Business opportunity recommendation method and device based on deep learning
By employing a deep learning-based business opportunity recommendation method, leveraging multimodal data and real-time state awareness, we have addressed the shortcomings of traditional methods, such as insufficient semantic understanding and cold start issues. This approach enables more accurate and real-time business opportunity recommendations, enhancing user experience and system adaptability.
Patent Information
- Application Number
- CN202511465277.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-10-14
- Publication Date
- 2026-01-16
Smart Images
Figure CN121350752A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of computer technology, and in particular to a business opportunity recommendation method and apparatus based on deep learning. Background Technology
[0002] In Customer Relationship Management (CRM) systems, opportunity recommendation is a core function that helps users (such as sales personnel) quickly identify potential customers and improve sales conversion rates. However, existing traditional methods based on rule matching or two-dimensional matrices lack deep semantic parsing capabilities, leading to biases in similarity calculations. In particular, methods based on two-dimensional matrices only consider structured data, ignoring unstructured data such as text descriptions, resulting in an inability to effectively integrate multimodal features. Furthermore, existing methods are sensitive to the cold start problem of new opportunities or users and have weak generalization capabilities. Therefore, there is an urgent need for an opportunity recommendation solution that can fully mine multimodal data features and possess deep semantic understanding capabilities. Summary of the Invention
[0003] In view of this, embodiments of the present invention provide a business opportunity recommendation method and apparatus based on deep learning, which can at least solve the problems of limited semantic understanding ability, insufficient data utilization, and prominent cold start problem in the prior art.
[0004] To achieve the above objectives, according to one aspect of the present invention, a business opportunity recommendation method based on deep learning is provided, comprising:
[0005] Acquire multimodal data of target business opportunities; wherein, the multimodal data includes structured data and unstructured data;
[0006] The multimodal data is input into a pre-trained model to encode semantic vectors;
[0007] Based on the semantic vector, a list of similar business opportunities is determined, and business opportunities with valid current status information are selected from the list of similar business opportunities for recommendation.
[0008] To achieve the above objectives, according to another aspect of the present invention, a business opportunity recommendation device based on deep learning is provided, comprising:
[0009] An acquisition module is used to acquire multimodal data of target business opportunities; wherein, the multimodal data includes structured data and unstructured data;
[0010] The encoding module is used to input the multimodal data into the pre-trained model to encode semantic vectors;
[0011] The filtering module is used to determine a list of similar business opportunities based on the semantic vector, and to filter out business opportunities whose current status information is valid from the list of similar business opportunities for recommendation.
[0012] To achieve the above objectives, according to another aspect of the present invention, a business opportunity recommendation electronic device based on deep learning is provided.
[0013] The electronic device of this invention includes: one or more processors; and a storage device for storing one or more programs, which, when executed by the one or more processors, enable the one or more processors to implement any of the deep learning-based business opportunity recommendation methods described above.
[0014] To achieve the above objectives, according to another aspect of the present invention, a computer-readable medium is provided having a computer program stored thereon, which, when executed by a processor, implements any of the deep learning-based business opportunity recommendation methods described above.
[0015] To achieve the above objectives, according to another aspect of the present invention, a computing program product is provided. One computing program product of the present invention includes a computer program, which, when executed by a processor, implements the deep learning-based business opportunity recommendation method provided in the present invention.
[0016] According to the solution provided by the present invention, one embodiment of the invention has the following advantages or beneficial effects: by acquiring multimodal data of target business opportunities and using a pre-trained model for semantic encoding, it is possible to deeply understand the structured and unstructured features of target business opportunities, thus exhibiting stronger semantic understanding capabilities compared to traditional methods. Simultaneously, through real-time state awareness and dynamic filtering mechanisms, invalid business opportunities are automatically eliminated, ensuring the timeliness and practicality of business opportunity recommendation results. This effectively solves the problem in traditional recommendation methods where recommendation results include already bound or expired business opportunities, thereby improving user experience and recommendation accuracy.
[0017] The further effects of the aforementioned unconventional alternative methods will be explained below in conjunction with specific implementation methods. Attached Figure Description
[0018] The accompanying drawings are provided to better understand the invention and are not intended to unduly limit the scope of the invention. Wherein:
[0019] Figure 1 This is a schematic diagram of the main process of a business opportunity recommendation method based on deep learning according to an embodiment of the present invention;
[0020] Figure 2 This is a flowchart illustrating an optional deep learning-based business opportunity recommendation method according to an embodiment of the present invention.
[0021] Figure 3 This is a flowchart illustrating the process of processing structured and unstructured data.
[0022] Figure 4 This is a flowchart illustrating another optional deep learning-based business opportunity recommendation method according to an embodiment of the present invention;
[0023] Figure 5 This is a flowchart illustrating the process of generating positive and negative samples;
[0024] Figure 6 This is a flowchart illustrating another optional deep learning-based business opportunity recommendation method according to an embodiment of the present invention;
[0025] Figure 7 This is a schematic diagram of the main modules of a business opportunity recommendation device based on deep learning according to an embodiment of the present invention;
[0026] Figure 8 This is an exemplary system architecture diagram in which embodiments of the present invention can be applied;
[0027] Figure 9 This is a schematic diagram of the structure of a computer system suitable for implementing the embodiments of the present invention, such as a mobile device or server. Detailed Implementation
[0028] The following description, in conjunction with the accompanying drawings, illustrates exemplary embodiments of the present invention, including various details to aid understanding. These details should be considered merely exemplary. Therefore, those skilled in the art will recognize that various changes and modifications can be made to the embodiments described herein without departing from the scope and spirit of the invention. Similarly, for clarity and brevity, descriptions of well-known functions and structures are omitted in the following description.
[0029] It should be noted that the collection, gathering, updating, analysis, processing, use, transmission, and storage of user personal information involved in this disclosed technical solution all comply with relevant laws and regulations, are used for legitimate purposes, and do not violate public order and good morals. Necessary measures are taken to prevent unauthorized access to user personal information data and to safeguard user personal information security, network security, and national security.
[0030] Business opportunities are essentially customer needs, such as a structured or unstructured description of the need for 10 computers, including their models and memory. The main current problem is that although a large number of business opportunities are collected, most are customer inquiries, with relatively few truly effective opportunities. Furthermore, opportunity collection typically focuses on basic customer information, while actual processing emphasizes specific product details. Therefore, to improve the efficiency of browsing and processing business opportunities, the system needs recommendation capabilities: after a user processes a current opportunity, they should be able to directly access and recommend similar opportunities, avoiding frequent searching and querying. Currently, the two main methods for business opportunity recommendation are as follows:
[0031] Option 1: Recommendation method based on rule or keyword matching. This method manually sets classification rules (such as industry category, customer size) or extracts keywords from business opportunity descriptions (such as "cloud computing" and "finance") to calculate the field matching degree between the current business opportunity and other business opportunities, thereby filtering out similar business opportunities. However, this method has obvious limitations: 1) Its semantic understanding ability is insufficient, and it cannot effectively distinguish synonyms (such as "AI" and "artificial intelligence") or handle polysemy (such as "cloud" may refer to "cloud computing" or "cloud storage"), resulting in frequent mismatches; 2) The rules and keywords rely on expert experience to formulate and maintain, making it difficult to adapt to dynamic business needs (such as the rapid iteration of emerging industry classifications); 3) For business opportunity information containing complex semantic details, this method has weak long text processing capabilities, making it difficult to accurately capture the key semantics, resulting in the loss of key information.
[0032] Option 2: Recommendation methods based on collaborative filtering or two-dimensional matrix similarity. This method transforms opportunity features (such as customer industry and budget amount) into a two-dimensional matrix and uses algorithms such as cosine similarity and Jaccard coefficient to calculate the similarity between opportunities, or makes recommendations based on historical user behavior data (such as binding records). This method also has certain limitations: 1) Two-dimensional matrices are difficult to effectively integrate multiple modal features such as text and numerical values, and are sensitive to sparse data (such as newly added opportunities), resulting in sparsity issues in feature representation; 2) The cold start problem is prominent, and the recommendation effect is significantly reduced due to the lack of historical behavior data for new opportunities or new users; 3) Poor real-time performance, as matrix updates and similarity recalculation usually rely on offline batch processing, which cannot respond promptly to changes in opportunity status (such as binding status updates), causing recommendation results to lag. It should be noted that the matrix dimension is determined by the total number of users or opportunities and the number of features, which are usually stable and do not increase or decrease rows and columns due to changes in the status of a single opportunity. What truly changes dynamically are the matrix element values. However, due to the high complexity and strong data dependency of real-time similarity calculation, the timed update mechanism may lead to consistency issues (such as conflicts caused by concurrent modifications).
[0033] In addition, both of the above methods share the following common drawbacks: First, there is a significant semantic gap, as traditional technologies struggle to deeply understand the contextual semantics of business opportunity texts, leading to biases in similarity calculations (e.g., misjudging "AI solutions" and "smart hardware" as highly similar); second, there is a lack of dynamic filtering mechanisms, as the recommendation process does not consider the real-time status of business opportunities (e.g., whether they have been bound or are still valid), resulting in the easy recommendation of invalid business opportunities; finally, there is insufficient scalability, making it difficult to adapt to complex scenarios such as multilingual and multimodal data (e.g., unstructured + structured data), limiting the system's generalization ability and future development potential.
[0034] See Figure 1 The diagram shows the main flowchart of a business opportunity recommendation method based on deep learning provided by an embodiment of the present invention, which includes the following steps:
[0035] S101: Obtain multimodal data of the target business opportunity; wherein, the multimodal data includes structured data and unstructured data;
[0036] S102: Input the multimodal data into the pre-trained model to encode semantic vectors;
[0037] S103: Based on the semantic vector, determine a list of similar business opportunities, and select business opportunities with valid current status information from the list of similar business opportunities for recommendation.
[0038] In the above implementation, for step S101, the user selects a target business opportunity through the CRM system interface, or after processing the target business opportunity, triggers the system to respond to this operation and automatically extracts the unique identifier ID of the selected target business opportunity. Based on this ID, the system retrieves the multimodal data of the target business opportunity from the business opportunity database. This multimodal data mainly includes two categories: structured data and unstructured data.
[0039] Structured data refers to information with a clearly defined format and organizational structure, such as industry classifications, budget amounts, customer size, and geographical location. This data is typically stored in key-value pairs, facilitating direct computer processing and analysis. Unstructured data, on the other hand, refers to information without a predefined organizational structure, such as requirement descriptions, customer backgrounds, business context descriptions, and communication records. This data exhibits characteristics of natural language, similar to spoken language, and requires parsing and understanding using natural language processing (NLP) techniques. By integrating these two different types of data, the system can comprehensively understand and characterize the features of target business opportunities.
[0040] In step S102, the system inputs the acquired multimodal data of the target business opportunity into a pre-trained model (or multimodal business opportunity encoding model) to generate a semantic vector for the target business opportunity. The pre-trained model is an end-to-end learning framework based on deep learning, capable of automatically learning text features and gaining deep semantic understanding through training on large-scale corpora. The model includes an input layer and an encoding layer. The input layer receives and processes multimodal data to form a sequence representation. Specifically, it concatenates structured and unstructured data into a unified sequence, such as [CLS] Requirement Description: {text} [SEP] Industry: {industry} [SEP] Budget: {budget} [SEP]. The encoding layer initializes based on BERT-base-chinese, encodes the sequence, and finally outputs a semantic vector representation, such as outputting a [CLS] position vector as the semantic vector representation of the business opportunity (e.g., 768-dimensional).
[0041] This sequence generation and encoding operation effectively solves the problem of similarity calculation bias caused by the inability of existing methods to deeply understand the contextual semantics of business opportunity text. For example, traditional methods cannot accurately handle synonym phenomena (such as "AI" and "artificial intelligence") and eliminate the interference of polysemy (such as "cloud" referring to "cloud computing" or "cloud storage"). This invention, through deep semantic encoding technology, can accurately capture the contextual semantic information of business opportunity text, achieving an accurate understanding of the deeper meaning of business opportunity data, thereby significantly improving the accuracy of semantic matching and recommendation performance.
[0042] In one optimized implementation, after generating the sequence, the input layer also needs to normalize the sequence length to ensure it meets a preset requirement (e.g., 512 tokens). When the sequence length is less than the preset length, the system pads it using methods such as adding zeros or repeating data. When the sequence length exceeds the preset length, the system performs truncation to retain the core content. The core content can be determined in the following ways: identifying high-frequency words in the sequence using TF-IDF or TextRank algorithms, retaining sentences containing high-frequency words, and removing sentences containing low-frequency words; or identifying sentences that hit preset key fields. Through the above length normalization process, the input sequence is ensured to meet the preset length requirement, thereby guaranteeing the stability and accuracy of the model processing.
[0043] The model can be selected based on the actual situation. This solution preferably uses BERT (Bidirectional Encoder Representations from Transformers) as the base model. It should be noted that BERT belongs to a broad category of pre-trained language models, which includes various variant models. The specific variant model chosen in this solution is not limited; only BERT is used as an example. In practical applications, the BERT model needs to be fine-tuned to effectively integrate the multimodal data features of business opportunities. The specific fine-tuning process will be discussed later. Figure 4 The detailed description shown is omitted here.
[0044] In step S103, after obtaining the semantic vector of the target business opportunity, the system retrieves its similarity to the semantic vectors of other business opportunities in the business opportunity database. A certain threshold is set, such as 0.8 (configurable). Similarity exceeding this value indicates similarity, thus obtaining a list of similar business opportunity IDs. To improve retrieval efficiency, the system can set up a vector database to store the semantic vector of each business opportunity in the database, facilitating rapid retrieval and matching. Alternatively, the business opportunity database and the vector dataset can be integrated into one system; there are no specific restrictions. The system can sort the obtained list of similar business opportunity IDs from highest to lowest similarity and extract the Top-K similar business opportunity IDs as recommendation candidates.
[0045] For similarity calculation, this scheme preferably uses the cosine similarity algorithm. Alternatively, other distance metrics such as Euclidean distance and Manhattan distance can be used. It's important to note that Euclidean distance is sensitive to vector scale and may be more suitable for vector calculations before normalization, but this requires a redesigned vector normalization strategy, which may deviate from the matching degree with the output characteristics of the BERT model. Considering both computational performance and implementation complexity, this scheme preferably uses the cosine similarity method for similarity calculation.
[0046] In terms of retrieval, this solution can also use Annoy and HNSW (Hierarchical Navigable SmallWorld) to replace Faiss for approximate nearest neighbor search. Annoy is more suitable for memory-constrained scenarios, with lower memory usage; HNSW supports dynamic incremental indexing, better adapting to the real-time update requirements of data. However, different retrieval algorithms have different trade-offs between recall and search speed, requiring re-evaluation and selection based on specific business scenarios and performance requirements to achieve optimal retrieval results.
[0047] Traditional business opportunity recommendation methods do not link with the real-time status information of business opportunities, resulting in the inclusion of invalid business opportunities that have already been bound by other users or have expired, affecting recommendation effectiveness and user experience. To address this, this invention proposes a real-time status awareness and dynamic filtering mechanism that automatically removes invalid business opportunities during the recommendation process, reducing the cost of manual secondary screening and improving the practicality and accuracy of the recommendation results.
[0048] Specifically, the system retrieves multimodal data and status information for each similar business opportunity from the business opportunity database. The status information includes key attributes such as binding status and validity period. Considering that a user can bind multiple business opportunities, but a business opportunity is typically bound to only one user, the system only recommends business opportunities with valid current status information—that is, business opportunities that are not currently bound by other users and are within their validity period—and displays the multimodal data of these valid business opportunities. When the current status information of all similar business opportunities is invalid (e.g., all are bound or expired), the system will trigger a corresponding prompt message, informing the user that there are currently no available business opportunities to recommend, thereby avoiding invalid recommendations and improving user satisfaction.
[0049] As can be seen, this solution employs a separate design for similarity calculation and status filtering. Real-time status is queried via a lightweight API (Application Programming Interface), avoiding the latency issues caused by traditional offline updates and ensuring the timeliness of recommendation results. Furthermore, similarity thresholds and filtering conditions (such as binding status and validity period) support dynamic configuration, flexibly adapting to changing needs in different business scenarios and improving the system's scalability and business adaptability.
[0050] It is understood that the specific values listed in the above hypothetical description of business opportunity recommendation based on deep learning are merely illustrative examples and not restrictive. In practice, other values can be set according to needs. Similarly, the specific values listed in the subsequent embodiments are also just examples and are not the only limitation.
[0051] The method provided in the above embodiments, by acquiring multimodal data of target business opportunities and using pre-trained models for semantic encoding, can deeply understand the structured and unstructured features of target business opportunities, exhibiting stronger semantic understanding capabilities and higher recommendation accuracy compared to traditional methods. Simultaneously, through real-time state awareness and dynamic filtering mechanisms, invalid business opportunities are automatically eliminated, ensuring the timeliness and practicality of business opportunity recommendation results, effectively solving the problem in traditional recommendation methods where recommended results include already bound or expired business opportunities.
[0052] Furthermore, existing similarity matrix update mechanisms based on offline batch calculation cannot respond in real time to changes in business opportunity status (such as binding status updates), resulting in delayed recommendation results. To address this real-time insufficiency caused by delayed recommendation results, this invention utilizes online real-time semantic encoding and similarity calculation technology to achieve millisecond-level response, ensuring that recommendation results are dynamically synchronized with business scenarios, thereby significantly improving user experience and recommendation efficiency.
[0053] Furthermore, existing rule-matching methods rely on manually defined keywords and industry classification rules, making it difficult to adapt to dynamic business needs (such as the emergence of new industries) and resulting in high manual maintenance costs. To address this issue, this invention employs automated semantic modeling technology, applicable to historical business opportunities, new business opportunities, and both new and existing users. This reduces manual intervention, enabling the system to adapt to business changes, significantly lowering maintenance costs and improving system scalability.
[0054] See Figure 2 The diagram illustrates an optional deep learning-based business opportunity recommendation method according to an embodiment of the present invention, comprising the following steps:
[0055] S201: Extract the first field value corresponding to the first preset field from the unstructured data, perform word segmentation on the first field value to determine high-frequency words according to their frequency of occurrence, and establish a first semantic unit based on the first preset field and the sentence in which the high-frequency words are located.
[0056] S202: Extract the second field value corresponding to the second preset field from the structured data, determine the type of the second field value, process the second field value using the strategy corresponding to the type, and establish a second semantic unit based on the second preset field and the processed second field value;
[0057] S203: The first semantic unit and the second semantic unit are concatenated to obtain a target sequence, and the target sequence is encoded to obtain a semantic vector.
[0058] This implementation describes how to generate semantic vectors based on multimodal data encoding. For step S201, regarding the processing of unstructured data, the first field value under the first preset field is extracted from the unstructured data. Assuming the first preset field is requirement description and customer background, the extraction result can be: Requirement description: A financial company needs cloud computing services with a budget of 1 million; Customer background: A listed company with 500 employees.
[0059] After the first field value is extracted, the system performs a series of preprocessing operations on it. First, data cleaning is carried out, including removing HTML tags, special symbols, and stop words (such as common function words like "of" and "is"); then term standardization is performed to unify different expressions of the same concept, for example, unifying "AI" into "artificial intelligence";接着, a tokenizer such as Jieba or BERT-wwm is used to tokenize the preprocessed first field value, and high-frequency words are identified through algorithms such as TF-IDF or TextRank, and then the sentences where the high-frequency words are located are determined and retained. Finally, based on the first preset field and the processed first field value, a first semantic unit is constructed, for example, the requirement description: A certain financial company needs cloud computing services with a budget of 1 million.
[0060] For step S202, for the processing of structured data, first, the second field value corresponding to the second preset field needs to be extracted from it, such as keyword fields like industry and budget. Subsequently, the system analyzes the data type of the second field value and adopts a strategy matching this type to process the second field value, and finally, a second semantic unit is established based on the second preset field and the processed second field value.
[0061] When the field value type is numerical (such as budget amount, creation timestamp, etc.), the system uses normalization or discretization processing methods. Normalization processing maps the numerical value to the interval [0, 1], for example, dividing the budget amount by the preset maximum budget value to achieve standardization; discretization processing buckets the continuous values, for example, dividing the budget into discrete intervals such as "<500,000", "500,000 - 1 million", ">1 million".
[0062] When the field value type is a categorical field (such as industry classification, customer scale, etc.), the system uses a categorical feature encoding method for processing. Among them, the categorical feature encoding methods are such as label encoding or one-hot encoding. Label encoding is used to convert categorical variables into numerical forms of 0, 1, 2, 3..., and each category is identified by a unique integer, which is suitable for ordered categorical variables (such as small, medium, large, etc.). For example, the original data: ["red", "blue", "green", "red"], after encoding: [0, 1, 2, 0]. One-hot encoding is used to convert categorical variables into binary vector representations, and each category corresponds to a new feature column, which is suitable for unordered categorical variables. After encoding the above example, it is obtained: red: [1, 0, 0], blue: [0, 1, 0], green: [0, 0, 1], red: [1, 0, 0].
[0063] In step S203, the system concatenates the first and second semantic units based on a preset delimiter to generate a standardized target sequence of target business opportunities. The model's encoding layer directly processes this target sequence, encoding it into a 768-dimensional business opportunity semantic vector representation. For the overall implementation process, please refer to [link to implementation details]. Figure 3 As shown.
[0064] The sequence employs a specific tagging structure: [CLS] (Classification Token) serves as the sequence start tag, and [SEP] (Separator Token) is used as the separator between semantic units. In Transformer-type models such as BERT, the hidden state at the [CLS] position is typically used as the semantic representation of the entire sequence, while [SEP], as a special separator tag, clearly marks the boundaries of each semantic unit, effectively avoiding confusion between fields (e.g., preventing the model from merging "Budget: 1 million" and "Industry: Finance"). Without explicit separators, the model may not accurately perceive field boundaries, especially when processing long texts, which can easily lead to misunderstandings.
[0065] The concatenation order can be flexibly set according to actual needs, or a fixed format can be used, for example: [CLS] Requirement Description: {Requirement Text} [SEP] Industry: {Industry Classification} [SEP] Budget: {Budget Amount} [SEP] Customer Size: {Size Label} [SEP] Customer Background: {Background Information} [SEP]. A specific concatenation result is as follows: [CLS] Requirement Description: A financial company needs cloud computing services, with a budget of 1 million [SEP] Industry: Finance [SEP] Budget: 100 [SEP] Customer Size: Large Enterprise [SEP] Customer Background: Approximately 500,000 entries [SEP]. This standardized sequence construction method can convert multimodal business opportunity data into a unified input format that the model can process, laying the foundation for subsequent tasks such as semantic similarity calculation and classification.
[0066] The method provided in the above embodiments effectively integrates semantic information from multimodal data by extracting key fields from unstructured data and constructing semantic units based on high-frequency words, combined with typified processing of structured data, thus achieving a more accurate representation of business opportunity feature sequences. Simultaneously, the use of sequence concatenation and encoding enhances the model's adaptability to different multimodal data, and the resulting semantic vectors better capture the essential features of business opportunities, providing a more reliable semantic foundation for subsequent similarity calculations and recommendations.
[0067] See Figure 4 The diagram illustrates another optional deep learning-based business opportunity recommendation method according to an embodiment of the present invention, including the following steps:
[0068] S401: Obtain a sample set and divide the sample set into a training set, a validation set, and a test set; wherein, the samples are divided into positive samples and negative samples, and each sample consists of a business opportunity pair;
[0069] S402: Train the preset model using the training set, and adjust the model parameters while keeping the model hyperparameters unchanged to obtain the first model;
[0070] S403: Train the first model using the validation set, and adjust the model hyperparameters while keeping the model parameters unchanged to obtain the second model;
[0071] S404: Test the second model using the test set, and in response to the test results meeting the preset conditions, use the second model as the pre-trained model.
[0072] This implementation describes the model training process. First, a sample set is obtained, which can be pairs of all business opportunities in the business opportunity database. This solution first acquires multimodal data and behavioral data for each business opportunity. The multimodal data includes structured data (such as industry classification, budget amount, customer size, etc.) and unstructured data (such as demand description, customer background, etc.), while the behavioral data includes user binding records, business opportunity conversion status, etc.
[0073] Unlike existing technologies, the samples in this scheme consist of opportunity pairs, such as opportunity pairs consisting of opportunities A and B. The sample generation process is as follows: historical opportunities are paired to obtain multiple opportunity pairs. Opportunity pairs that are bound to the same user and have all successfully converted are used as positive samples, and the remaining opportunity pairs are used as negative samples. The reason for adopting this sample construction strategy is that opportunities processed by the same user have high similarity in actual business, but these opportunities may not be close in terms of semantic distance and semantic similarity. Therefore, the model training goal of this scheme is to improve the similarity of opportunities in positive samples and reduce the similarity of opportunities in negative samples, so that the model can more accurately identify truly similar opportunities.
[0074] Suppose the business opportunity database includes business opportunities A, B, C, D, E, and F. Salesperson X is associated with business opportunities A, B, and C, and all three business opportunities have been successfully converted (e.g., a successful contract signing). Then, pairwise combinations of A, B, and C yield three business opportunity pairs: AB, AC, and BC, which are the three positive samples. The remaining business opportunity pairs are considered negative samples. As an optimized implementation method, the range of negative samples can be further refined and narrowed down, specifically into three cases, thereby improving the relevance and effectiveness of model training. See [link to details] for further information. Figure 5 As shown:
[0075] 1) Cross-user negative samples: Business opportunities that are bound to different users and have all successfully converted are paired up to generate negative sample business opportunity pairs;
[0076] 2) Random Negative Samples: Opportunities from different industries (i.e., the third preset field) and / or budget ranges (i.e., the fourth preset field) are randomly selected and paired to generate negative sample opportunity pairs. Budget ranges include <500,000, 500,000-1,000,000, and >1,000,000. Opportunities from different industries or budget ranges can be considered, or opportunities from both different industries and budget ranges can be considered; there are no specific restrictions. Furthermore, to control the size of the negative sample, a certain number of opportunities can be randomly selected from these eligible opportunities and paired again.
[0077] 3) Difficult negative sample: Business opportunity pairs in the same or similar industries but not converted. For example, business opportunity D and business opportunity E are in the same industry but have very different demand descriptions and neither has been converted. They can be used as difficult negative sample business opportunity pairs.
[0078] The sample set needs to be divided into training, validation, and test sets according to a certain ratio. For example, 80% of the sample set is used as the training set to update model parameters, including the weight matrix (W), bias vector (b), and attention head parameters of BERT in the neural network. 10% of the sample set is used as the validation set for hyperparameter tuning and early stopping, including hyperparameters such as learning rate, batch size, hidden layer dimension, dropout rate, and number of training epochs. The remaining 10% of the sample set is used as the test set for evaluating the final results.
[0079] It's important to note that during the training phase, the system only updates the model parameters, keeping the hyperparameters constant. During the validation phase, only hyperparameters are adjusted or early stopping conditions are monitored; no model parameters are updated. During the testing phase, all parameters and hyperparameters remain fixed and are used solely to evaluate model performance; for example, the model optimization process can be stopped when the test accuracy reaches 80%. This phased training strategy effectively avoids overfitting and ensures the model's generalization ability and stability.
[0080] The early stopping monitoring mechanism involves continuously monitoring performance metrics on the validation set during model training to determine whether training should be terminated prematurely, thus preventing overfitting. Specifically, the system evaluates the validation set performance after each training epoch. Early stopping is triggered when validation set metrics show no significant improvement or a downward trend over several consecutive epochs. The optimal performance value is recorded as a benchmark, ensuring timely training stoppage before overfitting begins, thereby saving computational resources and improving model generalization ability.
[0081] As Figure 2 and Figure 4 One optimized implementation involves data augmentation after preprocessing business opportunity data but before inputting it into the model to enhance its generalization ability. A pre-defined thesaurus can be used to replace words in semantic units with synonyms, such as "cloud computing" → "cloud service," increasing data diversity while preserving core semantics. For structured data, perturbation can be applied, such as adding ±10% random noise to numerical fields (e.g., budget 1 million → 900,000 or 1.1 million), and randomly replacing categorical fields with values from similar categories (e.g., "finance" → "bank"). This data augmentation strategy effectively expands the diversity of training samples, improving the model's adaptability and robustness to different types of data.
[0082] The method provided in the above embodiments constructs positive and negative samples in the form of business opportunity pairs, dividing the sample set into training, validation, and test sets for hierarchical model training. This achieves separate optimization of model parameters and hyperparameters, avoiding overfitting. Hyperparameter tuning and early stopping monitoring are performed using the validation set to ensure the model stops training in its optimal state, improving its generalization ability and stability. Finally, the model performance is validated using the test set, ensuring the accuracy and reliability of the recommendation results, thereby significantly improving the overall effect of business opportunity recommendations and user experience.
[0083] Furthermore, existing technologies, due to their limited feature representation capabilities, struggle to effectively integrate multimodal data such as text, numerical values, and classification labels (e.g., "budget amount" and "industry description"), and are sensitive to the cold start problem of new business opportunities or new users (lacking historical behavioral data). To address this multimodal data fusion and cold start issue, this invention constructs a multimodal business opportunity coding model, jointly modeling structured and unstructured business opportunity data. This significantly improves the accuracy of new business opportunity recommendations, effectively alleviates the cold start problem, and enhances the system's adaptability to new business scenarios.
[0084] See Figure 6 The diagram illustrates another optional deep learning-based business opportunity recommendation method according to an embodiment of the present invention, including the following steps:
[0085] S601: Determine all positive samples and all negative samples corresponding to the first business opportunity, determine the second business opportunity in the positive samples other than the first business opportunity, and the third business opportunity in the negative samples other than the first business opportunity;
[0086] S602: Based on the semantic vectors of the first business opportunity, the second business opportunity, and the third business opportunity, calculate the triplet loss of the first business opportunity;
[0087] S603: Based on the semantic vector similarity between the first business opportunity and the second business opportunity, and the semantic vector similarity between the first business opportunity and the third business opportunity, calculate the deep learning loss of the first business opportunity;
[0088] S604: Determine the mask modeling loss of the first business opportunity, and combine the triplet loss and the deep learning loss to obtain the total loss of the first business opportunity;
[0089] S605: Train the model with the goal of minimizing the total loss of all business opportunities.
[0090] This implementation describes how to train a model with the goal of minimizing the total loss. For steps S601-S602, the formula for calculating the Triplet Loss is designed as follows:
[0091]
[0092] in, Let p be the semantic vector of the anchor opportunity, and n be the set of positive and negative samples corresponding to the anchor opportunity, respectively. The semantic vector of the positive sample. The semantic vector of the negative sample. This is the margin hyperparameter. It is a commonly used loss function in deep learning, mainly used to measure the similarity between samples. It trains the model by considering three samples simultaneously (anchor, positive sample, and negative sample), making samples of the same class closer in the feature space and samples of different classes farther apart.
[0093] Taking the first business opportunity A as an example, we determine the positive samples (AB) and negative samples (AD) in which the first business opportunity A is located. We calculate the semantic distance (or semantic vector similarity, where smaller the semantic distance indicates greater similarity) between AB and the semantic distance between AD, and then calculate the triplet loss value for the first business opportunity A based on the triplet loss function. When optimizing the model parameters using this loss function, we need to ensure that the distance between A and the positive sample B is less than the distance between A and the negative sample D, and that the difference between the two is not less than the margin hyperparameter. This enhances the model's semantic discrimination ability. When the number of positive and negative samples corresponding to opportunity A is large, the triplet loss for opportunity A can be the average of the obtained triplet losses.
[0094] For step S603, the formula for calculating the deep learning loss Circle Loss is as follows:
[0095]
[0096] in, Let represent the semantic vector similarity between the i-th positive sample and the anchor opportunity. This represents the semantic vector similarity between the j-th negative sample and the anchor opportunity. γ is a scaling factor (hyperparameter) used to control the strength of the gradient, typically set to 80. The margin hyperparameter m (typically set to 0.25) is also set. =1-m, =m. =max(0, - ), =max(0, - ).
[0097] Taking the aforementioned business opportunity A as an example, by calculating its semantic vector similarity with all positive samples and with all negative samples, and combining this with the parameters mentioned above, the deep learning loss for business opportunity A can be obtained. Specifically, the deep learning loss measures the model's ability to classify positive and negative samples, and needs to comprehensively consider the similarity relationship between the anchor business opportunity and all relevant samples, rather than just the similarity with a single positive or negative sample.
[0098] When the number of both positive and negative samples corresponding to business opportunity A is large, the deep learning loss for business opportunity A can be the average of all deep learning losses. By averaging the losses of multiple sample pairs, the stability and representativeness of the loss calculation are improved. It is worth noting that Circle Loss, as an improved loss function, can accelerate model convergence and improve learning performance by applying higher weights to difficult samples.
[0099] For step S604, this scheme also sets an auxiliary loss function, mainly considering mask modeling loss, to improve the robustness of semantic understanding by reconstructing the input sequence through the decoder (MLM (Masked Language Model) task). As an optional implementation, classification loss can also be introduced to predict the probability of business opportunity conversion (such as success / failure) through binary classification cross-entropy, further enhancing the discriminativeness and business relevance of semantic representation.
[0100] For masking modeling loss, random masking can be performed during sequence generation, such as replacing non-critical entities (e.g., company names) with a 15% probability [MASK]. Therefore, masking is typically used for unstructured data. The masked sequence is input into the model, and through multiple layers of processing, a corresponding hidden state vector is generated for each position. For each masked position, the hidden state vector is used to predict the word distribution through a linear transformation, outputting a logits vector with dimensions equal to the vocabulary size (each element in the logits vector represents the original score of the corresponding word at that position, which is converted to a probability distribution using softmax for word prediction). Then, the cross-entropy loss between the predicted and true words is calculated to obtain the masking modeling loss for business opportunities.
[0101] The cross-entropy loss is calculated as Loss = -Σ(yi * log(pi)), which only calculates the loss at the masked positions. Other unmasked positions are not included in the calculation. yi is the one-hot encoding of the true word at the masked position, and pi is the probability of each word predicted by the model at that position (after softmax). The cross-entropy loss is calculated separately for each masked position, then summed and averaged. Assuming that opportunity A has three masked positions, its masking modeling loss is the average of the losses at these three positions.
[0102] The MLM task forces the model to deeply understand the relationship between unstructured and structured input data. This is achieved by: masking the input sequence → BERT encoding → predicting the masked words → calculating cross-entropy loss. This approach improves the model's robustness to noise, missing fields, and cold-start scenarios, enhancing its generalization ability.
[0103] For step S605, this scheme learns the objective function through multiple tasks, and the formula for calculating the total loss function of the model is as follows:
[0104]
[0105] in, The total loss function is the target of the final optimization during model training. The triplet loss is used to shorten the distance between opportunity pairs in positive samples and widen the distance between opportunity pairs in negative samples, thereby enhancing the discriminative power of semantic vectors. It is a deep learning loss used to supervise the model's prediction of business opportunity categories and improve classification accuracy. It is a masking modeling loss, which enhances the model's ability to understand the context by predicting the masked words or fields.
[0106] Its weight is 0.7, which plays a dominant role in the overall optimization. The weight is 0.2, which is used to help optimize the ability to identify business opportunity categories. The weight is 0.1, used to improve the model's language understanding ability. If the masking loss for business opportunities is not considered, the total loss function can be optimized as follows: Where w1 + w2 = 1. By designing corresponding weights for different loss functions and combining them in a weighted manner, multi-task collaborative training can be achieved, enabling the model to capture the semantic features of business opportunities while also adapting to classification and contextual understanding requirements, thereby improving overall recommendation performance. Then, by summing and averaging the losses of all training samples, the overall loss is obtained. With the goal of minimizing this overall loss, the model parameters are iteratively updated using the backpropagation algorithm until the model converges, thus completing the model training process.
[0107] The method provided in the above embodiments, by constructing a multi-task learning framework that includes triplet loss, deep learning loss, and mask modeling loss, can optimize model performance from multiple dimensions: triplet loss ensures that similar business opportunities are clustered in the vector space, improving semantic discrimination ability; deep learning loss enhances the model's accuracy in predicting the conversion probability of business opportunities; and mask modeling loss improves the model's robustness to understanding text semantics. By jointly optimizing these three loss functions, the model can simultaneously learn better semantic representation, classification ability, and language understanding ability, thereby significantly improving the accuracy and practicality of business opportunity recommendations.
[0108] See Figure 7 The diagram shows the main modules of a deep learning-based business opportunity recommendation device 700 provided in an embodiment of the present invention, including:
[0109] The acquisition module 701 is used to acquire multimodal data of the target business opportunity; wherein, the multimodal data includes structured data and unstructured data;
[0110] The encoding module 702 is used to input the multimodal data into the pre-trained model to encode semantic vectors;
[0111] The filtering module 703 is used to determine a list of similar business opportunities based on the semantic vector, and to filter out business opportunities whose current status information is valid from the list of similar business opportunities for recommendation.
[0112] In the apparatus of this invention, the encoding module 702 is used for:
[0113] Extract the first field value corresponding to the first preset field from the unstructured data, perform word segmentation on the first field value to determine high-frequency words according to their frequency of occurrence, and establish a first semantic unit based on the first preset field and the sentence in which the high-frequency words are located.
[0114] Extract the second field value corresponding to the second preset field from the structured data, determine the type of the second field value, process the second field value using the strategy corresponding to the type, and establish a second semantic unit based on the second preset field and the processed second field value;
[0115] The first semantic unit and the second semantic unit are concatenated to obtain a target sequence, and the target sequence is encoded to obtain a semantic vector.
[0116] In the apparatus of the present invention, the encoding module 702 further includes one or more of the following:
[0117] Based on a preset thesaurus, the characters in the target sequence are replaced with their corresponding synonyms;
[0118] The processed second field value is perturbed; wherein, in response to the type being a numeric type, the processed second field value is numerically perturbed; and in response to the type being a classification type, the processed second field value is replaced with a field value of a similar category.
[0119] The apparatus of this invention further includes a training module for training a preset model to obtain the pre-trained model. The training process includes:
[0120] Obtain a sample set and divide the sample set into a training set, a validation set, and a test set; wherein, the samples are divided into positive samples and negative samples, and each sample consists of a business opportunity pair;
[0121] The preset model is trained using the training set, and the model parameters are adjusted while keeping the model hyperparameters unchanged to obtain the first model.
[0122] The first model is trained using the validation set, and the hyperparameters are adjusted while keeping the model parameters unchanged to obtain the second model;
[0123] The second model is tested using the test set, and if the test results meet the preset conditions, the second model is used as the pre-trained model.
[0124] In the apparatus of this invention, the process of generating positive and negative samples includes:
[0125] Acquire multimodal data and behavioral data of historical business opportunities; wherein, the behavioral data includes user binding records and business opportunity conversion status;
[0126] By combining historical business opportunities in pairs, multiple business opportunity pairs can be obtained;
[0127] Among the multiple business opportunity pairs, those that are bound to the same user and have all successfully converted are considered as positive samples, and the remaining business opportunity pairs other than the positive samples are considered as negative samples.
[0128] In the apparatus of this invention, the step of treating the remaining business opportunity pairs other than the positive samples as negative samples includes one or more of the following:
[0129] Identify business opportunity pairs that are linked to different users and all of them have successfully converted;
[0130] Identify business opportunity pairs in multimodal data where the field values under the third and / or fourth preset fields differ;
[0131] Identify business opportunity pairs in multimodal data whose field values are the same or similar under the third preset field and have not been converted.
[0132] In the apparatus of this invention, the training module is used for:
[0133] Identify all positive and negative samples corresponding to the first business opportunity, identify the second business opportunity in the positive samples other than the first business opportunity, and identify the third business opportunity in the negative samples other than the first business opportunity;
[0134] Based on the semantic vectors of the first business opportunity, the second business opportunity, and the third business opportunity, calculate the triplet loss of the first business opportunity;
[0135] Based on the semantic vector similarity between the first business opportunity and the second business opportunity, and the semantic vector similarity between the first business opportunity and the third business opportunity, the deep learning loss of the first business opportunity is calculated.
[0136] The total loss of the first business opportunity is obtained through the triplet loss and the deep learning loss. The model is trained with the goal of minimizing the total loss of all business opportunities.
[0137] In the apparatus of this invention, the training module is further configured to: perform masking processing on specific information in each business opportunity sequence; and
[0138] Determine the mask modeling loss for the first business opportunity, and combine the triplet loss and the deep learning loss to obtain the total loss for the first business opportunity;
[0139] The mask modeling loss for determining the first business opportunity includes:
[0140] Determine the hidden state vector of the mask position in the sequence of the first business opportunity, and predict the probability of each word in the preset vocabulary at the mask position based on the hidden state vector;
[0141] Calculate the cross-entropy loss between the true and predicted values at the mask positions, sum and average the cross-entropy losses for all mask positions to obtain the mask modeling loss for the first business opportunity.
[0142] In the device of this invention, the filtering module 703 is used to: filter out business opportunities that are currently not bound to users and / or are in validity from the set of similar business opportunities for recommendation.
[0143] Furthermore, the specific implementation details of the device described in the embodiments of the present invention have been described in detail in the above-described method, so the details will not be repeated here.
[0144] Figure 8 An exemplary system architecture 800 to which embodiments of the present invention can be applied is shown, including terminal devices 801, 802, 803, network 804, and server 805 (only an example).
[0145] Terminal devices 801, 802, and 803 can be various electronic devices with displays and web browsing capabilities, and can be equipped with various communication client applications. Users can use terminal devices 801, 802, and 803 to interact with server 805 via network 804 to receive or send messages, etc.
[0146] Network 804 is a medium used to provide a communication link between terminal devices 801, 802, 803 and server 805. Network 804 can include various connection types, such as wired or wireless communication links or fiber optic cables, etc.
[0147] Server 805 can be a server providing various services, such as a backend management server supporting shopping websites browsed by users using terminal devices 801, 802, and 803 (this is just an example). The backend management server can analyze and process data such as received product information query requests, and feed back the processing results (such as target push information and product information—this is just an example) to the terminal devices. It should be noted that the method provided in this embodiment of the invention is generally executed by server 805, and correspondingly, the apparatus is generally set in server 805.
[0148] It should be understood that Figure 8 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.
[0149] The following is for reference. Figure 9 It shows a schematic diagram of the structure of a computer system 900 suitable for implementing a terminal device of the present invention. Figure 9 The terminal device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the present invention.
[0150] like Figure 9As shown, the computer system 900 includes a central processing unit (CPU) 901, which can perform various appropriate actions and processes based on programs stored in read-only memory (ROM) 902 or programs loaded from storage section 908 into random access memory (RAM) 903. The RAM 903 also stores various programs and data required for the operation of the system 900. The CPU 901, ROM 902, and RAM 903 are interconnected via a bus 904. An input / output (I / O) interface 905 is also connected to the bus 904.
[0151] The following components are connected to I / O interface 905: an input section 906 including a keyboard, mouse, etc.; an output section 907 including a cathode ray tube (CRT), liquid crystal display (LCD), etc., and speakers, etc.; a storage section 908 including a hard disk, etc.; and a communication section 909 including a network interface card such as a LAN card, modem, etc. The communication section 909 performs communication processing via a network such as the Internet. A drive 910 is also connected to I / O interface 905 as needed. A removable medium 911, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on drive 910 as needed so that computer programs read from it can be installed into storage section 908 as needed.
[0152] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication section 909, and / or installed from removable medium 911. When the computer program is executed by central processing unit (CPU) 901, it performs the functions defined above in the system of this invention.
[0153] It should be noted that the computer-readable medium shown in this invention can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this invention, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this invention, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. Computer-readable signal media can also be any computer-readable medium other than computer-readable storage media, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wireless, wire, optical fiber, RF, etc., or any suitable combination thereof.
[0154] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.
[0155] The modules described in the embodiments of the present invention can be implemented in software or hardware. The described modules can also be housed in a processor; for example, a processor may be described as including an acquisition module, an encoding module, and a filtering module. The names of these modules do not necessarily limit the module itself; for example, the filtering module may also be described as a "recommendation module."
[0156] In another aspect, the present invention also provides a computer-readable medium, which may be included in the device described in the above embodiments; or it may exist independently and not assembled into the device. The computer-readable medium carries one or more programs that, when executed by the device, cause the device to perform any of the deep learning-based business opportunity recommendation methods described above.
[0157] The computer program product of the present invention includes a computer program that, when executed by a processor, implements the deep learning-based business opportunity recommendation method in the embodiments of the present invention.
[0158] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can occur depending on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A business opportunity recommendation method based on deep learning, characterized in that, The method comprises: acquiring multi-modal data of a target business opportunity; wherein the multi-modal data comprises structured data and unstructured data; inputting the multi-modal data into a pre-trained model to encode a semantic vector; determining a similar business opportunity list based on the semantic vector, and screening business opportunities with valid current state information from the similar business opportunity list for recommendation.
2. The method of claim 1, wherein, The method further comprises one or more of the following: extracting a first field value corresponding to a first preset field from the unstructured data, performing word segmentation processing on the first field value to determine high-frequency words according to the frequency of occurrence, and establishing a first semantic unit based on the first preset field and the high-frequency words in the sentence; extracting a second field value corresponding to a second preset field from the structured data, determining the type of the second field value, processing the second field value using a strategy corresponding to the type, and establishing a second semantic unit based on the second preset field and the processed second field value; performing splicing processing on the first semantic unit and the second semantic unit to obtain a target sequence, and performing encoding processing on the target sequence to obtain a semantic vector.
3. The method of claim 2, wherein, The method further comprises one or more of the following: replacing characters in the target sequence with corresponding synonyms based on a preset synonym library; performing perturbation processing on the processed second field value; wherein, in response to the type being a numerical type, performing numerical perturbation processing on the processed second field value; and in response to the type being a classification type, replacing the processed second field value with a field value of a similar category.
4. The method according to any one of claims 1 to 3, characterized in that, The method further comprises training a pre-set model to obtain the pre-trained model, and the training process comprises: acquiring a sample set, and dividing the sample set into a training set, a validation set, and a test set; wherein the samples are divided into positive samples and negative samples, and each sample is composed of a business opportunity pair; training the pre-set model using the training set, adjusting the model parameters while keeping the model hyperparameters unchanged to obtain a first model; training the first model using the validation set, adjusting the model hyperparameters while keeping the model parameters unchanged to obtain a second model; testing the second model using the test set, and in response to the test result meeting a preset condition, regarding the second model as the pre-trained model.
5. The method of claim 4, wherein, The process of generating positive samples and negative samples comprises: acquiring multi-modal data and behavior data of historical business opportunities; wherein the behavior data comprises user binding records and business opportunity conversion states; pairing the historical business opportunities to obtain a plurality of business opportunity pairs; regarding business opportunity pairs that bind to the same user and both successfully convert as positive samples, and regarding the remaining business opportunity pairs other than the positive samples as negative samples.
6. The method of claim 5, wherein, The process of regarding the remaining business opportunity pairs other than the positive samples as negative samples comprises one or more of the following: determining business opportunity pairs that bind to different users and both successfully convert; determining business opportunity pairs with different field values in the third preset field and / or the fourth preset field in the multi-modal data; determining business opportunity pairs with the same or similar field values in the third preset field in the multi-modal data and both not converting.
7. The method of claim 4, wherein, The process of training the model comprises: determine all positive samples and all negative samples corresponding to the first business opportunity, determine a second business opportunity in the positive samples except the first business opportunity, and determine a third business opportunity in the negative samples except the first business opportunity; calculate a triplet loss of the first business opportunity based on semantic vectors of the first business opportunity, the second business opportunity, and the third business opportunity; calculate a deep learning loss of the first business opportunity based on semantic vector similarities of the first business opportunity and the second business opportunity, and semantic vector similarities of the first business opportunity and the third business opportunity; obtain a total loss of the first business opportunity through the triplet loss and the deep learning loss, and train a model with a target of minimizing total losses of all business opportunities.
8. The method of claim 7, wherein, The method further includes: performing mask processing on specific information in each business opportunity sequence. The obtaining of the total loss of the first business opportunity through the triplet loss and the deep learning loss includes: determining a mask modeling loss of the first business opportunity, and obtaining the total loss of the first business opportunity by combining the triplet loss and the deep learning loss. The determination of the mask modeling loss of the first business opportunity includes: determining a hidden state vector of a mask position in a sequence of the first business opportunity, and predicting probabilities of each word in a preset vocabulary at the mask position based on the hidden state vector; calculating a cross-entropy loss of a true value and a predicted value of the mask position, summing and averaging cross-entropy losses of all mask positions to obtain the mask modeling loss of the first business opportunity.
9. The method of claim 1, wherein, The filtering of the business opportunity with valid current state information from the similar business opportunity set for recommendation includes: filtering the business opportunity that is not bound to a user and / or is in a valid period from the similar business opportunity set for recommendation. 10.A business opportunity recommendation apparatus based on deep learning, characterized by comprising: The method includes: an acquisition module configured to acquire multi-modal data of a target business opportunity; the multi-modal data includes structured data and unstructured data; an encoding module configured to input the multi-modal data into a pre-trained model to obtain semantic vectors through encoding; a filtering module configured to determine a similar business opportunity list based on the semantic vectors, and filter a business opportunity with valid current state information from the similar business opportunity list for recommendation.
11. An electronic device, comprising: The method includes: one or more processors; a storage device configured to store one or more programs, when the one or more programs are executed by the one or more processors, the one or more processors implement the method of any one of claims 1-9.
12. A computer readable medium having stored thereon a computer program, characterized in that, The program is executed by the processor to implement the method of any one of claims 1-9.
13. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the method of any one of claims 1-9.