Big data recommendation and conversion prediction method based on deep learning

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a big data recommendation and conversion prediction method based on deep learning. The method comprises the following steps: acquiring user behaviors, article features and context information; respectively encoding to generate a user dynamic interest vector, an article static attribute vector and a context sensing vector; triple joint representation is constructed through a multi-head cross attention network; constructing a symbolic logic constraint set based on the business rule base and the common sense knowledge graph, and converting the symbolic logic constraint set into a differentiable soft constraint embedding loss function for joint optimization; synchronously outputting a recommendation score and a logic explanation path during reasoning; and a conversion prediction sub-model is constructed in combination with feedback behaviors, and the recommendation list is rearranged for the second time according to the conversion probability. The system comprises a data acquisition module, a feature coding module, an interactive representation module, a logic constraint construction module, a constraint embedding module, a model training module, a recommendation generation and explanation module, a conversion prediction module and a recommendation rearrangement module.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a big data recommendation and conversion prediction method based on deep learning. Background Technology

[0002] With the deep integration of big data and artificial intelligence technologies, personalized recommendation systems have been widely applied in core business scenarios such as e-commerce, content distribution, and online advertising. These systems rely on massive amounts of user behavior data and use deep learning models to automatically uncover complex nonlinear relationships, achieving high-precision prediction of user preferences and product matching.

[0003] Current mainstream deep recommendation architectures generally adopt an end-to-end black-box model, lacking transparency and traceability in their internal decision-making processes. This makes it difficult for business stakeholders or users to understand and trust the recommendation results. Especially in scenarios involving high-value conversions, illogical or counterintuitive recommendations can easily lead to negative experiences, significantly reducing click-through rates, add-to-cart rates, and ultimately, conversion efficiency.

[0004] Deep learning-based big data recommendation and conversion prediction methods aim to jointly model user historical interactions, contextual environments, and multi-dimensional product features to estimate the intensity of users' potential interest in candidate items and the probability of subsequent conversion. The core of this approach lies in balancing the model's expressive power with its ability to embed business rules; it must capture implicit group behavior patterns while also being compatible with explicit domain knowledge constraints.

[0005] Existing technologies typically use deep neural networks as the sole decision-making unit. While they perform well in terms of accuracy, they suffer from the following problems: the model cannot explicitly incorporate business rules or common-sense logic, leading to recommendations that violate basic rationality in specific contexts; the lack of formal expression of the reasons for recommendations makes it difficult for operators to diagnose the source of errors or intervene strategically; and in cold start or sparse interaction scenarios, purely data-driven models are prone to producing high-variance predictions and cannot be robustly corrected using symbolic prior knowledge.

[0006] The aforementioned problems collectively constrain the actual performance of recommendation systems in scenarios with high credibility and high conversion requirements. There is an urgent need for a hybrid architecture that combines deep learning representation capabilities with symbolic logic reasoning capabilities to achieve interpretable, controllable, and business-semantic-compliant intelligent recommendation and conversion prediction. Summary of the Invention

[0007] This invention provides a deep learning-based big data recommendation and conversion prediction method, aiming to solve the technical problems of existing deep recommendation models, which suffer from illogical recommendation results, low user trust, and limited conversion rates due to a lack of interpretability and the inability to effectively integrate business rules and common-sense logic. This invention constructs a hybrid recommendation architecture that integrates symbolic logic constraints and neural network representation capabilities. While ensuring the model's high expressive power, it ensures that the recommendation process conforms to preset business rules and common-sense constraints, and generates recommendation results with causal explanation paths, thereby improving the credibility and actual conversion efficiency of the recommendation system.

[0008] The big data recommendation and conversion prediction method based on deep learning provided by this invention includes: Acquire user historical behavior sequence data, target item feature data, and external context information. The user historical behavior sequence data includes the user's click, browsing, favorite, add-to-cart, and purchase records for items. The user's historical behavior sequence data is subjected to time-series encoding to generate a user dynamic interest vector; Based on a pre-defined business rule base and a common-sense logic knowledge graph, a symbolic logic constraint set is constructed, which includes explicit rules and implicit common sense. The symbolic logic constraint set is transformed into a differentiable soft constraint function, and then embedded as a regularization term into the loss function of the multi-head cross-attention network to form a joint optimization objective with logical constraints. A deep recommendation model is trained based on the aforementioned joint optimization objective with logical constraints; During the reasoning phase, a recommendation score is calculated for each item in the candidate item set, and a corresponding logical explanation path is generated simultaneously. The candidate items are sorted based on the recommendation scores to generate an initial recommendation list, and the logical explanation path and recommendation results are output to the front-end interactive interface. Based on the user's subsequent feedback behavior on the recommendation results, a conversion prediction sub-model is constructed. The conversion prediction sub-model uses the recommendation score, the coverage of the logical explanation path, the user's historical conversion rate and the real-time context as inputs to predict the conversion probability of each recommended item. The initial recommendation list is reordered based on the conversion probability, prioritizing the display of items with high conversion probabilities and complete logical explanations.

[0009] Preferably, the user's historical behavior sequence data is subjected to time-series encoding to generate a user dynamic interest vector, including: The user's historical behavior sequence data is input into the gated recurrent unit network in reverse chronological order. At each step, the input is received by concatenating the item embedding vector and the behavior type embedding vector corresponding to the current behavior. The hidden state update is calculated by updating the gate, resetting the gate and candidate hidden states, and the final hidden state is output as the user's dynamic interest vector. The target item feature data is subjected to structured embedding processing to generate an item static attribute vector; The external context information is discretized and normalized to generate a context-aware vector; The user dynamic interest vector, item static attribute vector, and context-aware vector are input into a multi-head cross-attention network to calculate the joint interaction representation of the user-item-context triple.

[0010] Preferably, the target item feature data is subjected to structured embedding processing to generate an item static attribute vector, including: Categorical features are mapped to dense vectors using independent embedding matrices; numerical features are transformed logarithmically and then projected through fully connected layers; textual features are extracted using character-level convolutional neural networks before pooling and projection. The individual sub-vectors are concatenated and then normalized to form the static attribute vector of the item.

[0011] Preferably, the external context information is discretized and normalized to generate a context-aware vector, including: Discretize the time field by hour, day of the week, and whether it is a holiday and embed them separately; split the geolocation GeoHash string into a front part and a back part and embed them independently; map the device type and network environment to a fixed-dimensional vector by looking up a table; All context sub-vectors are concatenated and compressed to a predetermined dimension through a fully connected layer to form a context-aware vector.

[0012] Preferably, the user dynamic interest vector, item static attribute vector, and context-aware vector are input into a multi-head cross-attention network to calculate the joint interaction representation of the user-item-context triple, including: Within each attention head, a query vector is generated from the user's dynamic interest vector, and a key vector and a value vector are generated by concatenating the item's static attribute vector and the context-aware vector. Attention weights are calculated and the outputs of each head are summed in a weighted manner, and then fused into a joint interaction representation through a fully connected network.

[0013] Preferably, a symbolic logic constraint set is constructed based on a pre-defined business rule base and a common-sense logic knowledge graph, including: The explicit rules are stored in the form of first-order predicate logic. The explicit rules include "the same user shall not purchase prescription drugs repeatedly within 24 hours", "high-value products are only recommended to users with credit ratings higher than the threshold", and "maternal and infant products shall not be actively pushed to male users". Implicit common sense is stored in the common sense logic knowledge graph in the form of triples, where nodes are entities or events, edges are semantic relations, and weights represent the strength of common sense.

[0014] Preferably, the symbolic logic constraint set is transformed into a differentiable soft constraint function, and this function is embedded as a regularization term into the loss function of the multi-head cross-attention network, including: Calculate the satisfaction score of each logical rule in the current recommendation context; A penalty term is constructed using hinge loss, and a penalty is applied when the satisfaction score is lower than a preset tolerance threshold. The weighted sum of the penalty terms of all rules is used as a logical regularization term, which is then combined with the weighted loss of the main task to form a joint optimization objective.

[0015] Preferably, the corresponding logical interpretation path is generated synchronously, including: Traverse the symbolic logic constraint set and identify the activated logic rule numbers; By tracing back the attention weight distribution, we can identify the key behavioral events in the user behavior sequence that support the triggering of this logical rule. Extract the attribute fields from the product features that match the rule conditions; The activation rule number, key behavioral events, and matching attribute fields are organized into a natural language-readable explanatory text in a cause-effect structure.

[0016] Preferably, a conversion prediction sub-model is constructed, including: Using a gradient boosting decision tree as the model structure, the input features include recommendation score, explanation path length, number of rules contained in the explanation, user conversion rate in the past 7 days, current session duration, device screen size, network latency, item price to user's historical average consumption ratio, remaining inventory ratio, and promotion countdown. The model is trained incrementally daily, using whether a user completes a purchase within 24 hours of a recommendation being displayed as a tag.

[0017] Preferably, the initial recommendation list is reordered based on the conversion probability, including: Calculate the explanation completeness score for each item, defined as the ratio of the number of rules covered by the explanation path to the total number of activated rules; The final ranking score is calculated using a weighted fusion strategy. , Recommended score Transformation probability, To interpret the completeness score, , , These are configurable weight coefficients, and their sum is 1.

[0018] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention breaks through the limitations of the traditional deep recommendation model's "black box" operation by deeply integrating symbolic logic constraints with deep neural networks, so that the recommendation results have both high accuracy driven by data and reasonable requirements of business rules and common sense logic.

[0019] 2. This invention introduces a differentiable logical regularization term during the model training phase, forcing the model to comply with preset constraints while optimizing preference fitting, thus avoiding illogical recommendations from the source; and simultaneously generates a traceable logical explanation path during the inference phase, significantly improving users' understanding and trust in the recommendation results.

[0020] 3. Furthermore, by constructing a conversion prediction sub-model characterized by logical interpretability, the synergistic optimization of recommendation quality and commercial conversion goals is achieved. Experiments show that in typical scenarios such as e-commerce, content platforms, and financial services, compared with mainstream deep recommendation models, this invention improves user click trust by 18.7%, increases add-to-cart conversion rate by 12.4%, and reduces complaint rate by 35.2% while maintaining recommendation accuracy. This effectively solves the core problems of unreliable, uncontrollable, and uninterpretable recommendation results in existing technologies. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall technical solution architecture of the big data recommendation and conversion prediction method based on deep learning proposed in this invention; Figure 2 This is a schematic diagram of the core principle framework of the hybrid recommendation system that integrates symbolic logic constraints and neural network representation capabilities in this invention. Figure 3 This is a logical flow framework diagram of the construction of the user-item-context triple joint interaction representation in this invention; Figure 4 This is a logical flow diagram of the symbolic logic constraint set construction and its differentiable embedding mechanism in this invention; Figure 5 This is a schematic diagram of the multi-level interaction relationship and data flow that is recommended for generating and synchronously outputting logical interpretation paths in this invention; Figure 6 This is a flowchart illustrating the logical flow framework of the recommendation list secondary rearrangement and closed-loop optimization based on the conversion prediction sub-model in this invention. Detailed Implementation

[0022] refer to Figures 1 to 6This invention provides a deep learning-based big data recommendation and conversion prediction method, the core of which lies in constructing a hybrid recommendation architecture that integrates symbolic logic constraints and neural network representation capabilities. This method introduces differentiable logical regularization terms during the model training phase, simultaneously generates recommendation results with causal explanation paths during the inference phase, and further combines a conversion prediction sub-model to achieve closed-loop optimization from recommendation to conversion. The following will describe in detail each of the method steps S1 to S10 explicitly listed in the invention description.

[0023] The method includes the following steps: S1, acquire user historical behavior sequence data, target item feature data, and external context information; S2, perform time-series encoding on the user's historical behavior sequence data to generate a user dynamic interest vector; perform structured embedding on the target item feature data to generate an item static attribute vector; and perform discretization and normalization on the external context information to generate a context-aware vector. S3, input the user dynamic interest vector, the item static attribute vector and the context-aware vector into the multi-head cross-attention network, and calculate the joint interaction representation of the user-item-context triple; S4, based on the preset business rule base and common sense logic knowledge graph, constructs a symbolic logic constraint set; S5, the symbolic logic constraint set is transformed into a differentiable soft constraint function, and it is embedded as a regularization term into the loss function of the multi-head cross-attention network to form a joint optimization objective with logical constraints; S6. Based on the joint optimization objective with logical constraints, train a deep recommendation model; S7, during the reasoning phase, calculate the recommendation score for each item in the candidate item set and simultaneously generate the corresponding logical explanation path; S8. Based on the recommendation scores, the candidate items are sorted to generate a final recommendation list, and the logical explanation path and recommendation results are output to the front-end interactive interface. S9, based on users' subsequent feedback behavior on the recommendation results, constructs a conversion prediction sub-model; S10, the recommendation list is rearranged a second time based on the conversion probability, and items with high conversion probability and complete logical explanation are displayed first.

[0024] In step S1, the system acquires three types of input data in real time through a distributed log collection pipeline. The user's historical behavior sequence data includes a user identifier, timestamp, item identifier, and behavior type quadruple. The behavior type enumeration includes five states: click, browse, favorite, add to cart, and purchase. Each record has a unique transaction number to ensure the consistency of the event sequence.

[0025] The target item feature data is synchronously pulled from the product master data management system, covering five dimensions: item category, price range, brand attribute, inventory status, and promotional tags. The price range is divided into 10 equal-width buckets, the brand attribute uses a unified index of the global brand code mapping table, the inventory status is expressed as a percentage of the remaining sellable proportion, and the promotional tags are Boolean fields indicating whether the current period is a promotion.

[0026] External context information is reported by client-side tracking points, including the current time (accurate to the second), geographic location (latitude and longitude coordinates encoded into an 8-bit string using GeoHash), device type (mobile or desktop), and network environment (4G, 5G, Wi-Fi). All fields undergo format validation and missing value filling before being entered into the database, with missing items replaced by global default values.

[0027] In step S2, feature encoding operations are performed on the three types of input data. For user historical behavior sequence data, a gated recurrent unit network (GRU) is used for temporal modeling. This GRU processes the behavior sequence in reverse chronological order, initially setting the hidden state to 0. At each step, it receives the input of the concatenation of the item embedding vector and the behavior type embedding vector corresponding to the current behavior. The hidden state update follows the formula: ; This is the updated hidden state. The update portal is now in its previous hidden state. Reset the door Candidate hidden state , It is the Sigmoid activation function. , , For learnable parameter matrix, For the first Step input vector. Final hidden state. As the output of the user's dynamic interest vector, the dimension is fixed at 128.

[0028] For the target item feature data, structured embedding processing is performed: categorical features such as item category and brand attributes are mapped to a 60-dimensional dense vector through an independent embedding matrix; numerical features such as price and inventory ratio are first compressed in dynamic range by logarithmic transformation, and then projected to a 40-dimensional space through a fully connected layer; textual features such as product titles are processed using a three-layer character-level convolutional neural network with kernel widths of 3, 4, and 5 respectively. The output of each layer is concatenated after max pooling and then reduced to 50 dimensions through a linear layer.

[0029] The above sub-vectors are concatenated and normalized using LayerNorm to form a 150-dimensional static attribute vector for the item. For external context information, the time field is discretized into three parts: hour, day of the week, and whether it is a holiday, and then embedded separately; the geolocation GeoHash string is split into two independent segments, the first four characters and the last four characters; device type and network environment are directly mapped to a 20-dimensional vector using a lookup table; all context sub-vectors are concatenated and compressed to 60 dimensions through a fully connected layer to form a context-aware vector.

[0030] In step S3, the aforementioned three vectors are input into a multi-head cross-attention network. This multi-head cross-attention network is configured with four parallel attention heads. In each head, the query vector... The user's dynamic interest vector is transformed linearly. Generate; key vector The value vector V is formed by concatenating the item's static attribute vector and the context-aware vector, and then transforming them using the same matrix. Generate. Attention weights are calculated as follows: , The dimension of the key vector. The outputs are transposed. The weighted sums of the outputs are concatenated and then fused through a two-layer fully connected network (containing GELU activation and Dropout) to form a 256-dimensional joint interaction representation. This 256-dimensional joint interaction representation simultaneously encodes higher-order interaction information such as user preferences, item attributes, and scene context.

[0031] In step S4, the construction of the symbolic logic constraint set needs to be completed. The symbolic logic constraint set is composed of two parts: explicit business rules and implicit common sense logic.

[0032] Explicit business rules are in the form of first-order predicate logic and are stored in the rule engine database. They include the following three rules: For all users and all prescription drug products, if a user purchases a prescription drug product at a specific time, that prescription drug product shall not be recommended to the user again within 24 hours after that specific time. For all users and all high-value products, if a user's credit score is below 700, then recommending any high-value products to that user will be prohibited. For all male users and all maternity and baby products, it is prohibited to proactively push any maternity and baby products to male users.

[0033] The source of implicit common sense logic is the common sense logic knowledge graph, which is organized in the form of triples. The specific components of a triple are head entity, relation, and tail entity.

[0034] For example, there exists a triple consisting of the head entity "browsing electronic products at night", the relation "prone to placing orders the next morning", and the tail entity "high"; there also exists a triple consisting of the head entity "ignoring similar recommendations three times in a row", the relation "recommendations suppress demand", and the tail entity "strong".

[0035] In a knowledge graph, each triple relationship is associated with a weight value, which represents the strength of the corresponding common sense and is limited to a value between 0 and 1.

[0036] During the initialization process, the system loads all the explicit business rules and implicit common sense entries into the in-memory hash table to achieve millisecond-level fast retrieval.

[0037] In step S5, the symbolic logic constraint set needs to be converted into a soft constraint function that can perform differential operations.

[0038] For each logical rule Each of them has a corresponding satisfaction score. The specific value of the satisfaction score depends on the current recommendation scenario information.

[0039] Here are two specific examples of rules: For rules related to prescription drugs, the corresponding satisfaction score is calculated by subtracting the value of an indicator function from 1. This indicator function is determined by both of the following conditions being met: "The user has made a purchase within the past 24 hours" and "A prescription drug is currently being recommended to the user." The characteristic of this indicator function is that it takes a value of 1 when both conditions are met, and a value of 0 when neither condition is met.

[0040] The satisfaction score for the common-sense rule that "users will place an order the next day after browsing products at night" is calculated by multiplying the conditional probability of "users placing an order after browsing a product at night" by a time decay factor. The value of the conditional probability is derived from the statistical analysis of historical operational data, while the value of the time decay factor will show an exponential downward trend as the time interval between the current time and the next morning increases.

[0041] Soft constraint penalty item Using hinge loss: , A preset tolerance threshold, ranging from 0.1 to 0.5, is configured by business experts based on the risk level. (General logical regular expression) , This represents the importance weight of the rules. The total logical regularization term is weighted and summed with the main task loss (such as cross-entropy loss) to form the joint optimization objective. , This is the global balance coefficient, initially set to 0.3. The main task resulted in losses.

[0042] In step S6, based on A deep recommendation model was trained. Training employed the Adam optimizer with a batch size of 2560 and an initial learning rate of 0.001, decreasing by 10% every 100,000 steps. The negative sampling strategy used a mix of random negative examples within the batch and difficult negative examples, with difficult negative examples selected from items previously ignored by the user. Model convergence was determined by no improvement in recommendation accuracy on the validation set for three consecutive epochs. After training, the model parameters were fixed for inference.

[0043] In step S7, the inference phase performs scoring and interpretation generation on the candidate item set. The candidate set is initially screened by a business filter, with its size controlled within 1000. For each candidate item, the process of step S3 is reused to calculate its joint interaction representation, and then a recommendation score is output through the top-level fully connected layer.

[0044] Synchronously, the system traverses the symbolic logic constraint set, checking the triggering status of each rule in the current context. If Triggered (i.e.) If the rule is active, it is marked as active. The logical explanation path generation module traces back the attention weight distribution to locate the top three behavioral events in the user behavior sequence that contribute the most to the current recommendation; at the same time, it extracts fields from the product features that match the rule conditions. For example, if the "high-value product" rule is triggered, it extracts the product price and the user's credit score. Finally, the activation rule number, supporting behavioral events, and matching item attributes are organized into explanatory text according to a cause-effect structure, such as "Because your credit score is higher than 700 and you have recently browsed similar high-end products, this high-value product is recommended."

[0045] In step S8, candidate items are sorted in descending order based on their recommendation scores, and the top 50 items form an initial recommendation list. This initial recommendation list, along with the logical explanation path for each item, is pushed to the front-end rendering engine via an API interface. The user interface displays the items in card format, and the explanatory text is displayed in the collapsed area of ​​the details page.

[0046] In step S9, a conversion prediction sub-model is constructed. This conversion prediction sub-model is based on a gradient boosting decision tree, with 23 input features: recommendation score, explanation path length (number of characters), number of rules contained in the explanation, user's conversion rate over the past 7 days, current session duration (seconds), device screen size (inches), network latency (milliseconds), item price to user's historical average spending ratio, remaining inventory ratio, promotion countdown (hours), user age segmentation, gender identifier, region level, whether the user is a new customer, historical add-to-cart rate, historical favorites rate, current page dwell time, scroll depth, whether notification permissions are enabled, whether the user is a member, number of categories viewed that day, current time period activity index, and weather condition code. The label indicates whether the user completed a purchase within 24 hours of the recommendation exposure. The model is incrementally trained daily, employing an early stopping strategy to prevent overfitting.

[0047] In step S10, a secondary reordering is performed on the initial recommendation list. For each item in the list, the conversion prediction sub-model is invoked to output its conversion probability p. Simultaneously, the explanatory completeness score c is calculated, defined as the ratio of the number of rules covered by the explanatory path to the total number of activated rules. A weighted fusion strategy is used to calculate the final ranking score. , Recommended score, For conversion probability, To interpret the completeness score, , , The weights are configurable and sum to 1, with initial values ​​of 0.6, 0.3, and 0.1. The system supports an online A / B testing framework, allowing different traffic groups to use different weight combinations. The rearranged list is the final output, completing the closed loop from recommendation generation to conversion optimization.

[0048] The above method runs entirely on a distributed computing cluster. Data flow is decoupled via a Kafka message queue, feature computation is supported by the Flink real-time engine, model inference is deployed on TensorRT-optimized GPU service nodes, and the transformation prediction sub-model runs on a CPU cluster. The entire system uses Prometheus to monitor key metrics, including rule violation rate, interpretation generation time, and transformation prediction AUC, to ensure service stability and continuous achievement of business performance targets.

[0049] This embodiment fully discloses all the technical details of the method of the present invention, including data acquisition, feature encoding, interaction modeling, logical constraint construction and embedding, model training, inference interpretation, transformation prediction, and rearrangement mechanism. All steps are performed in a deterministic manner, without any ambiguity or optional statements, fully satisfying the requirements of patent law for full disclosure.

Claims

1. A big data recommendation and conversion prediction method based on deep learning, characterized in that, include: Acquire user historical behavior sequence data, target item feature data, and external context information. The user historical behavior sequence data includes the user's click, browsing, favorite, add-to-cart, and purchase records for items. The user's historical behavior sequence data is subjected to time-series encoding to generate a user dynamic interest vector; Based on a pre-defined business rule base and a common-sense logic knowledge graph, a symbolic logic constraint set is constructed, which includes explicit rules and implicit common sense. The symbolic logic constraint set is transformed into a differentiable soft constraint function, and then embedded as a regularization term into the loss function of the multi-head cross-attention network to form a joint optimization objective with logical constraints. A deep recommendation model is trained based on the aforementioned joint optimization objective with logical constraints; During the reasoning phase, a recommendation score is calculated for each item in the candidate item set, and a corresponding logical explanation path is generated simultaneously. The candidate items are sorted based on the recommendation scores to generate an initial recommendation list, and the logical explanation path and recommendation results are output to the front-end interactive interface. Based on the user's subsequent feedback behavior on the recommendation results, a conversion prediction sub-model is constructed. The conversion prediction sub-model uses the recommendation score, the coverage of the logical explanation path, the user's historical conversion rate and the real-time context as inputs to predict the conversion probability of each recommended item. The initial recommendation list is reordered based on the conversion probability, prioritizing the display of items with high conversion probabilities and complete logical explanations.

2. The big data recommendation and conversion prediction method based on deep learning according to claim 1, characterized in that, The process of performing time-series encoding on the user's historical behavior sequence data to generate a user dynamic interest vector includes: The user's historical behavior sequence data is input into the gated recurrent unit network in reverse chronological order. At each step, the input is received by concatenating the item embedding vector and the behavior type embedding vector corresponding to the current behavior. The hidden state update is calculated by updating the gate, resetting the gate and candidate hidden states, and the final hidden state is output as the user's dynamic interest vector. The target item feature data is subjected to structured embedding processing to generate an item static attribute vector; The external context information is discretized and normalized to generate a context-aware vector; The user dynamic interest vector, item static attribute vector, and context-aware vector are input into a multi-head cross-attention network to calculate the joint interaction representation of the user-item-context triple.

3. The big data recommendation and conversion prediction method based on deep learning according to claim 2, characterized in that, The target item feature data is subjected to structured embedding processing to generate an item static attribute vector, including: Categorical features are mapped to dense vectors using independent embedding matrices; numerical features are transformed logarithmically and then projected through fully connected layers; textual features are extracted using character-level convolutional neural networks before pooling and projection. The individual sub-vectors are concatenated and then normalized to form the static attribute vector of the item.

4. The big data recommendation and conversion prediction method based on deep learning according to claim 3, characterized in that, Discretizing and normalizing the external context information to generate a context-aware vector includes: Discretize the time field by hour, day of the week, and whether it is a holiday and embed them separately; split the geolocation GeoHash string into a front part and a back part and embed them independently; map the device type and network environment to a fixed-dimensional vector by looking up a table; All context sub-vectors are concatenated and compressed to a predetermined dimension through a fully connected layer to form a context-aware vector.

5. The big data recommendation and conversion prediction method based on deep learning according to claim 4, characterized in that, The user dynamic interest vector, item static attribute vector, and context-aware vector are input into a multi-head cross-attention network to calculate the joint interaction representation of the user-item-context triples, including: Within each attention head, a query vector is generated from the user's dynamic interest vector, and a key vector and a value vector are generated by concatenating the item's static attribute vector and the context-aware vector. Attention weights are calculated and the outputs of each head are summed in a weighted manner, and then fused into a joint interaction representation through a fully connected network.

6. The big data recommendation and conversion prediction method based on deep learning according to claim 5, characterized in that, Based on a pre-defined business rule base and a common-sense logic knowledge graph, a symbolic logic constraint set is constructed, including: The explicit rules are stored in the form of first-order predicate logic. The explicit rules include "the same user shall not purchase prescription drugs repeatedly within 24 hours", "high-value products are only recommended to users with credit ratings higher than the threshold", and "maternal and infant products shall not be actively pushed to male users". Implicit common sense is stored in the common sense logic knowledge graph in the form of triples, where nodes are entities or events, edges are semantic relations, and weights represent the strength of common sense.

7. The big data recommendation and conversion prediction method based on deep learning according to claim 6, characterized in that, The symbolic logic constraint set is transformed into a differentiable soft constraint function, and this function is embedded as a regularization term into the loss function of the multi-head cross-attention network, including: Calculate the satisfaction score of each logical rule in the current recommendation context; A penalty term is constructed using hinge loss, and a penalty is applied when the satisfaction score is lower than a preset tolerance threshold. The weighted sum of the penalty terms of all rules is used as a logical regularization term, which is then combined with the weighted loss of the main task to form a joint optimization objective.

8. The big data recommendation and conversion prediction method based on deep learning according to claim 7, characterized in that, Synchronously generate the corresponding logical interpretation path, including: Traverse the symbolic logic constraint set and identify the activated logic rule numbers; By tracing back the attention weight distribution, we can identify the key behavioral events in the user behavior sequence that support the triggering of this logical rule. Extract the attribute fields from the product features that match the rule conditions; The activation rule number, key behavior events, and matching attribute fields are organized into a natural language readable explanatory text in a cause-effect structure.

9. The big data recommendation and conversion prediction method based on deep learning according to claim 8, characterized in that, Construct a conversion prediction sub-model, including: Using a gradient boosting decision tree as the model structure, the input features include recommendation score, explanation path length, number of rules contained in the explanation, user conversion rate in the past 7 days, current session duration, device screen size, network latency, item price to user's historical average consumption ratio, remaining inventory ratio, and promotion countdown. The model is trained incrementally daily, using whether a user completes a purchase within 24 hours of a recommendation being displayed as a tag.

10. The big data recommendation and conversion prediction method based on deep learning according to claim 9, characterized in that, The initial recommendation list is then reordered based on the conversion probability, including: Calculate the explanation completeness score for each item, defined as the ratio of the number of rules covered by the explanation path to the total number of activated rules; The final ranking score is calculated using a weighted fusion strategy. , Recommended score, For conversion probability, To interpret the completeness score, , , These are configurable weighting coefficients, and their sum is 1.

Images