Information recommendation model training method and device, storage medium, and computer equipment

Abstract

The present invention belongs to the field of artificial intelligence, and discloses an information recommendation model training method and device, storage medium, and computer equipment. The present invention obtains a preliminary screening article set by screening articles to be pushed according to popular article information and user behavior data of users; constructs the articles in the preliminary screening article set into a graph data network of articles, and generates an article node sequence by random walk on the graph data network, and collects the articles on the article node sequence into a sample candidate set; finally, collects the preliminary screening article set and the sample candidate set into a training sample set and trains, to obtain an information recommendation model. The present invention constructs a graph data network of articles for each article in the preliminary screening article set, and then obtains a sample candidate set that meets the user's personalized needs by random walk on the graph data network, so that the information recommendation model trained by the training sample set collected by the preliminary screening article set and the sample candidate set increases user stickiness.

Description

Technical Field

[0001] The present invention relates to the field of information technology, and in particular to an information recommendation model training method and device, a storage medium, and a computer device. Background Art

[0002] The existing method of recommending articles for users to read is to train the model based on the reading behavior data of users on computer devices, using the data that was not clicked during exposure as negative samples and the data that was clicked as positive samples. This method will bring the following two problems: one is that it will lead to a serious Matthew effect, that is, popular articles will receive more attention, and unpopular articles will be ignored; the other is that the exposed data is only a subset after being screened through models and business configurations, so model training that relies on data and feedback selected through the above methods will produce selection bias. Due to the continuous accumulation of the above-mentioned popularity bias and selection bias, the Matthew effect in the recommendation system will become more and more serious, which will lead to a large deviation in the mining of user personalization, and then lead to low accuracy of articles recommended to users, and cause problems in the long run of online operation. Summary of the invention

[0003] In view of this, the present invention provides an information recommendation model training method and device, storage medium, and computer equipment, the main purpose of which is to obtain a sample candidate set that meets the user's personalized needs to increase the user stickiness of the trained information recommendation model.

[0004] According to one aspect of the present invention, a method for training an information recommendation model is provided, comprising:

[0005] Filter the articles to be pushed based on the popular article information and the user behavior data to obtain a preliminary screening article collection;

[0006] Constructing a graph data network of articles according to the relationship between the articles in the preliminary screening article set; wherein the graph data network includes interconnected article nodes;

[0007] Obtaining a sample candidate set by randomly walking on the graph data network;

[0008] The preliminary screening article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model.

[0009] Furthermore, the article to be pushed is screened according to the popular article information and the user behavior data to obtain a preliminary screened article set, including:

[0010] Determine, based on the user behavior data, the preference features that the user is interested in, and a list of articles that the user has historically clicked on;

[0011] Recalling the articles to be pushed according to the preference features of the user, to obtain a first recalled article set;

[0012] According to the list of articles clicked by the user in history, construct first article pairs of articles in the list, and calculate the similarity score of each of the first article pairs;

[0013] Recalling the articles to be pushed whose similarity scores of the first article pairs are greater than a preset threshold to obtain a second recalled article set;

[0014] Recalling the articles to be pushed according to the popular article information to obtain a third recalled article set;

[0015] The first recalled article set, the second recalled article set and the third recalled article set are combined into the initial screening article set.

[0016] Furthermore, the step of constructing first article pairs of articles in the list according to the list of articles clicked by the user in history, and calculating similarity scores of each of the first article pairs includes:

[0017] Counting the number of co-occurrences of each of the first article pairs in the list of articles clicked by the user in history and the number of article appearances in each of the first article pairs, and calculating an initial similarity score of each of the first article pairs;

[0018] The initial similarity score is weightedly calculated in sequence by the display position information weight, click time weight, and creation time weight corresponding to the first article to obtain a weighted similarity score, and the weighted similarity score is used as the similarity score of the first article pair.

[0019] Furthermore, constructing a graph data network of articles based on the relationship between the articles in the initially screened article set includes:

[0020] constructing a second article pair of articles in the initially screened article set according to the user's historical click records on each article in the initially screened article set;

[0021] The graph data network is constructed by using the second article pair as a neighbor edge of a graph data network and using the articles in the second article pair as two end nodes of the neighbor edge.

[0022] Furthermore, the step of using the second article pair as a neighbor edge of a graph data network and using the articles in the second article pair as two end nodes of the neighbor edge to construct the graph data network includes:

[0023] Calculating article similarity scores in the second article pair according to the number of co-occurrences of the second article pair, the number of occurrences of articles in the second article pair, and the popularity corresponding to the articles in the second article pair;

[0024] The article similarity scores in the second article pair are used as edge weights to perform weighted calculation on the neighbor edges of the graph data network corresponding to the second article to obtain weighted neighbor edges of the graph data network.

[0025] Furthermore, the method of generating an article node sequence by randomly walking on the graph data network and aggregating the articles on the article node sequence into a sample candidate set includes:

[0026] The edge weight corresponding to the neighbor edge between the two nodes is used as the jump probability;

[0027] Acquire a wandering path according to a node connection structure of the jump probability on the graph data network;

[0028] According to the walking path, an article node sequence is generated, and a sample candidate set is determined, wherein the sample candidate set includes at least one second article pair corresponding to the walking path.

[0029] Furthermore, after generating the article node sequence according to the walking path and determining the sample candidate set, the following steps are included:

[0030] Determine the articles to be predicted corresponding to the articles in the sample candidate set according to the historical click records of the user on the articles in the preliminary screening article set;

[0031] All the obtained articles to be predicted are added to the sample candidate set to obtain an updated sample candidate set.

[0032] According to another aspect of the present invention, there is provided an information recommendation model training device, comprising:

[0033] The recall module is used to filter the articles to be pushed based on the popular article information and the user behavior data, and obtain the initial screening article collection;

[0034] A graph data network construction module, used to construct a graph data network of articles according to the relationship between the articles in the preliminary screening article set; wherein the graph data network includes interconnected article nodes;

[0035] A candidate sample acquisition module, used to acquire a sample candidate set by randomly walking on the graph data network;

[0036] The model training module is used to collect the preliminary screening article set and the sample candidate set into a training sample set, and train the information recommendation model to be trained to obtain the information recommendation model.

[0037] According to another aspect of the present invention, a storage medium is provided, wherein the storage medium stores at least one executable instruction, and the executable instruction enables a processor to perform operations corresponding to any of the information recommendation model training methods described above.

[0038] According to another aspect of the present invention, there is provided a computer device, comprising: a processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus;

[0039] The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the information recommendation model training method as described above.

[0040] By means of the above technical solution, the technical solution provided by the embodiment of the present invention has at least the following advantages:

[0041] The present invention provides an information recommendation model training method and device, storage medium, and computer equipment. Compared with the prior art, the present invention obtains a preliminary screening article set by screening articles to be pushed according to popular article information and user behavior data of users; then, a graph data network of articles is constructed according to the relationship between each article in the preliminary screening article set, and a sample candidate set is obtained by randomly walking on the graph data network; finally, the preliminary screening article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model that can send recommended articles to the computer equipment. The present invention increases user stickiness by taking a preliminary screening article set that is screened according to user behavior data, and constructing a graph data network of articles for each article in the preliminary screening article set, and then obtaining a sample candidate set that meets the user's personalized needs by randomly walking on the graph data network, thereby increasing user stickiness of the information recommendation model trained by the training sample set aggregated from the preliminary screening article set and the sample candidate set.

[0042] The above description is only an overview of the technical solution of the present invention. In order to more clearly understand the technical means of the present invention, it can be implemented according to the contents of the specification. In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are listed below. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] Various other advantages and benefits will become apparent to those of ordinary skill in the art by reading the detailed description of the preferred embodiments below. The accompanying drawings are only for the purpose of illustrating the preferred embodiments and are not to be considered as limiting the present invention. Moreover, the same reference symbols are used throughout the accompanying drawings to represent the same components. In the accompanying drawings:

[0044] Figure 1 A flow chart of an information recommendation model training method provided by an embodiment of the present invention is shown;

[0045] Figure 2 A flow chart of another information recommendation model training method provided by an embodiment of the present invention is shown;

[0046] Figure 3 A flowchart of obtaining a preliminary screening article set provided by an embodiment of the present invention is shown;

[0047] Figure 4 A partial schematic diagram of a graph data network provided by an embodiment of the present invention is shown;

[0048] Figure 5 A flowchart of obtaining a sample candidate set provided by an embodiment of the present invention is shown;

[0049] Figure 6 A block diagram showing a composition of an information recommendation model training device provided by an embodiment of the present invention;

[0050] Figure 7 A block diagram showing a composition of another information recommendation model training device provided by an embodiment of the present invention is shown;

[0051] Figure 8 A schematic diagram of the physical structure of a computer device provided by an embodiment of the present invention is shown. DETAILED DESCRIPTION

[0052] Exemplary embodiments will be described in detail herein, examples of which are shown in the accompanying drawings. When the following description refers to the drawings, the same numbers in different drawings represent the same or similar elements unless otherwise indicated. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Instead, they are merely examples of devices and methods consistent with some aspects of the present invention as detailed in the appended claims.

[0053] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, device, product or server that includes a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0054] The information recommendation model training method provided in the present application can be applied in an application environment where a computer device and a server communicate through a network, wherein the server receives user behavior data of the computer device, and then screens the articles to be pushed according to the popular article information and the user behavior data of the user, obtains a preliminary screening article set, and then constructs a graph data network of the articles according to the relationship between the articles in the preliminary screening article set, and then obtains a sample candidate set by randomly walking on the graph data network, and finally collects the preliminary screening article set and the sample candidate set into a training sample set, and trains the information recommendation model to be trained to obtain an information recommendation model that recommends information to the computer device. Among them, the computer device can be, but is not limited to, various personal computers, laptops, smart phones or tablet computers. The server can be implemented with an independent server or a server cluster composed of multiple servers, and the execution subject of the information recommendation model training method provided in the embodiment of the present disclosure is generally a computer device with a certain computing power, such as a server corresponding to the computer device. The following takes the execution subject as an example of a server to illustrate the information recommendation model training method provided in the embodiment of the present disclosure.

[0055] Figure 1 The flowchart of the training method of the information recommendation model provided in the embodiment of this specification specifically includes the following steps:

[0056] 101. Filter the articles to be pushed according to the popular article information and the user behavior data to obtain a preliminary screened article set.

[0057] User behavior data includes various historical behavior data generated by the user on the computer device for articles in the list of the user's historical clicks. These historical behavior data can be used as behavioral feature data for analyzing the user when viewing the list of the user's historical clicks. Based on this, the server can obtain the historical behavior data of the target user when browsing or clicking on each article in the list of the user's historical clicks.

[0058] It should be noted that, according to the popular article information and the user's user behavior data, a recall strategy can be determined, and then the articles to be pushed can be recalled according to the recall strategy, and the recalled articles can be screened according to preset rules to obtain a preliminary screening article set; here, since the popular article information and the user's user behavior data are updated regularly every day, the articles to be pushed according to the popular article information are screened by recall screening through regular statistics. When screening the articles to be pushed, the top k articles of the recalled articles can be taken as the articles in the preliminary screening article set.

[0059] 102. Construct a graph data network of articles based on the relationship between the articles in the initially screened article set; wherein the graph data network includes interconnected article nodes.

[0060] Here, based on the list of articles clicked by the user in the past, the relationship between the articles in the initial screening article set is determined, and then a graph data network from article to article is constructed. The graph data network constructed in this way can mine more unbiased data designed according to user personalization in the following steps, thereby achieving the purpose of alleviating selective bias.

[0061] 103. Obtain a sample candidate set by randomly walking on the graph data network.

[0062] It should be noted that here, random walks are performed on the graph data network through breadth-first traversal and depth-first traversal to obtain the isomorphism and homogeneity of the graph data network, thereby achieving an unbiased sample candidate set.

[0063] 104. The initially screened article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model.

[0064] The information recommendation model to be trained here is the deepfm model. By aggregating the initial screening article set and the sample candidate set into a training sample set and training the information recommendation model to be trained, the final most accurate information recommendation model can be obtained. This information recommendation model can send recommended articles to various computer devices to complete the information recommendation service for users.

[0065] The present invention provides an information recommendation model training method. Compared with the prior art, the embodiment of the present invention obtains a preliminary screening article set by screening articles to be pushed according to popular article information and user behavior data of users; then, a graph data network of articles is constructed according to the relationship between the articles in the preliminary screening article set, and a sample candidate set is obtained by randomly walking on the graph data network; finally, the preliminary screening article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model that can send recommended articles to a computer device. The present invention increases user stickiness by taking a preliminary screening article set that is screened according to user behavior data, and constructing a graph data network of articles for each article in the preliminary screening article set, and then obtaining a sample candidate set that meets the user's personalized needs by randomly walking on the graph data network, thereby increasing user stickiness of the information recommendation model trained by the training sample set that is aggregated by the preliminary screening article set and the sample candidate set.

[0066] The embodiment of the present invention provides another information recommendation model training method, such as Figure 2 As shown, the method includes:

[0067] 201. Filter the articles to be pushed according to the popular article information and the user behavior data to obtain a preliminary screened article set.

[0068] It should be noted that if the user is an old user, the user behavior data of the old user will be directly obtained; if the user is a new user, it is necessary to fill in the article category through the new user. For example, if the new user fills in the article category of "urban", "recipe", or "inspirational", the articles to be pushed will be filtered according to the above article categories to obtain a preliminary screened article set including "urban", "recipe", and "inspirational" articles.

[0069] In order to make the articles in the preliminary screening article set more in line with the user's personalized needs, the articles to be pushed are screened according to the popular article information and the user's user behavior data to obtain the preliminary screening article set. Figure 3 , which may include:

[0070] 2011. Determine, based on the user behavior data, the preference features that the user is interested in, and a list of articles that the user has historically clicked on.

[0071] Among them, for old users, their own reading habits can be obtained based on user behavior data, that is, weighted sampling is performed through the categories of articles that users have clicked on in the past to obtain N articles, and then the preference features that users are interested in are determined by analyzing these N articles; for new users, the preference features that new users are interested in are determined through user behavior data with optional categories.

[0072] 2012. Recall the articles to be pushed according to the preference characteristics of the user, and obtain a first recalled article set.

[0073] Since the user behavior data is counted daily, the preference features of the user determined based on the user behavior data are also updated daily. Here, in order to avoid too many articles in the first recalled article set, the number of articles in the first recalled article set is screened to be the same as the preset threshold.

[0074] 2013. According to the list of articles clicked by the user in history, construct the first article pairs of the articles in the list, and calculate the similarity scores of each of the first article pairs.

[0075] Here, the similarity score of the first article pair is calculated, so that the recommended articles can be recalled according to the i2i collaborative filtering recall method, so that the articles recalled in this way can better meet the personalized needs of users.

[0076] In order to alleviate the display position deviation and time deviation of articles recalled through the i2i collaborative filtering recall method, the list of articles clicked by the user in history is used to construct the first article pairs of articles in this list, and the similarity scores of each article pair are calculated. This can include: counting the number of co-occurrences of each first article pair in the list of articles clicked by the user in history, and the number of article appearances in each first article pair, and calculating the initial similarity score of each first article pair; weighting the initial similarity score in turn by the display position information weight, click time weight, and creation time weight corresponding to the first article to obtain a weighted similarity score, and using the weighted similarity score as the similarity score of the first article pair.

[0077] When calculating the initial similarity score of each of the first article pairs, the article similarity score of the first article pair is calculated based on the number of co-occurrences of the article pair in the list of articles clicked by the user in history, and the number of article appearances in the first article pair. For example, the formula for calculating the similarity of articles i and j is: Wij = (number of co-occurrences / number of article appearances).

[0078] When considering the influence of the display position information, article click time, and article creation time of the article in the first article pair on the initial similarity score, the initial similarity score is weightedly calculated by the display position information weight, click time weight, and creation time weight corresponding to the first article, and the similarity score of the article after the three weight calculations is obtained; wherein, the formula for calculating the similarity between article i and j is: Wij = (number of co-occurrences / number of article appearances)*loc_weight*click_time_weight*createtime_weight; wherein, the display position weight coefficient is calculated as loc_weight = 0.93*(abs(loci-locj)-1), loc represents the position of the article in the display list; the click time weight is calculated as: click_time_weight = e*(abs(click_timei-click_timej)), click_timei represents the time when the user clicks on the article; the creation time weight is calculated as: createtime_weight = e**(abs(createtimei-createtimej)), createtimei is the time when article i was created.

[0079] Here, when calculating the initial similarity scores of each of the first article pairs, the weight calculation of multiple factors is considered. The weights of the three main factors are the display position information weight, the click time weight, and the article creation time weight. The initial similarity scores are weightedly calculated using these three weights, which can alleviate the article display position deviation and time deviation.

[0080] 2014. Recall the articles to be pushed whose similarity scores of the first article pairs are greater than a preset threshold to obtain a second recalled article set.

[0081] For example, k1 is a preset threshold, where the articles whose similarity scores of the first article pair are greater than the previous k1 articles are recalled, and the first k1 most similar articles obtained are used as articles in the second recalled article set.

[0082] 2015. According to the popular article information, the articles to be pushed are recalled to obtain a third recalled article set.

[0083] It should be noted that the hot article information is counted daily, so the articles to be pushed recalled according to the hot article information are also updated daily. Here, in order to avoid too many articles in the third recalled article set, the number of articles in the third recalled article set is screened to be the same as the preset threshold.

[0084] 2016. The first recalled article set, the second recalled article set and the third recalled article set are combined into the preliminary screening article set.

[0085] Based on the information of popular articles, the preference characteristics of users' interests, and the articles clicked by users in history, the recalled articles obtained are more in line with the users' personalized needs.

[0086] 202. Construct a second article pair of articles in the initially screened article set according to the user's historical click records on each article in the initially screened article set.

[0087] For example: based on the user's historical click records on each article in the initial screening article set, it is determined that the user has read [1, 2, 3, 4] articles in chronological order, then the second article pair obtained is: [[1,2],[2,3][3,4]].

[0088] 203. Use the second article pair as a neighbor edge of a graph data network, and use the articles in the second article pair as two end nodes of the neighbor edge to construct the graph data network.

[0089] In order to expand the sample set by mining more unbiased data through the graph data network, thereby achieving the purpose of alleviating selective bias, the second article pair is used as the neighbor edge of the graph data network, and the articles in the second article pair are used as the two end nodes of the neighbor edge to construct the graph data network, which may include: calculating the article similarity score in the second article pair according to the number of co-occurrences of the second article pair, the number of appearances of the articles in the second article pair, and the popularity corresponding to the articles in the second article pair; using the article similarity score in the second article pair as the edge weight to perform weighted calculation on the neighbor edge of the graph data network corresponding to the second article, to obtain the weighted neighbor edge of the graph data network.

[0090] In the process of building a graph data network, a popularity penalty is introduced when calculating the article similarity score in the second article pair. Introducing a popularity penalty when making the edge weights of the graph data network can make it more likely to reach low-popularity articles during random walks, thereby alleviating the Matthew effect, accurately mining user interests and optimizing the subsequent recommendation ecosystem.

[0091] Among them, the calculation formula for the article similarity score in the second article pair is: Wij = Cij / Ci, where Wij is the article similarity score, Cij is the co-occurrence frequency number of the two articles in the ij pair, and Ci is the click occurrence number of the i-th article; Since the popularity corresponding to the articles in this embodiment corrects the article similarity score in the second article pair, therefore, the calculation formula for the article similarity score in the second article pair in this embodiment is: Wij = Wij / log(1 + pipj), where Pi represents the popularity of the i-th article.

[0092] It should be noted that in this embodiment, the text content similarity vectors of the two articles in each second article pair learned through word2Vector are used to calculate the k articles with the closest similarity according to the similarity vectors, and k edges are constructed for these k articles with the closest similarity. Among them, the edge weights of the k edges are respectively the article similarity scores in the second article pair corresponding to the two articles.

[0093] In addition, during the construction of the graph data network, to meet the requirements of bfs and dfs random walks, it also includes constructing a probability transition matrix between articles, and then modifying the edge weights through the probability transition matrix to adjust the random walk strategy to meet the requirements of DFS or BFS random walks. The following is an example for illustration, see Figure 4 , X1 is the common neighbor of node V and node t, and the weight of the edge V->X1 is set to 1; t is the previous node, and the weight of the edge V->t is set to the return parameter p: if p>1, then the next step tends to visit the common neighbor; if p<1, then the next step tends to return to the previous node. X2 and X3 are other first-order neighbor nodes of node V, and the weights of the edges are set to the in-out parameter q: if q>1, then the next step tends to visit the common neighbor; if q<1, then the next step tends to visit other first-order neighbor nodes. Furthermore, by using these two parameters p and q, the random walk strategy can be adjusted to achieve DFS or BFS.

[0094] Among them, during DFS random walk, let p = 1 and q = 0.5, making it more likely to visit its neighbors. At this time: P(visiting other first-order neighbor nodes)>P(returning to the previous node) = P(visiting the common neighbor); During BFS random walk, let p = 1 and q = 2.0, making it more likely to visit its previous node and the common neighbor. At this time: P(visiting other first-order neighbor nodes)<P(returning to the previous node) = P(visiting the common neighbor).

[0095] 204. Obtain a sample candidate set by randomly walking on the graph data network.

[0096] In order to construct an unbiased sample candidate set, the article node sequence is generated by randomly walking on the graph data network, and the articles on the article node sequence are aggregated into a sample candidate set. Figure 5 , which may include:

[0097] 2041. Use the edge weight corresponding to the neighbor edge between the two nodes as the jump probability.

[0098] 2042. Obtain a wandering path according to the node connection structure of the jump probability on the graph data network.

[0099] It should be noted that the edge weights corresponding to the neighbor edges between nodes can be understood as the probability value of the next jump. For example, an article in the graph is connected to 5 other articles, so there are 5 edges and 6 nodes. For example, the weights of the 5 edges are 0.23, 0.56, 0.67, 0.12, and 0.45 respectively. Then the next article of the random walk will jump based on this weight as a jump probability. For example, the edge weight of 0.67 is more likely than 0.12.

[0100] 2043. Generate an article node sequence according to the wandering path, and determine a sample candidate set, wherein the sample candidate set includes at least one pair of second article pairs corresponding to the wandering path.

[0101] Here, when performing the BFS and DFS random walks, random walks are performed according to the probability transfer matrices corresponding to steps 2041 and 2042 to generate n candidate sets.

[0102] In order to eliminate the problem of selectivity bias of articles in the sample candidate set, after generating an article node sequence according to the wandering path and determining the sample candidate set, the method includes: determining the articles to be predicted corresponding to each article in the sample candidate set according to the user's historical click records on each article in the preliminary screening article set; and adding all the obtained articles to be predicted to the sample candidate set to obtain an updated sample candidate set.

[0103] It should be noted that: based on the user's historical click records on each article in the initial screening article set, the articles to be predicted corresponding to each article in the sample candidate set are determined. It can be understood that: the user clicked on 5 articles, in order a->b->c->d->e, and after clicking d, the next article that the user may like to read is predicted. Based on the time point d, the articles that the user may read can be obtained as a sample set through random walks, such as e, h, g, k. Among them, e is a positive sample, and h, g, k are its negative samples. This implementation can replace the general method of directly constructing a set of positive and negative samples using click and exposure data.

[0104] 205. The initially screened article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model.

[0105] In addition, when obtaining the training sample set, the training sample set is also combined with the customer's life cycle to explore the customer's reading interests at different stages.

[0106] For example, the customer's life cycle can be divided into the initial free insurance transfer payment, the mid-term family protection conversion period, the up-sales period, the late period, the churn warning period and the insurance cancellation recovery period. By introducing the customer life cycle, we can more accurately explore the reading interests and needs of customers at different stages.

[0107] Here, the training feature data of the sample is constructed based on user information, user behavior data, and article content information. In addition to using the user's basic information, such as age, gender, etc., the basic information of the article, such as the number of words, category, etc., and the user's behavior data, such as the number of readings, etc., it also combines the user's marketing life cycle stage of health insurance to explore the differentiated reading needs of users at different stages. Generally speaking, it is divided into 5 stages, namely, the free insurance conversion payment period, the family protection conversion period, the upward sales period, the loss warning period, and the expansion and recovery period. There will be 5 models to score users in these 5 stages, and the model with the highest score will be used as the sample feature to enter the model to reflect the user's marketing life cycle stage.

[0108] An embodiment of the present invention provides another information recommendation model training method. Compared with the prior art, the embodiment of the present invention constructs a second article pair of articles in the preliminary screening article set based on the user's historical click records on each article in the preliminary screening article set, and then uses the second article pair as the neighbor edge of the graph data network, and uses the articles in the second article pair as the two end nodes of the neighbor edge to construct a graph data network; finally, by randomly walking on the graph data network, more unbiased data is mined to expand the sample set, thereby achieving the purpose of alleviating selective bias.

[0109] Furthermore, as a response to the above Figure 1 The implementation of the method shown in the figure, the embodiment of the present invention provides an information recommendation model training device, such as Figure 6 As shown, the device comprises:

[0110] The recall module 61 is used to screen the articles to be pushed according to the popular article information and the user behavior data of the user, and obtain a preliminary screened article set;

[0111] A graph data network construction module 62, configured to construct a graph data network of articles according to the relationships between the articles in the initially screened article set;

[0112] A candidate sample acquisition module 63, configured to acquire a sample candidate set by randomly walking on the graph data network;

[0113] The model training module 64 is used to collect the preliminary screening article set and the sample candidate set into a training sample set, and train the information recommendation model to be trained to obtain an information recommendation model.

[0114] The embodiment of the present invention provides an information recommendation model training device. Compared with the prior art, the embodiment of the present invention obtains a preliminary screening article set by screening articles to be pushed according to popular article information and user behavior data of users; then, a graph data network of articles is constructed according to the relationship between each article in the preliminary screening article set, and a sample candidate set is obtained by randomly walking on the graph data network; finally, the preliminary screening article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model that can send recommended articles to a computer device. The present invention increases user stickiness by taking a preliminary screening article set that is screened according to user behavior data, and constructing a graph data network of articles for each article in the preliminary screening article set, and then obtaining a sample candidate set that meets the user's personalized needs by randomly walking on the graph data network, thereby increasing user stickiness of the information recommendation model trained by the training sample set that is aggregated by the preliminary screening article set and the sample candidate set.

[0115] As the above Figure 2 The implementation of the method shown in the figure, the embodiment of the present invention provides another information recommendation model training device, such as Figure 7 As shown, the device comprises:

[0116] The recall module 71 is used to screen the articles to be pushed according to the popular article information and the user behavior data of the user, and obtain a preliminary screening article set;

[0117] A second article pair construction unit 721, configured to construct a second article pair of articles in the primary screening article set according to the user's historical click records on each article in the primary screening article set;

[0118] A graph data network construction unit 722, configured to use the second article pair as a neighbor edge of the graph data network and the articles in the second article pair as two end nodes of the neighbor edge to construct the graph data network;

[0119] A candidate sample acquisition module 73, used to acquire a sample candidate set by randomly walking on the graph data network;

[0120] The model training module 74 is used to collect the preliminary screening article set and the sample candidate set into a training sample set, and train the information recommendation model to be trained to obtain an information recommendation model.

[0121] Furthermore, the recall module 71 includes:

[0122] A user information determination unit, configured to determine, based on the user behavior data, the preference features that the user is interested in, and a list of articles that the user has historically clicked on;

[0123] An interest recall unit, configured to recall the articles to be pushed according to the preference characteristics of the user, to obtain a first recalled article set;

[0124] A first article pair construction unit, configured to construct first article pairs of articles in a list of articles clicked by the user historically, and calculate a similarity score of each of the first article pairs;

[0125] A collaborative filtering recall unit, configured to recall the articles to be pushed whose similarity scores of the first article pairs are greater than a preset threshold, to obtain a second recalled article set;

[0126] A hot recall unit, used for recalling the articles to be pushed according to the hot article information, to obtain a third recalled article set;

[0127] The first recalled article set, the second recalled article set and the third recalled article set are combined into the initial screening article set.

[0128] Furthermore, the first article pair construction unit includes:

[0129] an initial similarity score calculation subunit, used to count the number of co-occurrences of each first article pair in the list of articles clicked by the user in history, and the number of article appearances in each first article pair, and calculate the initial similarity score of each first article pair;

[0130] The first similarity weighting subunit is used to perform weighted calculation on the initial similarity score in sequence according to the display position information weight, click time weight, and creation time weight corresponding to the first article, to obtain a weighted similarity score, and to use the weighted similarity score as the similarity score of the first article pair.

[0131] Furthermore, the graph data network construction unit 722 includes:

[0132] A second similarity weighting subunit is used to calculate the article similarity score in the second article pair according to the number of co-occurrences of the second article pair, the number of occurrences of the articles in the second article pair, and the popularity corresponding to the articles in the second article pair;

[0133] The neighbor edge weighting processing subunit is used to perform weighted calculation on the neighbor edges of the graph data network corresponding to the second article using the article similarity scores in the second article pair as edge weights to obtain weighted neighbor edges of the graph data network.

[0134] Furthermore, the candidate sample acquisition module 73 includes:

[0135] A jump probability determination unit, used to use the edge weight corresponding to the neighbor edge between two nodes as the jump probability;

[0136] A wandering path acquisition unit, configured to acquire a wandering path according to a node connection structure of the jump probability on the graph data network;

[0137] The sample candidate set determination unit is used to generate an article node sequence according to the walking path and determine a sample candidate set, wherein the sample candidate set includes at least one pair of second article pairs corresponding to the walking path.

[0138] Furthermore, the candidate sample acquisition module 73 also includes:

[0139] A unit for determining articles to be predicted, configured to determine articles to be predicted corresponding to the articles in the sample candidate set according to the historical click records of the user on the articles in the preliminary screening article set;

[0140] The sample candidate set updating unit is used to add all the obtained articles to be predicted to the sample candidate set to obtain an updated sample candidate set.

[0141] An embodiment of the present invention provides an information recommendation model training device. Compared with the prior art, the embodiment of the present invention constructs a second article pair of articles in the preliminary screening article set according to the user's historical click records on each article in the preliminary screening article set, and then uses the second article pair as the neighbor edge of the graph data network, and uses the articles in the second article pair as the two end nodes of the neighbor edge to construct the graph data network; finally, by randomly walking on the graph data network, more unbiased data is mined to expand the sample set, thereby achieving the purpose of alleviating selective bias.

[0142] According to an embodiment of the present invention, a storage medium is provided, wherein the storage medium stores at least one executable instruction, and the computer executable instruction can execute the information recommendation model training method in any of the above method embodiments.

[0143] Based on the above Figure 1 The method shown and Figure 6 The embodiment of the device shown in the figure, the embodiment of the present invention also provides a physical structure diagram of a computer device, such as Figure 8 As shown, the computer device includes: a processor 81, a memory 82, and a computer program stored in the memory 82 and executable on the processor, wherein the memory 82 and the processor 81 are both arranged on a bus 83, and the processor 81 implements the following steps when executing the program: based on the popular article information and the user behavior data of the user, the articles to be pushed are screened to obtain a preliminary screening article set; based on the relationship between the articles in the preliminary screening article set, a graph data network of the articles is constructed; by randomly walking on the graph data network, a sample candidate set is obtained; the preliminary screening article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model.

[0144] Through the technical solution of the present invention, the present invention obtains a preliminary screening article set by screening articles to be pushed according to popular article information and user behavior data of users; then, according to the relationship between each article in the preliminary screening article set, a graph data network of articles is constructed, and a sample candidate set is obtained by randomly walking on the graph data network; finally, the preliminary screening article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model that can send recommended articles to a computer device. The present invention increases user stickiness by combining the preliminary screening article set screened according to user behavior data, and constructing a graph data network of articles for each article in the preliminary screening article set, and then obtaining a sample candidate set that meets the user's personalized needs by randomly walking on the graph data network, thereby increasing user stickiness of the information recommendation model trained by the training sample set aggregated from the preliminary screening article set and the sample candidate set.

[0145] Obviously, those skilled in the art should understand that the above modules or steps of the present invention can be implemented by a general computing device, they can be concentrated on a single computing device, or distributed on a network composed of multiple computing devices, and optionally, they can be implemented by a program code executable by a computing device, so that they can be stored in a storage device and executed by the computing device, and in some cases, the steps shown or described can be executed in a different order than here, or they can be made into individual integrated circuit modules, or multiple modules or steps therein can be made into a single integrated circuit module for implementation. Thus, the present invention is not limited to any specific combination of hardware and software.

[0146] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for training an information recommendation model, characterized in that: include: According to the popular article information and the user's user behavior data, the articles to be pushed are screened and recalled, the popular article information, the preference features of the user's interest in the user behavior data, and the recalled article sets corresponding to the list of the user's historical clicked articles are obtained, and a preliminary screening article set is obtained based on multiple recalled article sets, wherein the similarity of the article pairs in the recalled article set corresponding to the list of the user's historical clicked articles is greater than a preset threshold; Constructing a graph data network of articles according to the relationship between the articles in the preliminary screening article set; wherein the graph data network includes interconnected article nodes; Generate an article node sequence by randomly walking on the graph data network, and collect the articles on the article node sequence into a sample candidate set; The preliminary screening article set and the sample candidate set are aggregated into a training sample set, and the information recommendation model to be trained is trained to obtain an information recommendation model.

2. The information recommendation model training method according to claim 1, characterized in that: The method of screening the articles to be pushed according to the popular article information and the user behavior data of the user to obtain the preliminary screening article set includes: Determine, based on the user behavior data, the preference features that the user is interested in, and a list of articles that the user has historically clicked on; Recalling the articles to be pushed according to the preference features of the user, to obtain a first recalled article set; According to the list of articles clicked by the user in history, construct first article pairs of articles in the list, and calculate the similarity score of each of the first article pairs; Recalling the articles to be pushed whose similarity scores of the first article pairs are greater than a preset threshold to obtain a second recalled article set; Recalling the articles to be pushed according to the popular article information to obtain a third recalled article set; The first recalled article set, the second recalled article set and the third recalled article set are combined into the initial screening article set.

3. The information recommendation model training method according to claim 2, characterized in that: The step of constructing first article pairs of articles in the list according to the list of articles clicked by the user historically, and calculating similarity scores of each of the first article pairs includes: Counting the number of co-occurrences of each of the first article pairs in the list of articles clicked by the user in history and the number of article appearances in each of the first article pairs, and calculating an initial similarity score of each of the first article pairs; The initial similarity score is weightedly calculated in sequence by the display position information weight, click time weight, and creation time weight corresponding to the first article to obtain a weighted similarity score, and the weighted similarity score is used as the similarity score of the first article pair.

4. The information recommendation model training method according to claim 1, characterized in that: The step of constructing a graph data network of articles according to the relationship between the articles in the initially screened article set includes: constructing a second article pair of articles in the initially screened article set according to the user's historical click records on each article in the initially screened article set; The graph data network is constructed by using the second article pair as a neighbor edge of a graph data network and using the articles in the second article pair as two end nodes of the neighbor edge.

5. The information recommendation model training method according to claim 4, characterized in that: The step of using the second article pair as a neighbor edge of a graph data network and using the articles in the second article pair as two end nodes of the neighbor edge to construct the graph data network includes: Calculating article similarity scores in the second article pair according to the number of co-occurrences of the second article pair, the number of occurrences of articles in the second article pair, and the popularity corresponding to the articles in the second article pair; The article similarity scores in the second article pair are used as edge weights to perform weighted calculation on the neighbor edges of the graph data network corresponding to the second article to obtain weighted neighbor edges of the graph data network.

6. The information recommendation model training method according to claim 5, characterized in that: The method generates an article node sequence by randomly walking on the graph data network, and collects the articles on the article node sequence into a sample candidate set, including: The edge weight corresponding to the neighbor edge between the two nodes is used as the jump probability; Acquire a wandering path according to a node connection structure of the jump probability on the graph data network; An article node sequence is generated according to the walking path, and a sample candidate set is determined, wherein the sample candidate set includes at least one second article pair corresponding to the walking path.

7. The information recommendation model training method according to claim 6, characterized in that: After generating the article node sequence according to the walking path and determining the sample candidate set, the method includes: Determine the articles to be predicted corresponding to the articles in the sample candidate set according to the historical click records of the user on the articles in the preliminary screening article set; All the obtained articles to be predicted are added to the sample candidate set to obtain an updated sample candidate set.

8. An information recommendation model training device, characterized in that: include: A recall module is used to screen and recall articles to be pushed based on popular article information and user behavior data of the user, obtain the popular article information, and the recalled article sets corresponding to the preference features of the user's interest in the user behavior data and the list of articles that the user has historically clicked, and obtain a preliminary screening article set based on multiple recalled article sets, wherein the similarity of article pairs in the recalled article set corresponding to the list of articles that the user has historically clicked is greater than a preset threshold; A graph data network construction module, used to construct a graph data network of articles according to the relationship between the articles in the preliminary screening article set; wherein the graph data network includes interconnected article nodes; A candidate sample acquisition module, used to acquire a sample candidate set by randomly walking on the graph data network; The model training module is used to collect the preliminary screening article set and the sample candidate set into a training sample set, and train the information recommendation model to be trained to obtain the information recommendation model.

9. A storage medium, wherein at least one executable instruction is stored in the storage medium, and the executable instruction enables a processor to perform operations corresponding to the information recommendation model training method as described in any one of claims 1-7.

10. A computer device comprising: A processor, a memory, a communication interface and a communication bus, wherein the processor, the memory and the communication interface communicate with each other via the communication bus; The memory is used to store at least one executable instruction, and the executable instruction enables the processor to perform operations corresponding to the information recommendation model training method according to any one of claims 1-7.

Images