Product recommendation method and device
By constructing a multi-dimensional similarity network and integrating fund manager holdings, text, and performance data, the similarity assessment is dynamically updated, solving the problem of inaccurate fund manager similarity research in existing technologies and achieving personalized and accurate product recommendations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-22
- Publication Date
- 2026-03-27
AI Technical Summary
Current technologies for studying fund manager similarity rely on single performance indicators or qualitative labels, making it difficult to comprehensively and objectively capture the true situation of fund managers, leading to inaccurate product recommendations.
We construct a multi-dimensional similarity network based on holdings, text, and performance, and integrate the fund manager's holdings similarity network, text similarity network, and performance similarity network. We dynamically update the network through a Bayesian fusion mechanism to generate a fused network that reflects comprehensive similarity and recommends products based on user needs.
It enables a comprehensive and objective assessment of fund managers' investment behavior and style, improves the accuracy and reliability of product recommendations, can identify implicit correlations, and provides personalized product recommendations.
Smart Images
Figure CN121746083A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of data processing, in particular to a product recommendation method and device. BACKGROUND
[0002] In the field of financial technology and quantitative investment, fund manager similarity research is an important direction. It can classify fund managers with similar investment styles, such as value investment, growth investment, and trend tracking, through multi-dimensional analysis. Based on this classification, fund managers and fund products can be accurately recommended according to the needs of investors.
[0003] The existing technology relies on a single performance indicator or qualitative label for fund manager similarity research. However, the decision-making mode of fund managers is complex and may change over time. These methods are difficult to capture the true situation comprehensively and objectively, and thus cannot accurately recommend products to investors. SUMMARY
[0004] Based on the above problems, the present application provides a product recommendation method and device to overcome the limitations of traditional methods, provide accurate and reliable fund manager similarity research, and accurately recommend products to investors.
[0005] The present application discloses a product recommendation method, which comprises:
[0006] Constructing a holding position similarity network based on the holding position data of each fund manager to be recommended;
[0007] Constructing a text similarity network based on the text data of each fund manager to be recommended; the text data includes periodic reports, customer communication records, public speeches, and network social content;
[0008] Constructing a performance similarity network based on the performance indicators of each fund manager to be recommended; the performance indicators include yield, Sharpe ratio, and Calmar ratio;
[0009] Defining each fund manager to be recommended as a network node, and the edge weight between each network node as the node similarity. The holding position similarity network, the text similarity network, and the performance similarity network are fused to obtain a fusion network;
[0010] Inputting user demand into the fusion network to generate a product corresponding to the user demand; the product is a fund manager and / or a fund.
[0011] Optionally, the method of constructing a holding position similarity network based on the holding position data of each fund manager to be recommended comprises:
[0012] According to the fund classification standard, the funds under the fund managers to be recommended are screened out as recommended funds.
[0013] extracting periodic report holding data of the to-be-recommended funds;
[0014] According to the periodic report holding data, the to-be-recommended funds under the same to-be-recommended fund manager are merged and weight-normalized to generate a weight set; the weight set is used to represent the holding idea of the to-be-recommended fund manager;
[0015] Based on semantic matching, the holding similarity of each element in the weight set is calculated;
[0016] Based on the holding similarity, the holding similarity network is constructed.
[0017] Optionally, the text similarity network is constructed based on the text data of each to-be-recommended fund manager, comprising:
[0018] The text data of the to-be-recommended fund manager in a preset time period is collected;
[0019] The Doc2Vec model and the FinBERT model are fused to calculate the cosine similarity between the text data;
[0020] The average value of the cosine similarity is taken as the text similarity;
[0021] Based on the text similarity, the text similarity network is constructed.
[0022] Optionally, the performance similarity network is constructed based on the performance indicators of each to-be-recommended fund manager, comprising:
[0023] The historical performance data of the to-be-recommended fund manager in the preset time period is collected;
[0024] The performance indicators are calculated for the historical performance data;
[0025] The performance similarity of the performance indicators of each to-be-recommended fund manager is calculated;
[0026] Based on the performance similarity, the performance similarity network is constructed.
[0027] Optionally, after the fusion network is obtained, the method further comprises:
[0028] The historical fusion network at time t-1 and the fusion network at time t are assigned weights; the time t is the current time;
[0029] Based on the Bayesian fusion mechanism, the historical fusion network is updated by the fusion network at time t.
[0030] Based on the product recommendation method, the application further discloses a product recommendation device, comprising a first construction unit, a second construction unit, a third construction unit, a fusion unit and a generation unit.
[0031] The first construction unit is configured to construct a holding similar network based on holding data of each fund manager to be recommended.
[0032] The second construction unit is configured to construct a text similar network based on text data of each fund manager to be recommended; the text data comprises periodic reports, customer communication records, public speeches and network social content.
[0033] The third construction unit is configured to construct a performance similar network based on performance indicators of each fund manager to be recommended; the performance indicators comprise yield, Sharpe ratio and Calmar ratio.
[0034] The fusion unit is configured to define each fund manager to be recommended as a network node, and the edge weight between each network node is the node similarity; the holding similar network, the text similar network and the performance similar network are fused to obtain a fusion network.
[0035] The generation unit is configured to input user demand into the fusion network to generate a product corresponding to the user demand; the product is a fund manager and / or a fund.
[0036] Optionally, the first construction unit comprises:
[0037] The screening subunit is configured to screen funds under the fund managers to be recommended according to fund classification standards as funds to be recommended.
[0038] The extraction subunit is configured to extract periodic report holding data of the funds to be recommended.
[0039] The normalization subunit is configured to combine and weight normalize the funds to be recommended under the same fund manager to be recommended according to the periodic report holding data to generate a weight set; the weight set is used to represent the holding ideas of the fund manager to be recommended.
[0040] The first calculation subunit is configured to calculate the holding similarity of each element in the weight set based on semantic matching.
[0041] The first construction subunit is configured to construct the holding similar network based on the holding similarity.
[0042] Optionally, the second construction unit comprises:
[0043] The collection subunit is configured to collect the text data of the fund managers to be recommended within a preset time period.
[0044] The second computational subunit is used to fuse the Doc2Vec model and the FinBERT model to calculate the cosine similarity between the text data.
[0045] The average subunit is used to take the average value of the cosine similarity as the text similarity.
[0046] The second construction subunit is used to construct the text similarity network based on the text similarity.
[0047] Optionally, the third building unit includes:
[0048] The acquisition sub-unit is used to collect historical performance data of the fund manager to be recommended for the preset time period.
[0049] The indicator calculation subunit is used to calculate the performance indicator based on the historical performance data.
[0050] The third calculation subunit is used to calculate the performance similarity of the performance indicators of each of the fund managers to be recommended.
[0051] The third construction subunit is used to construct the performance similarity network based on the performance similarity.
[0052] Optionally, the device further includes:
[0053] An allocation unit is used to allocate weights to the historical fusion network at time t-1 and the fusion network at time t; where time t is the current time.
[0054] The update unit is used to update the historical fusion network by weighting the fusion network at time t based on the Bayesian fusion mechanism.
[0055] This application discloses a method and apparatus for product recommendation. Based on the holdings data, text data, and performance indicators of each fund manager to be recommended, a holdings similarity network, a text similarity network, and a performance similarity network are constructed respectively. Breaking through the limitations of traditional single, static evaluation indicators, this method constructs three independent similarity networks based on performance, holdings, and publicly available text information, achieving a comprehensive and objective assessment of fund managers' investment behavior and style. Specifically, by constructing the text network, the method fully analyzes publicly available text such as fund managers' communication records and social media statements to extract their investment philosophies and identify implicit relationships. Each fund manager to be recommended is defined as a network node, and the edge weights between network nodes represent node similarity. The holdings similarity network, text similarity network, and performance similarity network are fused to obtain a comprehensive and robust similarity matrix, achieving complementarity and unification of multi-source information and improving the accuracy and reliability of the assessment. By inputting user needs into the fused network, personalized and accurate recommendations of fund managers and funds that meet the user's needs can be provided. Attached Figure Description
[0056] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of this application. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.
[0057] Figure 1 This is a flowchart illustrating a product recommendation method disclosed in an embodiment of this application;
[0058] Figure 2 This is a flowchart illustrating another product recommendation method disclosed in an embodiment of this application;
[0059] Figure 3 This is a schematic diagram of the structure of a product recommendation device disclosed in an embodiment of this application. Detailed Implementation
[0060] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.
[0061] Example 1: This application discloses a method for product recommendation.
[0062] For details, please refer to Figure 1 The product recommendation method disclosed in this embodiment includes the following steps:
[0063] Step 101: Construct a holdings similarity network based on the holdings data of each fund manager to be recommended.
[0064] In this embodiment, funds under the names of potential fund managers are selected as potential recommended funds based on fund classification standards (specifically, standards determined by experience or industry-specific standards). These potential recommended funds include those directly managed by a single fund manager and those jointly managed by multiple fund managers. Subsequently, the periodic report holdings data of these potential recommended funds are extracted, and all constituent securities and their corresponding weights are aggregated. Different potential recommended funds under the same fund manager are merged and their weights are normalized to generate a weight set. The multiple weight vector elements in this weight set are standardized to represent the overall portfolio allocation strategy of the potential recommended fund manager, providing a unified and comparable data foundation for subsequent similarity analysis.
[0065] In the method of this embodiment, the semantic alignment of each element in the weight set can be calculated based on iterative semantic matching. The element with the highest semantic alignment during the iterative calculation is selected to quantify the similarity between the two weight sets, which is to calculate the holding similarity.
[0066] As a feasible approach, we can define X as the reference weight set and Y as the candidate weight set, with each element in both sets associated with a positive weight w. x (i), w ᵧ (j) and a semantic representation. For each pair of elements (x) in the set i y j The selected component-level similarity metric is used to calculate the pairwise similarity (position similarity) score S. ij .
[0067] Specifically, this can be done by first treating all elements in sets X and Y as unmatched, with all weights of all elements in an available state. In each iteration, the similarity score S of the unmatched element pairs is selected. ij The highest element pair (x i y j The contribution of this match to the overall similarity score is min(w). x (i),w ᵧ (j)×S ij , where w x (i) and w ᵧ (j) represent elements x respectively i and element y j Current available weights. From element x i weight w x (i) and element y j weight w ᵧ Subtract the weight used in this match from (j) to complete the weight update. When the weight of an element drops to 0, that element will no longer participate in subsequent matches. The iteration process stops when there are no more valid weights available for matching. The final similarity score is the sum of all weighted pairwise similarities during the matching process. A holding similarity network can be constructed based on the calculated holding similarity. In addition, a residual term is generated after the algorithm terminates to record the sum of the remaining unmatched weights in the two sets.
[0068] Step 102: Construct a text similarity network based on the text data of each of the fund managers to be recommended.
[0069] In this embodiment, the text data originates from all publicly available language materials generated by the fund manager to be recommended, mainly including four categories. The first is statutory periodic reports, such as the "Manager's Report" and "Investment Strategy and Operation Analysis" sections of fund quarterly and annual reports; these are the most standardized and continuous core texts. The second is roadshow minutes and client communication records; these texts reflect more immediate market views and operational strategies. The third is public speeches and media interviews, which often contain in-depth explanations of their investment philosophy. The fourth is online social content, such as official self-media articles, which have a more flexible style. Before constructing the text similarity network, all the above-mentioned texts generated by a fund manager to be recommended within a preset time period (e.g., a single year) need to be aggregated into a (annual) comprehensive document, serving as the benchmark unit for the text analysis of that fund manager in that (annual) period.
[0070] In this embodiment, the method first collects text data from the fund managers to be recommended within a preset time period, i.e., the comprehensive document mentioned above. The key to constructing the text similarity network is to transform the unstructured comprehensive document into vectors with computable similarity and to define the weights of the edges. As an feasible solution, this embodiment innovatively integrates two complementary natural language models, the Doc2Vec model and the FinBERT model, to calculate text similarity.
[0071] Specifically, the Doc2Vec model can be used, trained with hyperparameters based on well-established research on financial texts (such as 10-X documents), to generate a fixed-dimensional vector representation for each comprehensive document. To overcome the limitation of the Doc2Vec model ignoring word order and context, a FinBERT model pre-trained on financial corpora can be introduced to further enhance the discriminative power of document vectors. Specifically, the self-supervised SimCSE contrastive learning method can be used to fine-tune FinBERT, enabling it to learn to converge vectors of semantically similar sentences and distance vectors of dissimilar sentences.
[0072] In this embodiment, the document is first segmented into sentences, and the FinBERT embedding of each sentence is obtained. Then, average pooling is used to generate a contextualized vector for the entire document. This vector can more precisely capture the logical differences in investment arguments. Subsequently, for any two documents from fund managers to be recommended in a given year, the cosine similarity between their document vectors is independently calculated using a pre-trained Doc2Vec model and a fine-tuned FinBERT model, respectively. This yields two similarity scores, and the average of these two scores is taken as the text similarity. Finally, a text similarity network is constructed based on the text similarity.
[0073] Step 103: Construct a performance similarity network based on the performance indicators of each of the fund managers to be recommended.
[0074] In this embodiment, historical performance data of the fund manager to be recommended over a preset time period is collected, and performance indicators are calculated based on the historical performance data. Specifically, the adjusted net asset value per unit of the fund can be obtained from a professional financial data platform, and the return series for different periods can be calculated. To ensure comparability, funds can be grouped by category and compared within the same category. Furthermore, net asset value anomalies (such as large dividends, massive redemptions, etc.) are handled by using sufficiently long historical data (such as the past year or longer) to reflect persistent characteristics.
[0075] The performance metrics in this embodiment include return, excess return, Sharpe ratio, volatility, maximum drawdown, and Calmar ratio.
[0076] In this embodiment, the fund manager (or individual fund) to be recommended is first considered as a node in a network. The weight of the edge between any two nodes is defined by the weighted average of the correlation coefficients of their performance indicator sequences. The higher the correlation, the more similar the two nodes are in their historical performance trends, and the greater the weight of the edge. Then, the performance similarity, i.e., the correlation coefficient, of the performance indicators of each fund manager to be recommended is calculated, and a performance similarity network is constructed based on this performance similarity. As an feasible approach, a complete similarity matrix can be calculated. This matrix is essentially a mathematical expression of the performance similarity network and can be directly used for subsequent analysis.
[0077] In the method of this embodiment, to enhance the depth of analysis, operations such as "similarity neutralization" can be performed on the performance similarity network. Specifically, this can involve calculating the average performance indicators of a fund manager to be recommended and its K most similar neighbors (local network), and then subtracting this local mean from the fund manager's original indicators to obtain the residual. This process can effectively remove the part of the performance influenced by common market styles, thereby more purely identifying the individual management capabilities of the fund manager to be recommended.
[0078] Step 104: Define each of the fund managers to be recommended as a network node, and the edge weights between the network nodes as node similarity. Merge the holdings similarity network, the text similarity network, and the performance similarity network to obtain a fused network.
[0079] In this embodiment, to achieve deep integration and dynamic updating of multi-dimensional correlation information of the fund managers to be recommended, a three-dimensional network fusion based on Bayesian principles can be performed. This embodiment treats three independent networks—holding similarity network, text similarity network, and performance similarity network—as the same set of behavioral data on the relationships between the fund managers to be recommended under different observation dimensions. Through iterative updates, a robust network reflecting comprehensive similarity can be obtained.
[0080] As a feasible approach, each fund manager to be recommended is considered a node, and the connection strength (edge weight) between any two nodes is defined as their similarity across various observation dimensions (similar networks). Specifically, the text similarity network is constructed based on the vector similarity of investment documents, revealing the underlying logic. The portfolio similarity network is constructed based on semantic matching of components and weights, reflecting practical competence. The performance similarity network is constructed based on the correlation of return series, capturing market co-variation patterns.
[0081] Subsequently, the historical fusion network at time t-1 (the time preceding the current time) is considered the prior network, and its edge weight distribution represents the comprehensive similarity based on all information up to time t-1. At time t (the current time), the latest text data, position data, and performance indicators are obtained, and three similar networks are constructed accordingly. Dynamic weights (e.g., calculated based on the recent information entropy or stability of each network) are then assigned to each of the three similar networks to characterize their relative information reliability. The edge weights of the fusion network at time t are updated by performing a Bayesian weighted average of the edge weights of the prior network and the three weighted similar networks.
[0082] Step 105: Input the user's requirements into the fusion network to generate a product corresponding to the user's requirements.
[0083] In this embodiment, the product can be any one or both of a fund manager and a fund. In practical applications, users can request the fusion network to recommend a fund manager or a fund. For example, if user A is a new customer and there are no fund purchase records for user A in the system, user A can input the target fund they want to buy or query. The fusion network in this embodiment can then query which fund manager manages the target fund, and then query other fund managers with high similarity to that fund manager, as well as similar funds under other fund managers, for recommendations. A list sorted from high to low similarity can be generated, or a preset number of recommended products can be generated.
[0084] In another scenario, User B is a long-term customer, meaning their fund purchase records are available in the system. Based on these records, the system can identify User B's investment portfolio and recommend fund managers whose investment style aligns with User B's, as well as similar funds managed by other fund managers. Alternatively, User B can input their desired fund or fund manager, and the system will then search and recommend funds using the methods described above.
[0085] In the method of this embodiment, users A and B can also input their investment style requirements, and the fusion network of this embodiment will automatically match funds or fund managers that meet their requirements.
[0086] The fusion network described in this embodiment not only balances information across three dimensions—surface performance correlation, overlapping actual holdings, and deep logical similarity—but also continuously learns and iterates on historical information through a Bayesian framework. This allows for more sensitive identification of "implicit correlations" that are not obvious in a single dimension but consistently point across dimensions. For example, two fund managers may have low performance correlation coefficients and low holding overlap, but their text networks reveal highly similar investment philosophies and decision-making logic. This strong cross-dimensional signal is amplified and captured during the fusion process, revealing potential correlation opportunities stemming from consistent underlying logic that have not yet been fully recognized by the market. Specifically, the holding network uses a "semantic matching + weight perception + greedy aggregation" algorithm to accurately quantify the similarity of "partially overlapping" combinations that are related but have different weights. The text network integrates Doc2Vec and a FinBERT model fine-tuned by SimCSE to extract investment philosophies from massive amounts of unstructured data, thereby identifying implicit strategy correlations where "philosophies overlap but holdings or performance differ." The performance network, through similarity neutralization, removes common market influences and focuses on differences in individual capabilities.
[0087] Example 2: This application discloses another method for product recommendation; please refer to [link / reference]. Figure 2 The method described in this embodiment introduces the entire process of product recommendation.
[0088] Step 201: Obtain the current holdings data, text data, and performance indicators of each fund manager to be recommended.
[0089] Step 202: Construct a holdings similarity network based on holdings data, a text similarity network based on text data, and a performance similarity network based on performance indicators.
[0090] Step 203: Merge the holdings similarity network, text similarity network, and performance similarity network to obtain the fused network.
[0091] Step 204: Update the fusion network from the previous time step using the fusion network. Repeat step 201.
[0092] Step 205: Receive user input request: Recommend a fund manager.
[0093] Step 206: Determine if the user's fund purchase records can be obtained. If yes, proceed to step 207. If no, proceed to step 208.
[0094] Step 207: Generate a recommendation list containing 10 fund managers whose fund purchase records are more similar to the user's than a preset value.
[0095] Step 208: Request the user to input their investment style, target fund, or target fund manager, etc.
[0096] Step 209: Receive the user's supplementary request and generate a recommendation list containing 10 fund managers whose investment style, target fund, or target fund manager is more similar to the one in the supplementary request than a preset value.
[0097] Step 210: Return the recommendation list, along with the funds managed by the fund managers in the recommendation list, to the user.
[0098] Based on the product recommendation method disclosed in the above embodiments, this embodiment correspondingly discloses a product recommendation apparatus. Please refer to... Figure 3 The product recommendation apparatus includes: based on the above-mentioned product recommendation method, this application also discloses a product recommendation apparatus, including: a first building unit, a second building unit, a third building unit, a fusion unit, and a generation unit;
[0099] The first construction unit is used to construct a holdings similarity network based on the holdings data of each fund manager to be recommended;
[0100] The second construction unit is used to construct a text similarity network based on the text data of each of the fund managers to be recommended; the text data includes periodic reports, client communication records, public speeches, and online social content;
[0101] The third construction unit is used to construct a performance similarity network based on the performance indicators of each of the fund managers to be recommended; the performance indicators include return, Sharpe ratio and Calmar ratio.
[0102] The fusion unit is used to define each of the fund managers to be recommended as a network node, and the edge weights between the network nodes are the node similarity. The holdings similarity network, the text similarity network, and the performance similarity network are fused to obtain a fused network.
[0103] The generation unit is used to input user requirements into the fusion network and generate products corresponding to the user requirements; the products are fund managers and / or funds.
[0104] Optionally, the first building unit includes:
[0105] The screening sub-unit is used to screen out funds under the name of the fund manager to be recommended, based on the fund classification standard, and to select them as funds to be recommended.
[0106] Extraction sub-unit, used to extract the periodic report holding data of the fund to be recommended;
[0107] The normalization sub-unit is used to merge and normalize the weights of the funds under the same fund manager based on the periodic report holding data, and generate a weight set; the weight set is used to represent the holding strategy of the fund manager to be recommended.
[0108] The first calculation subunit is used to calculate the holding similarity of each element in the weight set based on semantic matching;
[0109] The first construction subunit is used to construct the holding similarity network based on the holding similarity.
[0110] Optionally, the second building unit includes:
[0111] The data collection subunit is used to collect the text data of the fund manager to be recommended within a preset time period.
[0112] The second computational subunit is used to fuse the Doc2Vec model and the FinBERT model to calculate the cosine similarity between the text data.
[0113] The average subunit is used to take the average value of the cosine similarity as the text similarity.
[0114] The second construction subunit is used to construct the text similarity network based on the text similarity.
[0115] Optionally, the third building unit includes:
[0116] The acquisition sub-unit is used to collect historical performance data of the fund manager to be recommended for the preset time period.
[0117] The indicator calculation subunit is used to calculate the performance indicator based on the historical performance data.
[0118] The third calculation subunit is used to calculate the performance similarity of the performance indicators of each of the fund managers to be recommended.
[0119] The third construction subunit is used to construct the performance similarity network based on the performance similarity.
[0120] Optionally, the device further includes:
[0121] An allocation unit is used to allocate weights to the historical fusion network at time t-1 and the fusion network at time t; where time t is the current time.
[0122] The update unit is used to update the historical fusion network by weighting the fusion network at time t based on the Bayesian fusion mechanism.
[0123] The embodiments in this specification are described in a progressive manner. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant details can be found in the method section.
[0124] It should also be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0125] The steps of the methods or algorithms described in conjunction with the embodiments disclosed herein can be implemented directly by hardware, a software module executed by a processor, or a combination of both. The software module can be located in random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, registers, hard disk, removable disk, CD-ROM, or any other form of storage medium known in the art.
[0126] The features described in the embodiments of this specification can be substituted for or combined with each other, so that those skilled in the art can implement or use this application.
[0127] The above description of the disclosed embodiments enables those skilled in the art to make or use this application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for product recommendation, characterized in that, include: Construct a holdings similarity network based on the holdings data of each fund manager to be recommended; A text similarity network is constructed based on the text data of each of the fund managers to be recommended; The text data includes periodic reports, customer communication records, public speeches, and online social content; A performance similarity network is constructed based on the performance indicators of each of the fund managers to be recommended; the performance indicators include return, Sharpe ratio, and Calma ratio. Each of the fund managers to be recommended is defined as a network node, and the edge weight between the network nodes is the node similarity. The holding similarity network, the text similarity network, and the performance similarity network are fused to obtain a fused network. User requirements are input into the fusion network to generate products corresponding to those user requirements; the products are fund managers and / or funds.
2. The method according to claim 1, characterized in that, The construction of a holdings similarity network based on the holdings data of each fund manager to be recommended includes: Funds managed by the fund managers to be recommended were selected based on the fund classification criteria and designated as funds to be recommended. Extract the periodic report holdings data of the fund to be recommended; Based on the holdings data in the periodic reports, the funds under the same fund manager to be recommended are merged and their weights are normalized to generate a weight set; the weight set is used to represent the holding strategy of the fund manager to be recommended. Based on semantic matching, the similarity of holdings of each element in the weight set is calculated; The holdings similarity network is constructed based on the holdings similarity.
3. The method according to claim 1, characterized in that, The construction of a text similarity network based on the text data of each of the fund managers to be recommended includes: Collect the text data of the fund manager to be recommended within a preset time period; The cosine similarity between the text data is calculated by fusing the Doc2Vec model and the FinBERT model. The average value of the cosine similarity is taken as the text similarity. The text similarity network is constructed based on the text similarity.
4. The method according to claim 1, characterized in that, The construction of a performance similarity network based on the performance indicators of each of the fund managers to be recommended includes: Collect historical performance data of the fund manager to be recommended for the preset time period; The performance indicators are calculated based on the historical performance data. Calculate the performance similarity of the performance indicators of each of the fund managers to be recommended; The performance similarity network is constructed based on the performance similarity.
5. The method according to claim 1, characterized in that, After obtaining the fused network, the method further includes: Assign weights to the historical fusion network at time t-1 and the fusion network at time t; where time t is the current time. Based on the Bayesian fusion mechanism, the historical fusion network is updated by weighting the fusion network at time t.
6. A product recommendation device, characterized in that, include: First building unit, second building unit, third building unit, fusion unit, and generation unit; The first construction unit is used to construct a holdings similarity network based on the holdings data of each fund manager to be recommended; The second construction unit is used to construct a text similarity network based on the text data of each of the fund managers to be recommended; the text data includes periodic reports, client communication records, public speeches, and online social content; The third construction unit is used to construct a performance similarity network based on the performance indicators of each of the fund managers to be recommended; the performance indicators include return, Sharpe ratio and Calmar ratio. The fusion unit is used to define each of the fund managers to be recommended as a network node, and the edge weights between the network nodes are the node similarity. The holdings similarity network, the text similarity network, and the performance similarity network are fused to obtain a fused network. The generation unit is used to input user requirements into the fusion network and generate products corresponding to the user requirements; the products are fund managers and / or funds.
7. The apparatus according to claim 6, characterized in that, The first building unit includes: The screening sub-unit is used to screen out funds under the name of the fund manager to be recommended, based on the fund classification standard, and to select them as funds to be recommended. Extraction sub-unit, used to extract the periodic report holding data of the fund to be recommended; The normalization sub-unit is used to merge and normalize the weights of the funds under the same fund manager based on the periodic report holding data, and generate a weight set; the weight set is used to represent the holding strategy of the fund manager to be recommended. The first calculation subunit is used to calculate the holding similarity of each element in the weight set based on semantic matching; The first construction subunit is used to construct the holding similarity network based on the holding similarity.
8. The apparatus according to claim 6, characterized in that, The second building unit includes: The data collection subunit is used to collect the text data of the fund manager to be recommended within a preset time period. The second computational subunit is used to fuse the Doc2Vec model and the FinBERT model to calculate the cosine similarity between the text data. The average subunit is used to take the average value of the cosine similarity as the text similarity. The second construction subunit is used to construct the text similarity network based on the text similarity.
9. The apparatus according to claim 6, characterized in that, The third building unit includes: The acquisition sub-unit is used to collect historical performance data of the fund manager to be recommended for the preset time period. The indicator calculation subunit is used to calculate the performance indicator based on the historical performance data. The third calculation subunit is used to calculate the performance similarity of the performance indicators of each of the fund managers to be recommended. The third construction subunit is used to construct the performance similarity network based on the performance similarity.
10. The apparatus according to claim 6, characterized in that, The device further includes: An allocation unit is used to allocate weights to the historical fusion network at time t-1 and the fusion network at time t; where time t is the current time. The update unit is used to update the historical fusion network by weighting the fusion network at time t based on the Bayesian fusion mechanism.