the associated score sa and similarity score s corresponding to the candidate product are read. s ,in accordance with Calculate the fusion score sf, and bind the product identification information corresponding to the candidate product with the fusion score sf to form a fusion score record entry. Repeatedly perform segmented interval judgment, candidate weight combination reading and fusion score calculation on each score difference information in the candidate product score record entry to generate fusion score record entries corresponding to all candidate products.

[0124] S505: Sort the merged candidate product set according to the fusion score to obtain shadow recommendation results information.

[0125] In this embodiment, the shadow recommendation result information refers to the sorting result information formed after sorting the merged candidate product set according to the fusion score. The shadow recommendation result information includes at least a sequence of product identification information sorted by fusion score and the fusion score corresponding to each product identification information in the product identification information sequence.

[0126] Specifically, when sorting the merged candidate product set based on the fusion score and obtaining the shadow recommendation result information, the process first reads the fusion score sf corresponding to each product identifier in the fusion score record entry and forms a sorted key-value pair set with the product identifier and the fusion score sf. Then, the sorted key-value pair set is sorted in descending order according to the fusion score sf to form a sorted key-value pair sequence. If there are two or more product identifiers with the same fusion score sf value, the preset parallel rule information is read and the arrangement order in the key-value pair sequence is determined according to the preset parallel rule information. Then, the product identifier information in the sorted key-value pair sequence is written into the product identifier information sequence in order, and the corresponding fusion score sf is written into the fusion score sequence in order. Finally, the product identifier information sequence and the fusion score sequence are bound to form the shadow recommendation result information.

[0127] In one embodiment, after step S50, i.e., the method further includes:

[0128] S60: Associate the main parameter combination, shadow parameter combination, main recommendation result information and shadow recommendation result information and write them into the sample information.

[0129] In this embodiment, sample information refers to structured record information used to record the parameter selection results and recommendation ranking results corresponding to a recommendation request. The sample information includes at least the binding relationship information between the main parameter combination, the shadow parameter combination, the main recommendation result information, and the shadow recommendation result information.

[0130] Specifically, when writing the main parameter combination, shadow parameter combination, main recommendation result information, and shadow recommendation result information into the sample information, the process first reads the member identifier information corresponding to the target member's recommendation request and the request time information corresponding to the recommendation request. Then, it reads the association recall weight, similarity recall weight, and preset segmentation rule information from the main parameter combination, and encapsulates the association recall weight, similarity recall weight, and preset segmentation rule information into fields to form the main parameter combination record field. Then, it reads the association recall weight, similarity recall weight, and preset segmentation rule information from the shadow parameter combination, and encapsulates the association recall weight, similarity recall weight, and preset segmentation rule information into fields to form the shadow parameter combination record. The system then reads the product identification information sequence and fusion score sequence from the main recommendation result information and encapsulates them into fields to form the main recommendation result information record field. Next, it reads the product identification information sequence and fusion score sequence from the shadow recommendation result information and encapsulates them into fields to form the shadow recommendation result information record field. Finally, it writes the member identification information, request time information, main parameter combination record field, shadow parameter combination record field, main recommendation result information record field, and shadow recommendation result information record field into the sample record entries according to a preset field order, and then writes the sample record entries into the sample information set to complete the association writing.

[0131] S70: Read the feedback result information corresponding to the sample information from the normalized feedback information, compare and evaluate the main recommendation result information and the shadow recommendation result information based on the feedback result information to obtain the reward value information, and update the fusion decision strategy based on the reward value information to obtain the updated fusion decision strategy.

[0132] In this embodiment, the feedback result information refers to the feedback record information formed after the target member generates behavioral feedback on the product identification information corresponding to the main recommendation result information or the shadow recommendation result information. The reward value information refers to the numerical information obtained after comparing and evaluating the main recommendation result information and the shadow recommendation result information based on the feedback result information. The fusion decision strategy refers to the strategy information used to select the main parameter combination from the preset parameter candidate set. The updated fusion decision strategy refers to the strategy information formed after updating the value function information in the fusion decision strategy.

[0133] Specifically, when reading the feedback result information corresponding to the sample information from the normalized feedback information, the member identification information, request time information, product identification information sequence from the main recommendation result information, and product identification information sequence from the shadow recommendation result information are read first from the sample information. Then, feedback record entries are filtered based on the member identification information and request time information in the normalized feedback information to obtain a candidate feedback record set. Next, the product identification information in the candidate feedback record set is matched with the product identification information sequence from the main recommendation result information, and the matched feedback record entries are marked as main feedback record entries. Simultaneously, the product identification information in the candidate feedback record set is matched with the product identification information sequence from the shadow recommendation result information, and the matched feedback record entries are marked as shadow feedback record entries. The main feedback record entries and shadow feedback record entries are then summarized to obtain the feedback result information. When comparing and evaluating the main recommendation result information and the shadow recommendation result information based on the feedback result information, the feedback type identification information in the feedback result information is read first and mapped to a feedback identification score (bj). Then, the fusion score corresponding to each product identification information in the main recommendation result information is read, and the fusion score and the feedback identification score (bj) are correlated according to the product identification information to obtain the main chain evaluation value. Then, the fusion score corresponding to each product identifier in the shadow recommendation results information is read, and the fusion score and the feedback identifier score are correlated and calculated according to the product identifier information to obtain the shadow chain evaluation value. Then based on Calculate the difference score between the control and the control group. And based on When calculating the reward value information R(q) and updating the fusion decision strategy based on the reward value information, the current value function value of the main chain in the fusion decision strategy is read. Current value function value of the shadow chain And read the subsequent value function value of the main chain. With the value function value of the shadow chain ,in accordance with Calculate the gating coefficient g(q) based on Calculate the updated main chain value function. And incorporated into the fusion decision-making strategy, based on }) Calculate the updated shadow chain value function value The fusion decision strategy is written and then the fusion decision strategy after writing is determined as the updated fusion decision strategy.

[0134] In one embodiment, in step S70, the main recommendation result information and the shadow recommendation result information are compared and evaluated based on the feedback result information to obtain reward value information, including:

[0135] S701: Generate feedback scores based on feedback results, and generate main chain evaluation values ​​and shadow chain evaluation values ​​respectively. Generate comparison difference scores based on main chain evaluation values ​​and shadow chain evaluation values, and generate reward value information based on comparison difference scores.

[0136] In this embodiment, the feedback score refers to the score information formed by mapping the feedback type identifier information in the feedback result information to a numerical value; the main chain evaluation value refers to the score information formed by performing a summary calculation on the product identifier information sequence corresponding to the main recommendation result information based on the feedback score; the shadow chain evaluation value refers to the score information formed by performing a summary calculation on the product identifier information sequence corresponding to the shadow recommendation result information based on the feedback score; and the comparison difference score refers to the difference calculation result information between the main chain evaluation value and the shadow chain evaluation value.

[0137] Specifically, when generating a feedback score based on the feedback result information, the feedback type identifier information in the feedback result information is read first, and the preset feedback identifier coefficient is read second. ,in accordance with Calculate the feedback score y(q) and map the feedback type identifier information to the feedback identifier score. When generating the main chain evaluation value and shadow chain evaluation value respectively, first read the product identification information sequence in the main recommendation result information and then read the feedback identification score that matches the product identification information sequence. Read the preset position coefficient And based on Calculate the main chain evaluation value Then, read the product identifier information sequence from the shadow recommendation results and read the feedback identifier score that matches the product identifier information sequence. Read the preset position coefficient And based on Calculate the shadow chain evaluation value When generating a comparison score based on the main chain evaluation value and the shadow chain evaluation value, positive parameters are read. And based on Calculate the difference score between the control and the control group. When generating reward value information based on the comparison difference score, the time interval parameter is read. With time constant parameter And based on Calculate the time decay factor d(q) and read the scaling factor. And based on Calculate the reward value information R(q) and write the reward value information into the reward value information record entry to complete the generation.

[0138] In one embodiment, step S70, namely updating the fusion decision strategy based on reward value information to obtain the updated fusion decision strategy, includes:

[0139] S701: Based on the current value function value of the main chain and the current value function value of the shadow chain in the fusion decision strategy, read the subsequent value function value of the main chain and the subsequent value function value of the shadow chain.

[0140] In this embodiment, the current value function value of the main chain refers to the value function value information corresponding to the main parameter combination in the fusion decision strategy; the current value function value of the shadow chain refers to the value function value information corresponding to the shadow parameter combination in the fusion decision strategy; the subsequent value function value of the main chain refers to the value function value information corresponding to the subsequent parameter combination in the fusion decision strategy; and the subsequent value function value of the shadow chain refers to the value function value information corresponding to the subsequent parameter combination in the fusion decision strategy.

[0141] Specifically, when reading the current value function value of the main chain and the current value function value of the shadow chain based on the fusion decision strategy, and when reading the subsequent value function values ​​of the main chain and the shadow chain, the parameter combination identifier information and corresponding value function value information recorded in the fusion decision strategy are read first to form a value function mapping table. Then, the main parameter combination and shadow parameter combination in the sample information are read, and the value function value corresponding to the main parameter combination is retrieved from the value function mapping table based on the parameter combination identifier information as the current value function value of the main chain. The value function value corresponding to the shadow parameter combination is retrieved as the current value function value of the shadow chain. Then, the candidate parameter combinations in the preset parameter candidate set are read, and the value function values ​​corresponding to the candidate parameter combinations are read and sorted to form a subsequent candidate parameter combination sequence. Then, the first candidate parameter combination after the main parameter combination is read from the subsequent candidate parameter combination sequence, and the corresponding value function value is retrieved from the value function mapping table as the subsequent value function value of the main chain. Finally, the first candidate parameter combination after the shadow parameter combination is read from the subsequent candidate parameter combination sequence, and the corresponding value function value is retrieved from the value function mapping table as the subsequent value function value of the shadow chain.

[0142] S702: Based on the control difference score, read the temperature parameters, and according to... Calculate the gating coefficient g(q).

[0143] In this embodiment, the gating coefficient g(q) refers to the coefficient information used to adjust the update amplitude of the main chain value function value and the update amplitude of the shadow chain value function value, the control difference score refers to the difference score information calculated from the main chain evaluation value and the shadow chain evaluation value, and the temperature parameter refers to the parameter information used to adjust the scaling amplitude of the control difference score.

[0144] Specifically, when reading temperature parameters and calculating the gate coefficient g(q) based on the control difference score, the control difference score is read first. And read the temperature parameter T, and score the difference between the control and the control. Perform a ratio calculation with the temperature parameter T to obtain the scaling difference value. Then based on Calculate the gating coefficient g(q) and write the gating coefficient g(q) into the gating coefficient record entry to complete the generation of the gating coefficient g(q).

[0145] S703: Obtain the update coefficient and discount coefficient, and calculate the updated main chain value function value and the updated shadow chain value function value based on the reward value information.

[0146] In this embodiment, the update coefficient refers to the parameter information used to control the update step of the value function value, the discount coefficient refers to the parameter information used to control the participation of subsequent value function values ​​in the update, the updated main chain value function value refers to the update result information calculated based on the current value function value of the main chain, combined with the reward value information and the subsequent value function value of the main chain, and the updated shadow chain value function value refers to the update result information calculated based on the current value function value of the shadow chain, combined with the reward value information and the subsequent value function value of the shadow chain.

[0147] Specifically, when obtaining the update coefficient and discount coefficient, and calculating the updated main chain value function value and the updated shadow chain value function value based on the reward value information, the update coefficient is read first. With discount factor It also reads the reward value information R(q) and the current value function value of the main chain. With the subsequent value function value of the main chain And read the gating coefficient g(q), based on Calculate the updated main chain value function. Then read the current value function value of the shadow chain. With the value function value of the shadow chain And read the gating coefficient g(q), based on Calculate the updated shadow chain value function. And update the main chain value function value Compared with the updated shadow chain value function value Write the value function record corresponding to the fusion decision strategy to complete the writing of the calculation results.

[0148] S704: Write the updated main chain value function value into the fusion decision strategy to update the main chain value function information, and write the updated shadow chain value function value into the fusion decision strategy to update the shadow chain value function information, so as to obtain the updated fusion decision strategy.

[0149] In this embodiment, the main chain value function information refers to the value function value record information corresponding to the main parameter combination in the fusion decision strategy, the shadow chain value function information refers to the value function value record information corresponding to the shadow parameter combination in the fusion decision strategy, and the updated fusion decision strategy refers to the updated strategy information in which the main chain value function information and the shadow chain value function information are written into the fusion decision strategy.

[0150] Specifically, when writing the updated main chain value function value into the fusion decision strategy to update the main chain value function information, and when writing the updated shadow chain value function value into the fusion decision strategy to update the shadow chain value function information, the value function mapping table in the fusion decision strategy is read first, and the parameter combination identifier information corresponding to the main parameter combination and the parameter combination identifier information corresponding to the shadow parameter combination are read. Then, the updated main chain value function value is written into the fusion decision strategy. Write the value function value field in the value function mapping table that matches the parameter combination identifier information corresponding to the main parameter combination to overwrite the current value function value of the main chain, and then update the shadow chain value function value. The value function value field that matches the parameter combination identifier information corresponding to the shadow parameter combination is written into the value function mapping table to cover the current value function value of the shadow chain. Then, a consistency check is performed on the value function mapping table to confirm that the value function value field corresponding to the main parameter combination and the value function field corresponding to the shadow parameter combination have been written and updated, and to confirm that the identifier information of each parameter combination in the value function mapping table maintains the unique constraint. Finally, the value function mapping table after writing and updating is backfilled into the fusion decision strategy, and the backfilled fusion decision strategy is determined as the updated fusion decision strategy.

[0151] It should be understood that the sequence number of each step in the above embodiments does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of the embodiments of this application.

[0152] In one embodiment, a product recommendation device is provided, which corresponds one-to-one with a product recommendation method described in the above embodiments. For example... Figure 2 As shown, the product recommendation device includes a normalization module, a mapping and representation module, a candidate generation module, a main sorting module, and a shadow sorting module.

[0153] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is used as an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the device can be divided into different functional units or modules to complete all or part of the functions described above.

[0154] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application, and should all be included within the protection scope of this application.

Claims

1. A product recommendation method, characterized in that, The method includes: Historical transaction information and behavioral feedback information are obtained, and the historical transaction information and behavioral feedback information are standardized to obtain standardized transaction information and standardized feedback information. Commodity co-occurrence mapping information is constructed based on standardized transaction information, and vectorized representation results are generated based on standardized transaction information; After obtaining the recommendation request from the target member, a set of associated candidate products is generated based on the product co-occurrence mapping information, and a set of similar candidate products is generated based on the vectorized representation results. Select a main parameter combination from the preset parameter candidate set, and merge and sort the associated candidate product set and the similar candidate product set based on the main parameter combination to obtain the main recommendation result information; While outputting the main recommendation result information, shadow parameter combinations are selected from the preset parameter candidate set. Based on the shadow parameter combinations, the related candidate product set and the similar candidate product set are fused and sorted to obtain the shadow recommendation result information.

2. The product recommendation method according to claim 1, characterized in that, The construction of commodity co-occurrence mapping information based on standardized transaction information includes: Extract incremental transaction records from standardized transaction information; The incremental transaction records are summarized by sliding according to a preset sliding time window to obtain window statistics; The window statistics are updated by weighting them using a preset time decay factor to obtain weighted statistics. The total number of transactions in the sliding window is determined based on window statistics, and the support count threshold is determined based on the total number of transactions in the sliding window. Low-frequency filtering is performed on the weighted statistics based on the support count threshold information to obtain weighted statistics for co-occurrence screening; Based on the weighted statistical information used for co-occurrence screening, a set of mapping items is generated between the source product identification information combination and the target product identification information combination, and the set of mapping items is summarized to obtain the product co-occurrence mapping information.

3. The product recommendation method according to claim 1, characterized in that, The information generated based on the normalized transaction information to produce a vectorized representation includes: Member identification information and product identification information are extracted from standardized transaction information, and member-product interaction relationship information is formed based on the correspondence between member identification information and product identification information; Based on the member-product interaction relationship information, determine the identifier jump set information, and generate sequence information along the identifier jump set information according to the preset sequence generation rule information; The training sample set information is constructed based on sequence information, and the parameters of the training sample set information are iteratively updated according to the preset vector training rules to obtain the vectorized representation result information.

4. The product recommendation method according to claim 1, characterized in that, Both the main parameter combination and the shadow parameter combination are selected from the preset parameter candidate set, and the main parameter combination and the shadow parameter combination include at least the association recall weight, the similarity recall weight and the preset segmentation rule information.

5. The product recommendation method according to claim 1, characterized in that, While outputting the main recommendation result information, the process involves selecting shadow parameter combinations from a preset parameter candidate set, and then fusing and sorting the associated candidate product set and the similar candidate product set based on the shadow parameter combinations to obtain shadow recommendation result information, including: Based on the preset perturbation rule information, perturbation generation is performed on the associated recall weight and similar recall weight, and perturbation generation is performed on the preset segmentation rule information according to the preset perturbation rule information to obtain the shadow parameter combination; The sets of related and similar candidate products are deduplicated and merged to obtain a merged candidate product set. Iterate through the candidate products in the merged candidate product set, generate the association score and similarity score for each candidate product, and calculate the score difference information for each candidate product. Candidate weight combinations are selected based on score difference information and preset segmentation rules. Based on the candidate weight combinations, the associated scores and similar scores are weighted and fused to obtain the fusion score corresponding to the candidate product. The candidate product set is sorted according to the fusion score to obtain the shadow recommendation results.

6. The product recommendation method according to claim 1, characterized in that, The method further includes: The main parameter combination, shadow parameter combination, main recommendation result information and shadow recommendation result information are associated and written into the sample information; The feedback result information corresponding to the sample information is read from the normalized feedback information. The main recommendation result information and the shadow recommendation result information are compared and evaluated based on the feedback result information to obtain the reward value information. The fusion decision strategy is updated based on the reward value information to obtain the updated fusion decision strategy.

7. A product recommendation method according to claim 6, characterized in that, The step of comparing and evaluating the main recommendation result information and the shadow recommendation result information based on the feedback result information to obtain reward value information includes: Feedback scores are generated based on the feedback results, and main chain evaluation values ​​and shadow chain evaluation values ​​are generated separately. Comparison difference scores are generated based on the main chain evaluation values ​​and shadow chain evaluation values, and reward value information is generated based on the comparison difference scores.

8. The product recommendation method according to claim 7, characterized in that, The process of updating the fusion decision strategy based on reward value information to obtain the updated fusion decision strategy includes: Based on the current value function value of the main chain and the current value function value of the shadow chain in the fusion decision-making strategy, and read the subsequent value function value of the main chain and the subsequent value function value of the shadow chain; Based on the difference score, temperature parameters were read and then... Calculate the gating coefficient g(q); Obtain the update coefficient and discount coefficient, and calculate the updated main chain value function value based on the reward value information. Compared with the updated shadow chain value function value; The updated main chain value function value is written into the fusion decision strategy to update the main chain value function information, and the updated shadow chain value function value is written into the fusion decision strategy to update the shadow chain value function information, so as to obtain the updated fusion decision strategy.

9. A product recommendation device, characterized in that, The product recommendation device includes: The standardization module is used to acquire historical transaction information and behavioral feedback information, and to standardize the historical transaction information and behavioral feedback information to obtain standardized transaction information and standardized feedback information. The mapping and representation module is used to construct commodity co-occurrence mapping information based on normalized transaction information and generate vectorized representation results based on normalized transaction information. The candidate generation module is used to generate a set of associated candidate products based on product co-occurrence mapping information after obtaining the recommendation request from the target member, and to generate a set of similar candidate products based on the vectorized representation result information. The main sorting module is used to select main parameter combinations from the preset parameter candidate set, and to merge and sort the related candidate product set and the similar candidate product set based on the main parameter combinations to obtain the main recommendation result information; The shadow ranking module is used to select shadow parameter combinations from a preset parameter candidate set while outputting the main recommendation result information. Based on the shadow parameter combinations, the related candidate product set and the similar candidate product set are merged and ranked to obtain the shadow recommendation result information.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the steps of a product recommendation method as described in any one of claims 1 to 8.

Images