An advertisement item recommendation method and system

Abstract

The application relates to an advertisement commodity recommendation method and system, wherein the method comprises the following steps: obtaining a plurality of commodities in a user log, and constituting a first commodity sequence in chronological order; generating a commodity weight graph according to a plurality of first commodity sequences of a plurality of users; obtaining a plurality of second commodity sequences from the commodity weight graph, wherein the second commodity sequence comprises a plurality of commodities connected in turn through weights; calculating commodity vectors in the plurality of second commodity sequences, and classifying the commodity vectors into common commodity vectors and advertisement commodity vectors; respectively calculating the distance between each common commodity vector and advertisement commodity vector, obtaining a preset number of advertisement commodity vectors with the minimum distance, and obtaining advertisement commodities similar to each common commodity; and when a condition of recommending advertisement commodities to a user is met, recommending advertisement commodities similar to commodities in the user log of the user to the user. The application improves the recommendation effect of advertisement commodities and improves the traffic utilization efficiency.

Description

Technical Field

[0001] This invention relates to the field of internet advertising, and in particular to a method and system for recommending advertised products. Background Technology

[0002] In the field of internet advertising recommendation, recommendation systems typically consist of two parts: product recall and ranking. Common recall schemes include collaborative filtering, content-based recall using natural language processing, and vector recall. Vector recall primarily involves obtaining embedded vectors of users and products, calculating vector distances to obtain the nearest neighbor vector set between users and products or between products, thus identifying the products corresponding to the vectors, and then ranking and recommending them in order. In the context of secondhand e-commerce, most ordinary and advertised products are unique items with distinctive characteristics. When using such unique items for modeling to obtain corresponding product vectors during product recall, the sparsity of the products results in product vectors that cannot accurately represent the products, leading to poor recommendation performance, low traffic utilization efficiency, and no increase in revenue for e-commerce platforms and advertisers. Summary of the Invention

[0003] In view of the technical problems existing in the prior art, the present invention proposes an advertising product recommendation method and system to improve the recommendation effect of advertising products and enhance the efficiency of advertising traffic utilization.

[0004] To address the aforementioned technical problems, according to one aspect of the present invention, an advertising product recommendation method is provided, comprising the following steps: obtaining multiple products from user logs and constructing a first product sequence in chronological order; generating a product weight graph based on multiple first product sequences from multiple users, wherein the product weight graph uses products as nodes and the weights corresponding to the similarity between two products as nodes connecting edges, and the products corresponding to the nodes include at least advertising products; obtaining multiple second product sequences from the product weight graph according to preset conditions, wherein the second product sequences include multiple products sequentially connected by weights; calculating product vectors in the multiple second product sequences and classifying them into ordinary product vectors and advertising product vectors; calculating the distance between each ordinary product vector and each advertising product vector, obtaining a preset number of advertising product vectors with the smallest distance, thereby obtaining advertising products similar to each ordinary product; and when the conditions for recommending advertising products to users are met, recommending advertising products similar to the products in the user's user logs.

[0005] To address the aforementioned technical problems, according to another aspect of the present invention, an advertising product recommendation system is provided, comprising: a product acquisition module, a graph construction module, a product sequence acquisition module, a vector acquisition module, a vector distance calculation module, and a recommendation module. The product acquisition module is configured to acquire multiple products from user logs, forming a first product sequence in chronological order. The graph construction module is configured to generate a product weight graph based on multiple first product sequences from multiple users. The product weight graph uses products as nodes and the weights corresponding to the similarity between two products as edges connecting the nodes. The products corresponding to the nodes include at least advertising products. The product sequence acquisition module... The module obtains multiple second product sequences from the product weight graph according to preset conditions. The second product sequence includes multiple products connected sequentially by weights. The vector acquisition module is configured to calculate product vectors in the multiple second product sequences and classify them into ordinary product vectors and advertising product vectors. The vector distance calculation module calculates the distance between each ordinary product vector and the advertising product vector, and obtains a preset number of advertising product vectors with the smallest distance, thereby obtaining advertising products similar to each ordinary product. When the conditions for recommending advertising products to users are met, the recommendation module recommends advertising products similar to the products in the user's user log.

[0006] The advertising product recommendation method and system provided by this invention addresses the issues of low inventory and sparse clicks in the secondhand market, as well as the unique characteristics of product information. It employs a co-construction scheme of advertising products and ordinary products, utilizing user click behavior information on ordinary products to improve the quality of advertising product vectors, enabling them to better represent the advertising products. This invention is linked to the business rules of secondhand goods and integrates influential features into the graph construction process in the form of weights, further improving the generalization ability of vectors in secondhand scenarios. This allows the recalled advertising products to adapt to the secondhand trading scenario, thereby improving recommendation effectiveness and increasing online click-through rate (CTR). This results in more clicks for the same exposure traffic, improving recommendation quality for users while increasing revenue for both the platform and advertisers. Attached Figure Description

[0007] The preferred embodiments of the present invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0008] Figure 1 This is a flowchart of an advertising product recommendation method according to an embodiment of the present invention;

[0009] Figure 2 This is a flowchart of recalling specific advertised products according to an embodiment of the present invention;

[0010] Figure 3A This is a schematic diagram of a first product sequence consisting of four products clicked by user1 according to an embodiment of the present invention;

[0011] Figure 3B This is a schematic diagram of a first product sequence consisting of products clicked by another user, provided according to an embodiment of the present invention;

[0012] Figure 4 This is a schematic diagram of a commodity weighting chart provided according to an embodiment of the present invention;

[0013] Figure 5 This is a schematic diagram of a second product sequence provided according to an embodiment of the present invention;

[0014] Figure 6 This is a schematic diagram illustrating the correspondence between a product and its corresponding similar advertised products according to an embodiment of the present invention;

[0015] Figure 7 This is a block diagram illustrating the principle of an advertising product recommendation system according to an embodiment of the present invention;

[0016] Figure 8 This is a block diagram illustrating the principle of a product acquisition module according to an embodiment of the present invention;

[0017] Figure 9 This is a schematic diagram of a graph construction module according to an embodiment of the present invention; and

[0018] Figure 10 This is a block diagram of a vector acquisition module according to an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] In the following detailed description, reference can be made to the accompanying drawings, which form part of this application and illustrate specific embodiments of the present application. In the drawings, similar reference numerals describe substantially similar components in different figures. Specific embodiments of the present application are described in sufficient detail below to enable those skilled in the art to implement the technical solutions of the present application. It should be understood that other embodiments may also be utilized, or structural, logical, or electrical changes may be made to the embodiments of the present application.

[0021] like Figure 1 The diagram shown is a flowchart according to an embodiment of the present invention. The advertising product recommendation method provided by the present invention includes the following steps:

[0022] Step S1: Recall specific advertised products to target users.

[0023] Step S2: Determine if the number of advertised products to be recommended is more than one. If so, sort them in step S3 to determine the target advertised product. If not, determine the current advertised product as the target advertised product in step S4 and proceed to step S5.

[0024] Step S5: Determine whether the recommendation conditions are met. If the recommendation conditions are met, recommend the target recommended advertising product to the target user in step S6. If the conditions are not met, wait and return to step S5.

[0025] In step S1, as follows Figure 2 One embodiment shown is a recall of specific advertised products.

[0026] Step S11: Query the user log to determine the first product sequence clicked by the user in history. The platform using this invention, such as a second-hand e-commerce platform, records user behavior to form a user log. This log records user actions on the platform, such as searching for products, clicking, viewing products, placing orders, posting comments after purchase, etc. For each action, corresponding information is recorded, such as the time and keywords of the search, the time and product information of the clicked product, including product name, product category, product price, city of the product, and transaction location. By querying the user log, the products clicked by the user within a preset time period, the click time, and the corresponding product information can be obtained. A product sequence can be obtained in chronological order; for distinction, this product sequence is called the first product sequence. The number of products in this sequence can be determined by the preset time period or a preset quantity value. For example, the first product sequence can be obtained by arranging all products clicked by the user the previous day in chronological order, or by generating the first product sequence from the latest 10 or 20 products in the user log.

[0027] Step S12: Obtain product information for each product in the first product sequence; the product information includes, for example, product category, product price, and city where the product is located, and identify whether the product is a regular product or an advertised product, and mark each product in the first product sequence as either a regular product or an advertised product. In some embodiments, when the number of users is small, to obtain a sufficient number of advertised products, the first product sequence must include at least one advertised product. After identification, if there are no advertised products in the sequence, products are re-obtained, such as by increasing the time period for obtaining products, increasing the number of products, or shifting the start time of obtaining products, until an advertised product is obtained.

[0028] Step S13: Divide the first product sequence into two or more sub-sequences according to preset rules. For example, divide the first product sequence into multiple sub-sequences according to user session information, with each sub-sequence containing multiple products. Generally, when a user has the intention to purchase a target product, such as a mobile phone, they are likely to click on multiple mobile phone products to select one within a short period of time. If the user repeats the behavior after a certain period of time, their intention may have changed, and they may no longer want to select a mobile phone. Therefore, by setting a reasonable threshold to segment the first product sequence, the products in the sub-sequences can be more correlated and have greater similarity.

[0029] like Figure 3A As shown, the four most recently clicked items by user1 were obtained by querying online logs. Identification reveals that items 1-3 are regular items, and ad1 is an advertised item. These four items constitute the first item sequence S1, and information such as the category, price, and city of each item is obtained. The dashed lines indicate that the time interval between user1's clicks on ad1 and item2 exceeded a set threshold, i.e., exceeded the duration of one session. Therefore, the first item sequence S1 is divided into two subsequences, S11 and S12.

[0030] Figure 3A This segmentation is based on user session information recorded in the system. However, it can also be based on other rules, such as the density of click time distribution for products. The click time of each product is obtained, the time intervals between two adjacent products are compared, and products with time interval differences less than a threshold are grouped together. This threshold matches the time intervals of the products in the current sequence. Figure 3B The image shows the first product sequence S1 for a user during a day. The time interval between the fifth and sixth products is much longer than the time interval between the first five products, so the first five products are grouped together. Similarly, the eighth and ninth products are divided. Figure 3B The first product sequence S1 shown is divided into three subsequences S11, S12, and S13. Of course, it is also possible that the first product sequence does not need to be divided.

[0031] Step S14: Combine multiple items in a subsequence into pairs in sequence to form item co-occurrence pairs. For example... Figure 3A As shown, there are two co-occurrence pairs of items: (item 1, ad1) and (item 2, item 3). Figure 3B As shown, the first subsequence consists of four co-occurrence pairs of items (item 1, ad1) ... (item 3, ad2), the second subsequence consists of two co-occurrence pairs of items (item 4, ad 3) and (ad3, item 5), and the third subsequence consists of one co-occurrence pair of items (item 6, item 7).

[0032] Step S15: Calculate the weights of co-occurrence pairs of goods based on the goods information. This involves comparing the differences in information about two goods of the same category; goods with greater differences have lower weights, and vice versa. For example, differences can be calculated separately for three categories of goods information: category, price, and city. For example: for category, if two goods belong to the same category, the weight is 1; for price, multiple price difference thresholds can be set to compare price differences, each corresponding to a weight. For example, if the price difference is less than the first price difference threshold, the weight is 1; if the price difference is greater than the first price difference threshold but less than the second price difference threshold, the weight is 0.5; for location, if two goods are located in the same location, the weight for that category is 1, and distance thresholds are set based on distance, with different distance thresholds corresponding to different weights. For example... Figure 3A In the first product sequence S1, the two co-occurrence pairs (item1, ad1) are of different categories, have a price difference within the first threshold, and are located in the same city, so their weight is 2. The two co-occurrence pairs (item2, item3) are of the same category, have a price difference within the threshold, and are located in the same city, so their weight is 3. The weights are then standardized to between 0 and 1.

[0033] Step S16: Obtain the co-occurrence pairs of products for multiple users and their weights, and construct a product weight graph with products as nodes and the weights between products as edges. For example... Figure 4 As shown. The numbers on the edges connecting the nodes represent the weights of the two items.

[0034] Step S17: Traverse each node and use a walk algorithm to obtain multiple second product sequences on the weight graph. The walk algorithm takes each node in the product weight graph as the starting point and uses the weight of the edge connected to that node as the probability of the next step. The sequence is recorded as soon as the next node is reached, until a set sequence length is reached or there are no more connected nodes. Figure 4 As shown, starting from item1, there is a 0.75 probability of reaching ad2 and a 0.25 probability of reaching ad1. Assuming ad2 is reached, it is recorded in the sequence. If the set sequence length is 2, the walk terminates, resulting in the sequence {item1, ad2}. If the sequence length is longer, such as 4 items, the walk continues from ad2, resulting in... Figure 5 The sequence shown. Taking node item1 as an example, starting with item2-item5, ad1, and ad2 respectively, yields a sequence. For Figure 4 The weighted graph shown yields six product sequences. To distinguish them from the product sequences obtained in step S11, the reconstructed product sequences are referred to as the second product sequences.

[0035] Step S18: Calculate the vector of each item in the second item sequence. In one embodiment, this is achieved using a vector such as... Figure 4 The product weight graph shown and the resulting multiple second product sequences are used as samples to train a Skip-Gram model. Skip-Gram is an open-source, general-purpose algorithm in the field of natural language processing. In the original algorithm, the Skip-Gram model has a neural network structure, where the weight matrix of the hidden layer is the word vector matrix. For the application scenario of this invention, the product sequences obtained from a product weight graph are used as natural language corpus, and the products are treated as words. The Skip-Gram model is trained using a certain amount of training data, and the product vectors are obtained through the weight matrix of the hidden layer. According to the weight graph, other products from different users are associated through the same product, which not only serves to augment the data (i.e., expand the data) but also helps the model better represent vectors. Therefore, the product vector in this invention is a product information representation that combines user behavior and the correlation between product information (i.e., the aforementioned weights). The similarity between two products can be abstracted as the proximity of two vectors.

[0036] Step S19: Divide the multiple vectors into a set of ordinary product vectors and a set of advertising product vectors according to the attribute identifiers of the tags.

[0037] Step S20: Calculate the distance between each ordinary product vector and the advertised product vector to obtain one or more similar advertised product vectors, and then obtain the corresponding similar advertised products. For example... Figure 6As shown. In one embodiment, the distance between a regular product vector and an advertised product vector is calculated using Faiss (Facebook AISimilarity Search), which performs nearest neighbor retrieval on the regular product vector within the set of advertised product vectors. This yields the advertised product vectors that are closest to the regular product. Nearest neighbor retrieval involves calculating the distance between a regular product vector and all advertised product vectors, and then selecting the top N advertised product vectors with the smallest distances. Faiss is a tool to accelerate this calculation process; other tools can also be used, such as Milvus (an open-source vector similarity search engine) or Annoy (Spotify's open-source approximate nearest neighbor query library).

[0038] Step S21: Based on the target user's historical clicked products (i.e., products in the first product sequence), obtain the corresponding similar advertising products and use them as advertising products to be recommended.

[0039] This invention uses a joint modeling approach for advertised products and ordinary products, which solves the problem of sparse click behavior for advertised products. Furthermore, it integrates product information that has a significant impact on second-hand e-commerce into the weight information, thereby improving scenario adaptability and generalization ability.

[0040] above Figure 2 The embodiment shown is a preferred embodiment, which calculates the similarity between two products by segmenting the product sequence to group products with stronger relevance together. Of course, in some scenarios, such as when the preset time period for obtaining the first product sequence is short, since the probability of a user maintaining a preference for one product within that short time is higher, it is not necessary to segment the first product sequence in this case, and the weights of the two products in the first product sequence can be calculated directly.

[0041] Figure 7 This is a block diagram illustrating the principle of an advertising product recommendation system according to an embodiment of the present invention. In this embodiment, the system includes a product acquisition module 1, a graph construction module 2, a product sequence acquisition module 3, a vector acquisition module 4, a vector distance calculation module 5, and a recommendation module 6. The product acquisition module 1 acquires multiple products from user logs according to a strategy and constructs a first product sequence in chronological order. The strategy can be a preset time period or a preset quantity. When the product acquisition module 1 acquires products using a preset time period, such as... Figure 8As shown, it includes a time point determination unit 11, a product retrieval unit 12, and a product identification unit 13. The time point determination unit 11 determines the product retrieval deadline from the user log according to a preset time period and the current time. For example, if the current time is 9:00 and the preset time period is 24 hours, then the time point determination unit 11 starts from the current time 9:00 and determines that 24 hours later, the time will be 9:00 the previous day. Figure 3B As shown, the time period for reading data is determined from the user log. The product extraction unit 12 is connected to the time point determination unit 11, and extracts multiple products and their information from the user log in chronological order between the current time and the product extraction deadline to obtain a first product sequence. The product identification unit 13 identifies the advertised products in the first product sequence S1 and marks the product attributes in the first product sequence, for example, by... Figure 3B The second, fifth, and seventh items are marked as advertised items, while the rest are ordinary items.

[0042] In another embodiment, the product extraction unit 12 extracts a predetermined number of products and their information from the user log, starting from the current moment, to obtain a first product sequence. For example, it determines that 10 products will be extracted.

[0043] Graph construction module 2 generates a product weight graph based on multiple first product sequences S1 from multiple users. The product weight graph uses products as nodes and the similarity weights between two products as connecting edges. The multiple users are, for example, registered users of a second-hand e-commerce platform. The product acquisition module 1 obtains the first product sequences S1 of all registered users on the platform, and graph construction module 2 integrates all users' first product sequences S1 together to generate the product weight graph. To generate the product weight graph, as follows... Figure 9As shown, the graph construction module 2 includes at least a weight calculation unit 21 and a graph construction unit 22. In one embodiment, the weight calculation unit 21 calculates the similarity between two products in each user's first product sequence S1 based on the product information, and uses the similarity as a weight. Product information includes, for example, product categories such as clothing, books, and furniture; and product location, such as the city or region. In the second-hand market, the city where the product is located is a crucial factor in second-hand transactions, therefore, this invention uses it as a factor in calculating similarity. Product information also includes, for example, price, which is also an important factor in second-hand transactions. When the time period of the user's first product sequence S1 is relatively short, or the number of products is small, the user's product preferences or the target product searched by the user usually remain unchanged, thus the relevance of products clicked within a short period is relatively high. For example, when a user's target product is an Apple mobile phone, their search results are mostly focused on Apple mobile phones. The similarity between two adjacent products in the first product sequence S1 is obtained by calculating the differences in category, price, and city, and this similarity is used as the weight between the two products. In another embodiment, to more accurately determine which products represent the user's product preferences, the graph construction module 2 further includes a sequence segmentation unit 23 and a co-occurrence pair composition unit 24. The sequence segmentation unit 23 divides the first product sequence S1 into multiple sub-sequences based on time information, such as through session information or differences in time intervals between products, grouping the products in the first product sequence S1. The co-occurrence pair composition unit 24 sequentially pairs the products in the sub-sequences into co-occurrence pairs; the weight calculation unit 21 calculates the weights of the two products in each co-occurrence pair. The graph construction unit 22 uses products as nodes and the weights of the two products as edges to connect the two nodes, thereby obtaining a product weight graph.

[0044] In one embodiment, the product sequence acquisition module 3 is a random walk unit. According to the random walk algorithm, starting from each node in the product weight graph, it randomly walks through a preset number of nodes or until there are no nodes to form a second product sequence.

[0045] like Figure 10 As shown, the vector acquisition module 4 includes a semantic model unit 41 and a vector classification unit 42. The semantic model unit 41 uses multiple second product sequences obtained from the product weight map as corpus and calculates the vector for each product using a semantic model, such as the Skip-Gram model. The vector classification unit 42 classifies the vectors output by the semantic model unit into ordinary product vectors and advertising product vectors based on the product's attribute identifier, thus obtaining a set of ordinary product vectors and a set of advertising product vectors. Generally speaking, the number of vectors in the advertising product vector set is much smaller than the number of vectors in the ordinary product vector set.

[0046] The vector distance calculation module 5 calculates the distance between each ordinary product vector and the advertising product vector, obtaining the preset number of advertising product vectors with the smallest distance, thereby obtaining the advertising products similar to each ordinary product. In one embodiment, Faiss (Facebook AI Similarity Search) is used to perform nearest neighbor retrieval based on the product vector in the advertising vector set to obtain the preset number of advertising product vectors with the smallest distance.

[0047] In the advertising system, the aforementioned product acquisition module 1, graph construction module 2, product sequence acquisition module 3, vector acquisition module 4, and vector distance calculation module 5 implement the recall of advertised products. When the recommendation conditions are met, the recommendation module 6 recommends similar advertised products to the user based on the products in the user's user log. The products in the user log can include currently clicked and historically clicked products. In one embodiment, when a user logs into a second-hand platform's app and clicks on a product, the aforementioned recall can retrieve similar advertised products and recommend them, for example, through a dropdown list on the homepage or through similar recommendations on a single product page. When there are multiple advertised products to recommend, they are sorted to obtain an ordered set of advertisements, and recommendations are made to the user starting from the first one; for example, different sorting indicators are used based on the payment method of the advertisements, or similarity values ​​are obtained based on the similarity of the advertised products, and the products are sorted according to the similarity values.

[0048] The advertising product recommendation method and system provided by this invention addresses the issues of low inventory and sparse clicks in the secondhand market, as well as the unique characteristics of product information. It employs a co-construction scheme of advertising products and ordinary products, utilizing user click behavior information on ordinary products to improve the quality of advertising product vectors, enabling them to better represent the advertised products. This invention integrates features that are closely related to and have a significant impact on the business rules of secondhand products into the graph construction process in the form of weights. This further enhances the generalization ability of vectors in the secondhand scenario, allowing the recalled advertising products to adapt to the secondhand trading scenario, thereby improving recommendation effectiveness and increasing online click-through rate (CTR). This results in more clicks for the same amount of exposure traffic, improving recommendation quality for users while simultaneously increasing revenue for the platform and advertisers. An A / B test conducted on the platform's homepage recommendation scenario, compared to the experimental group with 50% traffic, showed an increase of 1700 yuan in advertising costs and a relative increase of 2.78% in CTR.

[0049] The above embodiments are for illustrative purposes only and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the scope of the invention. Therefore, all equivalent technical solutions should also fall within the scope of the invention.

Claims

1. A method for recommending an advertised product, comprising: obtaining a plurality of products in a user log, and forming a first product sequence in chronological order; generating a product weight graph based on a plurality of first product sequences of a plurality of users, the product weight graph having products as nodes and weights corresponding to similarities between two products as edges connecting the nodes, wherein the products corresponding to the nodes at least include an advertised product; randomly walking from each node in the product weight graph to a preset number of nodes to form a second product sequence, the second product sequence including a plurality of products connected in sequence by the weights; calculating product vectors in the plurality of second product sequences obtained from the product weight graph using a natural language semantic model, and classifying the product vectors into common product vectors and advertised product vectors; performing nearest neighbor search of the common product vectors in a set of advertised product vectors, respectively calculating distances between each common product vector and an advertised product vector, and obtaining a preset number of advertised product vectors with minimum distances to obtain advertised products similar to each common product; and when a condition for recommending an advertised product to a user is met, recommending to the user an advertised product similar to a product in the user log of the user.

2. The method of claim 1, further comprising: obtaining a plurality of products in a user log within a preset time period; or obtaining a plurality of products in a user log within a preset number of products from the most recent product.

3. The method of claim 1, further comprising: obtaining product information of each product, and calculating a similarity between two products based on the product information, and using the similarity as a weight. 4.The method of claim 3, wherein the product information includes one or more of a product category, a product price, and a city where the product is located. 5.The method of claim 1, further comprising: dividing the first product sequence into a plurality of subsequences according to time information; grouping products in each subsequence into co-occurrence pairs in chronological order; and calculating a similarity between two products in a co-occurrence pair, and using the similarity as a weight. 6.The method of claim 5, wherein the time information is session information, and products in a session are grouped into a subsequence. 7.The method of claim 5, wherein the time information is a time interval between two products, and when a difference between time intervals of two adjacent pairs of products is greater than a threshold, the pair of products with a larger time interval is split and placed in two subsequences. 8.A system for recommending an advertised product, comprising: a product obtaining module configured to obtain a plurality of products in a user log, and form a first product sequence in chronological order; a graph constructing module configured to generate a product weight graph based on a plurality of first product sequences of a plurality of users, the product weight graph having products as nodes and weights corresponding to similarities between two products as edges connecting the nodes, wherein the products corresponding to the nodes at least include an advertised product; a product sequence obtaining module configured to randomly walk from each node in the product weight graph to a preset number of nodes to form a second product sequence, the second product sequence including a plurality of products connected in sequence by the weights; and ​ a vector obtaining module configured to utilize a natural language semantic model to calculate a commodity vector in a plurality of second commodity sequences obtained from the commodity weight graph, and classify the commodity vector into a common commodity vector and an advertisement commodity vector; a vector distance calculating module configured to calculate a distance between each common commodity vector and advertisement commodity vector, and obtain a preset number of advertisement commodity vectors with minimum distance, so as to obtain an advertisement commodity similar to each common commodity; and a recommendation module configured to recommend an advertisement commodity similar to a commodity in a user log of a user to the user when a condition of recommending an advertisement commodity to the user is met.

9. The system of claim 8, wherein the commodity obtaining module comprises: a time point determining unit configured to determine a commodity extraction cutoff time point from the user log according to a preset time period and a current time; a commodity extraction unit configured to extract a plurality of commodities and commodity information thereof in the user log in a time sequence between the current time and the commodity extraction cutoff time point to obtain a first commodity sequence, or extract a preset number of commodities and commodity information thereof in the user log in a time sequence from the current time to obtain a first commodity sequence; and a commodity identifying unit configured to identify an advertisement commodity in the first commodity sequence, and mark a commodity attribute in the first commodity sequence, the commodity attribute being a common commodity or an advertisement commodity.

10. The system of claim 8, wherein the graph constructing module comprises: a weight calculating unit configured to calculate a similarity between two commodities in a plurality of first commodity sequences of a plurality of users according to commodity information, and take the similarity as a weight; and a graph constructing unit configured to take a commodity as a node, and connect two nodes with a weight between the two commodities as an edge.

11. The system of claim 10, wherein the graph constructing module further comprises: a sequence dividing unit configured to divide the first commodity sequence into a plurality of sub-sequences according to time information; and a co-occurrence pair forming unit configured to combine commodities in a sub-sequence into co-occurrence pairs in order two by two; correspondingly, the weight calculating unit calculates a weight between two commodities in each co-occurrence pair.

12. The system of claim 9, wherein the vector obtaining module comprises: a semantic model unit configured to take a plurality of second commodity sequences obtained from the commodity weight graph as a corpus, and calculate a vector of each commodity through a natural language semantic model; and a vector classifying unit configured to classify the vector output by the semantic model unit into a common commodity vector and an advertisement commodity vector according to an attribute of the commodity.

Images