A method for recommending product conversation sequences based on global directed graph

Abstract

The present invention discloses a commodity conversation sequence recommendation method based on a global directed graph, comprising: step 1, extracting commodity items in an original data set to generate a graph database node set; generating a node relationship set according to the click relationship between items in the original data set and taking the number of occurrences of the click relationship as the attribute value of the node relationship; step 2, generating a global directed graph G=(V, E) of the conversation sequence; wherein V is the graph database node set and E is the node relationship set; step 3, establishing an m-element rating array with an initial value of 0 for each item in a to-be-recommended conversation, wherein each element in the rating array corresponds to a rating value of a to-be-recommended item, and updating the corresponding element in each rating array through the preference propagation of the item; wherein m is the number of items in the original data set; step 4, obtaining the total rating value of all to-be-recommended items in the to-be-recommended conversation, and screening out a plurality of items with high total rating values ​​as the final recommendation results of the to-be-recommended conversation.

Description

Technical Field

[0001] The present invention belongs to the technical field of commodity conversation sequence recommendation, and in particular relates to a commodity conversation sequence recommendation method based on a global directed graph. Background Art

[0002] In the recommendation system, the conversation sequence of goods is a special form of data. It refers to the order in which a user clicks on goods within a period of time. It has the characteristics of large data volume, strong randomness, and many user anonymity. The purpose of product recommendation based on conversation sequence is to predict the next product entity that the current user may be interested in through existing conversations. Traditional recommendation systems often only focus on the last item entity (referred to as item) in the conversation sequence or the item conversion in a few conversations. The strong local preference leads to unsatisfactory recommendation results and even easily arouses user disgust. With the continuous increase in the amount of conversation sequence data and users, the computational cost of making recommendations by traditional methods is also increasing. In the field of recommendation, collaborative filtering has always been enduring and is used by many e-commerce websites such as Amazon to achieve simple and accurate recommendations.

[0003] The most common collaborative filtering algorithms can be divided into two categories: user-based algorithms and item-based algorithms. The core of user-based collaborative filtering is to calculate the similarity between users and neighboring users, and predict the possible clicks of the current user based on the clicks of neighboring users. This requires the user's explicit preferences and historical interactions. In actual recommendation systems, a huge user-item interaction matrix needs to be maintained, which has serious matrix sparsity problems. The addition of new users or items will also lead to recalculation of user similarities, and poor scalability.

[0004] When users are more anonymous, user-based collaborative filtering cannot meet the needs. Unlike user-based collaborative filtering algorithms, item-based collaborative filtering is the best choice besides deep learning methods in session sequence recommendation. It has good recommendation accuracy and low algorithm complexity, but it also has a certain computational cost because it needs to calculate the similarity information of all commodity items. The Markov chain method is a method that predicts the next item based on the recent clicks of the current user's last clicked item. It cannot capture the overall characteristics of the session and the global item click information, and only focuses on the last clicked item. The accuracy of Markov chain-based recommendation is also worse than that of item-based collaborative filtering. In short, traditional recommendation methods are seriously challenged in the case of user anonymity and large-scale data. Therefore, algorithms that can meet the requirements of anonymous use, simple implementation, and high recommendation efficiency have become a hot issue in current research. Summary of the invention

[0005] In view of the defects of the existing commodity conversation sequence recommendation algorithm, the present invention provides a commodity conversation sequence recommendation method based on a global directed graph. One purpose of the present invention is to adopt a global preference propagation strategy to obtain the global influence of each item in the current conversation sequence, obtain the rating information of the items to be recommended in the current conversation, and realize fast and effective item recommendation.

[0006] Another object of the present invention is to further improve the accuracy of item recommendations by improving the parameters corresponding to the later items in the conversation sequence in the process of obtaining the ratings so as to consider the important influence of the time factor of the click on the recommendation.

[0007] The technical solution provided by the present invention is:

[0008] A method for recommending commodity conversation sequences based on a global directed graph comprises the following steps:

[0009] Step 1: extract the commodity items in the original data set and generate a graph database node set; generate a node relationship set based on the click relationship between the items in the session sequence in the original data set and the number of occurrences of the click relationship as the attribute value of the node relationship;

[0010] Step 2: obtaining a global directed graph G=(V, E) of the session sequence according to the graph database node set and the generated node relationship set;

[0011] Among them, V is the graph database node set, and E is the node relationship set;

[0012] Step 3: For each item in the session to be recommended, an m-element rating array with an initial value of 0 is established, wherein each element in the rating array corresponds to a rating value of the item to be recommended; and the corresponding element in the rating array is updated through the preference propagation of the items in the session to be recommended;

[0013] Wherein, m is the number of items in the original data set;

[0014] Step 4: Obtain the total score of each item to be recommended in the session to be recommended according to the score array obtained in step 3, and select multiple items with high total scores as recommended items for the session to be recommended.

[0015] Preferably, in step 3, obtaining the scoring value of the project includes:

[0016] For a session S{v1,v2...v n}Project v i , search for item v in the graph database i 1-hop, 2-hop, ... N-hop set, and update the corresponding score value to the project v according to the global preference propagation strategy iThe corresponding rating array s i middle;

[0017] Among them, v i N-hop collection The jth item in The rating is:

[0018]

[0019] In the formula, yes and The weight of the node relationship, s (N-1) yes 's rating value.

[0020] Preferably, in step 4, according to the session S{v1,v2...v n} Rating array s for each item i , get the global score array of session S:

[0021]

[0022] In the formula, k i Indicates project v i Parameter weights of

[0023] Among them, each element in the global score array s corresponds to an item v i The total rating value of .

[0024] Preferably, k i The setting rule is k n >k n-1 >…k1.

[0025] Preferably, in step 3, a cypher statement is used to search for item v in the graph database. i A collection of 1-hop, 2-hop...N-hop.

[0026] Preferably, in the step 2, after the graph database node set and the node relationship set are imported into the database through the neo4j-import api of Neo4j, a global directed graph of the session sequence is obtained.

[0027] Preferably, the elements in the global scoring array s are arranged in descending order according to the total score values, and the items corresponding to the first t elements in the global scoring array s are screened out as the recommended items for the session.

[0028] Preferably, t=20.

[0029] The beneficial effects of the present invention are:

[0030] The present invention models all conversation sequences of a data set by constructing a global directed graph, uses a preference propagation strategy for the current conversation and considers the time factor to obtain the global impact of the items in the conversation, thereby generating scores for all items to be recommended, and recommends items with high scores. A large number of experiments on two standard data sets have shown that compared with traditional item-based collaborative filtering and Markov chain methods, it performs better under both the P@20 evaluation indicator and the MRR@20 evaluation indicator. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a schematic diagram of an example of database node connection according to the present invention.

[0032] Figure 2 A schematic diagram of the node preference propagation “ripple” described in the present invention.

[0033] Figure 3 A schematic diagram of an example of node preference propagation according to the present invention.

[0034] Figure 4 This is a distribution diagram of the Diginetica dataset session sequence lengths described in the present invention.

[0035] Figure 5 This is a distribution diagram of the conversation sequence length of the Yoochoose dataset described in the present invention.

[0036] Figure 6 It is a schematic diagram comparing the recommended results of each group of Diginetica parameters described in the present invention.

[0037] Figure 7 It is a schematic diagram comparing the recommended results of various groups of parameters of Yoochoose described in the present invention. DETAILED DESCRIPTION

[0038] The present invention is further described in detail below in conjunction with the accompanying drawings so that those skilled in the art can implement the invention with reference to the description.

[0039] The present invention provides a commodity conversation sequence recommendation method based on a global directed graph, and the specific implementation process is as follows:

[0040] Data preprocessing

[0041] Each row of data in the data set contains the session_id (session id), item_id (item id), timeframe (time slice), and eventdate (date) of the click. First, the data set is preprocessed to extract the "graph database node set" and "node relationship set". The preprocessing steps are as follows:

[0042] 1) For each row of click data, extract the items to generate a "graph database node set".

[0043] 2) For each session sequence, sort the items by time slice. Define click relationship: In the same session, every two adjacent items have a click relationship, i.e., "item_id1 (item id1), item_id2 (item id2)".

[0044] 3) Traverse all click relationships in the data set, take the number of occurrences of the click relationship as the attribute value of the node relationship, and define the node relationship: "item_id1 (item id1), item_id2 (item id2), weight (attribute value)", where weight represents the number of occurrences of the click relationship. Extract all node relationships to generate a "node relationship set".

[0045] Finally, the original data set is converted into a “graph database node set” and a “node relationship set”.

[0046] Constructing a global directed graph

[0047] In this embodiment, a global directed graph is constructed based on graph database technology. The two data files obtained after data preprocessing are imported into the graph database through Neo4j's neo4j-importapi (it takes 4.57s to import the Diginetica dataset into the database, and 3.78s to import the Yoochoose dataset into the database). After importing all project nodes and relationships into the database, a global directed graph G = (V, E) of all session sequences of the dataset can be obtained, where V is the graph database node set and E is the node relationship set. Figure 1 This is an example of the connection between project entity nodes in the graph database. The edge between entities has the attribute "click", and the attribute value is the weight weight (the attribute value in the node relationship set, the type is int).

[0048] Item recommendation based on commodity conversation sequence recommendation method based on global directed graph

[0049] Ripple propagation is a typical network propagation model, which embodies the characteristics of one-to-many broadcasting in social networks. Based on the ripple topology model, the present invention obtains the global influence of the project through a global project preference propagation strategy. Global preference propagation is the propagation of the influence of the project in the global directed graph according to the preference propagation rules. One click generates a "water wave" that spreads outward in the directed graph. The "water wave" will have a superposition effect. There are multiple clicks in a session. Finally, the project that is most affected by the "water wave" is found in the graph to make recommendations. The "water wave" will have an attenuation mechanism during the propagation process, otherwise it will never stop spreading. The present invention uses the path length as the basis for the attenuation parameter. An example of the preference propagation "ripple" of a project node is as follows: Figure 2 As shown in the figure, the inner circle is a 1-hop node and the outer circle is a 2-hop node. The influence of propagation from the inside to the outside will gradually weaken.

[0050] The preference propagation and scoring rules of the items in the graph are as follows:

[0051] For each item, create an m-element array with an initial value of 0 (the size of m is the total number of items, and each subscript in the array corresponds to the item id). Each element in the array corresponds to the score value of an item node. The corresponding element value in the array is updated through the preference propagation of the item. The scoring strategy is as follows:

[0052] 1) For v i 1-hop collection For each item in the 1-hop set, the score corresponding to but in Represents v i and The weight of the node relationship Before update value.

[0053] 2)v i 2-hop collection The score for each project node in the 2-hop set is as follows in, yes and The weight of the node relationship, Before update Value, s (1) for 's rating value.

[0054] 3) For v i N-hop collection The score of each item is similar to formula (1):

[0055]

[0056] In the formula, yes and The weight of the node relationship, s (N-1) yes 's rating value.

[0057] The steps of the project recommendation algorithm are as follows:

[0058] 1) For the current session S{v1,v2…v n}Project v i , initialize the corresponding array s with m elements to 0 i , m is the total number of items.

[0059] 2) Use cypher statements to search for v in the graph database i The 1-hop and 2-hop sets of the project are updated according to the global preference propagation strategy. i In the array (1-hop traversal achieved better results in the experiment).

[0060] 3) Through the session S{v1,v2…v n} The n m-element arrays s1, s2…s obtained by node preference propagation n According to formula (2), we can obtain an m-element array s by adding them together. s is the global score of the current session S. We can make recommendations based on the items corresponding to the array indices of the top 20 elements in s.

[0061]

[0062] When considering the time factor, the most recently clicked item in the session sequence has a greater influence, so the parameter k in formula (2) i Set the rule to: n >k n-1 >…k1.

[0063] As a further preference, k i The value selection rules also include k i ≤2k i-1 , to prevent the scores of the top items from having too little influence and reducing the accuracy of the recommendation.

[0064] like Figure 3The following is an example of a session {v1, v2, v3}, where a, b, c, d, e are the relationship weights in the directed graph, and k2 and k3 are the parameter weights corresponding to v2 and v3. At this time, v4 and v5 are 1-hop nodes of v3, v3 and v5 are 1-hop nodes of v2, and v4 is a 2-hop node of v2. According to the global preference propagation strategy, the total score of v3 is v3 =k2×e, v5 total score v5 =k2×a+k3×b, v4 total score v4 =k3×c+k2×(a+d) / 2, and then the priority of recommendation is determined according to the score values ​​of v3, v4 and v5.

[0065] This embodiment uses the two most widely used standard datasets Diginetica and Yoochoose in the research of product conversation sequence recommendation. Diginetica is a public dataset of CIKMCup2016, which is the transaction data of all anonymous users of a European e-commerce platform within 6 months. Yoochoose is a public dataset of the RecSys Challenge 2015 competition, which contains the click stream of an e-commerce website for 6 months. We use the last 7 days of Diginetica's conversations as the test set, and store other conversation sequences as raw data in the graph database. The Yoochoose1 / 64 dataset takes the last day of membership as the test set, and stores other conversation sequences as raw data in the graph database. The overall information of the two datasets is shown in Table 1.

[0066] Table 1 Experimental data information

[0067]

[0068] Evaluation indicators:

[0069] P@20: P@20 (Precision@20) represents the proportion of correct recommendations among the top 20 recommendations. It is widely used in the recommendation field:

[0070]

[0071] Where N represents the total number of recommendations made, C hit Indicates the total number of correct recommendations among the top 20.

[0072] MRR@20: MRR@20 (Mean Reciprocal Rank) is the mean reciprocal ranking. It is the sum of the reciprocals of the top 20 correct recommendations (if there is no correct recommendation, the reciprocal value is set to 0). A larger MRR value indicates that the correct recommendation ranks higher in the recommended results:

[0073]

[0074] Where N represents the total number of recommendations made, and rank(t) represents the position of the correct recommended item in the recommendation list in each recommendation.

[0075] Comparative verification example

[0076] This paper compares the method based on item similarity, Item-KNN, the Bayesian ranking method, BPR-MF, and the Markov chain method, FPMC. The evaluation method is written with reference to the current mainstream evaluation method.

[0077] Among them, Item-KNN finds the k items that are most similar to the current item by calculating the similarity between items. There are generally two ways to calculate item similarity: one is through the co-occurrence rule, and the other is to calculate the cosine value between the item vector representations.

[0078] The co-occurrence rule states that if two items appear in the same session, then they co-occur, and the similarity is calculated based on the number of co-occurrences.

[0079] The general algorithm for calculating similarity by calculating cosine values ​​is as follows:

[0080] STEP 1: For the item being clicked in the conversation sequence, calculate the similarity with the item to be recommended by calculating the cosine value between the item vectors, and select the top k items most similar to the item, where k is also called the number of neighbors of the item.

[0081] STEP 2: Delete the set M of items that the user has clicked from the k neighbors that are most similar to the item, and save the remaining items in the set R.

[0082] STEP 3: Calculate the similarity between every two items in the set R and M, and finally sort the items in R according to the similarity and make recommendations.

[0083] The similarity between items can be calculated offline. In the actual recommendation process, only the item similarity comparison is needed. However, if the amount of item data is large, the cost of offline calculation will be very high. In addition, Item-KNN only calculates the similarity between items and generates predictions based only on the last click, without considering the click order of items in the current session sequence and other items.

[0084] After multiple experiments with different parameter settings on two datasets, the following experimental results are finally obtained and the comparison results with the existing methods are shown in Table 2:

[0085] Table 2 Comparison of recommendation accuracy results on two datasets

[0086]

[0087] As can be seen from Table 2, Item-KNN is superior to most Markov chain-based methods. Most Markov chain-based methods only use the item click information of the last item entity in the session in the most recent session. In the case of large-scale data, it cannot obtain the global click information of a certain item. The SRGDG algorithm provided in the present invention is superior to Item-KNN. On the Yoochoose1 / 64 and Diginetica data sets, P@20 is improved by 30.25% and 6.12%, respectively, and MRR@20 is improved by 33.88% and 15.04%, respectively. In comparison, the MRR@20 value is improved more, which also reflects the shortcomings of Item-KNN that only considers the similarity relationship of items and does not consider the order of precedence. The SRGDG algorithm provided in the present invention forms a global directed graph by modeling all sessions of the data set, diffuses the influence of each item in the session in the global directed graph, and increases the preference parameters of the later items in the session, and finally obtains the recommendation of the current session sequence. This algorithm considers both the order of items in the session and the influence of other items in the session. The experimental results prove that it is effective. Since Yoochoose1 / 64 has fewer items than Diginetica and the correlation between items is stronger, the accuracy of Yoochoose1 / 64 recommendations is much better than that on the Diginetica dataset, which is more similar to the actual situation of large-scale commodities and users. In addition, the graph database can increase the number of commodities and set the click relationship weights according to demand, so the SRGDG method provided in the present invention also has the natural advantages of solving the scalability of the recommendation system and avoiding the sparsity problem of the matrix. Since the global click relationships are visible, the use of the graph database also enhances the interpretability of the recommendation system recommendations.

[0088] Effectiveness test of preference communication strategy:

[0089] The global rating information is obtained in formula (2). A large number of experiments on Diginetica and Yoochoose datasets show that the global preference propagation strategy can improve the recommendation accuracy. n With k n >k n-1 >kn-2 >…k1 rule. And by Figure 4 and Figure 5 As shown in the distribution diagram of the length of conversation sequences of the Diginetica and Yoochoose datasets, the length of a large number of conversation sequences is less than 3 and 4. Therefore, in the experiment, the k corresponding to the last 3 item entities in each conversation are first analyzed. n-2 ,k n-1 ,k n The impact of k1,k on the other project entities n-2 …k n-3 Then set it to 0.

[0090] Different (k n-2 ,k n-1 ,k n The experimental results of the combination of ) on the data set are shown in Table 3 and Table 4 respectively:

[0091] Table 3 Comparison of recommendation accuracy results for each parameter setting in Diginetica

[0092]

[0093]

[0094] Table 4 Comparison of recommendation accuracy results for each group of parameter settings in Yoochoose

[0095]

[0096] The comparison of the experimental results of various parameter settings on the Diginetica dataset and the Yoochoose dataset is shown in the line graph in the order of groups 1-8. Figure 6 , Figure 7 shown.

[0097] In the experiment on the Diginetica dataset, the (0,0,1) group is equivalent to considering only the impact of the last item, with a P@20 value of 30.77%; the (0,0.4,1) group considers the impact of the last two items, but the penultimate item has a smaller impact due to its smaller coefficient, and the experimental results are basically the same as the (0,0,1) group; (0,0.6,1) increases the impact of the penultimate item, and P@20 is 19.07% higher than the previous group, and MRR@20 is 13.25% higher; the (0,0.8,1) group has little improvement compared to the (0,0.6,1) group, indicating that the impact of the penultimate item has achieved a relatively good effect; the (0.2,0.6,1) group has a 3.46% increase in P@20 and a 1.29% increase in MRR@20 compared to the (0,0.8,1) group, as shown in Figure 2. Figure 7As shown in the figure. This shows that considering the third-to-last item will also have an impact and can improve the accuracy of the recommendation. However, the improvement of the subsequent groups is not obvious. Since the improvement of the 5th, 6th, 7th, and 8th groups is smaller than that of the 2nd, 3rd, and 4th groups, considering only the global impact of the last three nodes of the session can already get a good recommendation result. If we continue to set parameters forward, it will not only increase the computational cost of making recommendations, but also the improvement is not significant. However, this has already demonstrated the effectiveness of the global preference propagation strategy that considers the influence of multiple nodes.

[0098] Similarly, in the Yoochoose1 / 64 dataset, considering the last three nodes at the same time can get a better value for the recommendation result. The best P@20 value is obtained in the combination (0.2, 0.4, 1), which is 3.71% higher than (0, 0, 1). Since a relatively good accuracy has been achieved on the Yoochoose1 / 64 dataset, the overall improvement in the comparison of MRR@20 values ​​is not large, or even decreases. After a large number of experiments on two actual datasets, it is confirmed that the preference propagation strategy of multiple entities can effectively improve the accuracy of recommendations.

[0099] Aiming at the disadvantage that the traditional recommendation algorithm in the commodity conversation sequence recommendation is overly dependent on the adjacent click status and loses the overall commodity access trend to a certain extent, this paper proposes a new commodity conversation sequence recommendation algorithm based on a global directed graph. First, a global directed graph of commodity conversation sequence is constructed, in which the nodes are commodities and the arcs between the nodes represent the click order, and the directed graph is stored in a graph database; then, a global preference propagation strategy on the directed graph is given, and the important influence of the time factor of the click on the recommendation is considered; finally, the score of the commodity to be recommended is obtained. On the Yoochoose1 / 64 and Diginetica standard datasets, the algorithm improves the recommendation accuracy by 30.25% and 6.12% compared with the traditional Item-KNN method according to the P@20 standard, and by 33.88% and 15.04% according to the MRR@20 standard. The experimental results show that the global directed graph search and scoring strategy provided by the present invention is effective.

[0100] Although the embodiments of the present invention have been disclosed as above, they are not limited to the applications listed in the specification and the implementation modes, and they can be fully applied to various fields suitable for the present invention. For those familiar with the art, additional modifications can be easily implemented. Therefore, without departing from the general concept defined by the claims and the scope of equivalents, the present invention is not limited to the specific details and the illustrations shown and described herein.

Claims

1. A method for recommending commodity conversation sequences based on a global directed graph, characterized in that: The steps include: Step 1: extract the commodity items in the original data set and generate a graph database node set; generate a node relationship set based on the click relationship between the items in the session sequence in the original data set and the number of occurrences of the click relationship as the attribute value of the node relationship; Step 2: obtaining a global directed graph G=(V, E) of the session sequence according to the graph database node set and the generated node relationship set; Among them, V is the graph database node set, and E is the node relationship set; Step 3: For each item in the session to be recommended, an m-element rating array with an initial value of 0 is established, wherein each element in the rating array corresponds to a rating value of the item to be recommended; and the corresponding element in the rating array is updated through the preference propagation of the items in the session to be recommended; Wherein, m is the number of items in the original data set; Step 4: Obtain the total score of each item to be recommended in the session to be recommended according to the score array obtained in step 3, and select multiple items with high total scores as recommended items for the session to be recommended; In step 3, the scoring value of the project is obtained, including: For a session S{v1,v2…v n }Project v i , search for item v in the graph database i 1-hop, 2-hop, ... N-hop set, and update the corresponding score value to the project v according to the global preference propagation strategy i The corresponding rating array s i middle; Among them, v i N-hop collection The jth item in The rating is: In the formula, yes and The weight of the node relationship, s (N-1) yes 's rating value.

2. The method for recommending commodity conversation sequences based on a global directed graph according to claim 1, characterized in that: In step 4, according to the session S{v1,v2…v n } Rating array s for each item i , get the global score array of session S: In the formula, k i Indicates project v i Parameter weights of Among them, each element in the global score array s corresponds to an item v i The total rating value of .

3. The method for recommending commodity conversation sequences based on a global directed graph according to claim 2, characterized in that: k i The setting rule is k n >k n-1 >…k1.

4. The method for recommending commodity conversation sequences based on a global directed graph according to claim 2 or 3, characterized in that: In step 3, a cypher statement is used to search for item v in the graph database. i A collection of 1-hop, 2-hop...N-hop.

5. The method for recommending commodity conversation sequences based on a global directed graph according to claim 4, characterized in that: In the step 2, the graph database node set and the node relationship set are imported into the database through the neo4j-import api of Neo4j to obtain a global directed graph of the session sequence.

6. The method for recommending commodity conversation sequences based on a global directed graph according to claim 5, characterized in that: The elements in the global scoring array s are arranged in descending order according to the total score values, and the items corresponding to the first t elements in the global scoring array s are screened out as the recommended items for the session.

7. The method for recommending commodity conversation sequences based on a global directed graph according to claim 6, characterized in that: t＝20。

Images