Content recommendation method and device, electronic equipment and storage medium

Abstract

The embodiment of the invention discloses a content recommendation method and device, electronic equipment and a storage medium. According to the embodiment of the invention, a historical behavior sequence of a target object is obtained, the historical behavior sequence comprises behavior items arranged according to a time sequence, and the behavior items represent interaction behaviors of the target object for contents; determining the similarity between two behavior items at a fixed interval in the historical behavior sequence; according to the similarity, an interest generic degree curve is constructed, and the interest generic degree curve reflects the change of the interest generic degree along with time; and determining recommended content adapted to the target object based on the interest generic degree curve. According to the embodiment of the invention, the historical behavior sequence of the target object is mined, and the interest extensive degree curve is proposed to effectively represent the change of the interest extensive degree of the object along with time, so that the recommendation content range is widened while the recommendation content adaptive to the object is accurately predicted, and the condition that the recommendation content is too single is avoided. Therefore, the scheme can improve the quality of content recommendation.

Description

Technical Field

[0001] This application relates to the field of computers, and particularly to a content recommendation method, apparatus, electronic device, and storage medium. Background Art

[0002] Content personalized recommendation is a technology that provides content highly matching the needs of an object by analyzing the object's information. Its core goal is to screen out the content that the object is most likely to be interested in from a large amount of content, thereby enhancing the object experience. However, the current content personalized recommendation methods overly rely on the object's historical records for recommending similar content, resulting in convergent recommended content and forming an information cocoon.

[0003] Therefore, there is an urgent need for a personalized content recommendation method that can both meet the object's interests and avoid the defect of convergent recommended content to improve the content recommendation effect. Summary of the Invention

[0004] Embodiments of this application provide a content recommendation method, apparatus, electronic device, and storage medium, which can improve the content recommendation effect.

[0005] Embodiments of this application provide a content recommendation method, including:

[0006] Obtain the historical behavior sequence of the target object, where the historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content;

[0007] Determine the similarity between two behavior items separated by a fixed interval in the historical behavior sequence;

[0008] According to the similarity, construct an interest breadth curve, where the interest breadth curve reflects the change of interest breadth over time, and the interest breadth represents the degree of the target object's interest breadth;

[0009] Based on the interest breadth curve, determine the recommended content suitable for the target object.

[0010] Embodiments of this application also provide a content recommendation apparatus, including:

[0011] A history unit for obtaining the historical behavior sequence of the target object, where the historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content;

[0012] A similarity unit for determining the similarity between two behavior items separated by a fixed interval in the historical behavior sequence;

[0013] A construction unit for constructing an interest breadth curve according to the similarity, where the interest breadth curve reflects the change of interest breadth over time, and the interest breadth represents the degree of the target object's interest breadth;

[0014] A recommendation unit, configured to determine recommended content suitable for a target object based on an interest popularity curve.

[0015] In some embodiments, the similarity unit includes:

[0016] A sequence sub-unit, configured to extract a first subsequence and a second subsequence from a historical behavior sequence, where the first subsequence includes the i-th to the j-th behavior items in the historical behavior sequence, and the second subsequence includes the (i + k)-th to the (j + k)-th behavior items in the historical behavior sequence, i and j are both positive integers, and k is a fixed interval;

[0017] A similarity sub-unit, configured to determine the similarity between corresponding item pairs in the first subsequence and the second subsequence, where a corresponding item pair is two behavior items with corresponding positions in the first subsequence and the second subsequence.

[0018] In some embodiments, the similarity sub-unit is configured to:

[0019] The corresponding item pair includes a first corresponding item in the first subsequence and a second corresponding item in the second subsequence. Determining the similarity between the corresponding item pairs in the first subsequence and the second subsequence includes:

[0020] Determining a first set, where the first set includes the first corresponding item in the first subsequence and its adjacent behavior items;

[0021] Determining a second set, where the second set includes the second corresponding item in the second subsequence and its adjacent behavior items;

[0022] Obtaining the similarity between the first corresponding item and the second corresponding item according to the mean value of the first set and the mean value of the second set.

[0023] In some embodiments, the sequence sub-unit is configured to:

[0024] Obtain the behavior activity of the target object;

[0025] Determine the window size of a sliding window based on the behavior activity;

[0026] Place the sliding window in the historical behavior sequence to obtain a first subsequence, where the first subsequence includes the i-th to the j-th behavior items in the historical behavior sequence, and the difference between i and j is one window size;

[0027] Move the sliding window in the historical behavior sequence with a step size of k to obtain a second subsequence.

[0028] In some embodiments, the recommendation unit includes:

[0029] A to-be-detected sub-unit, configured to obtain the content to be detected in the candidate set;

[0030] A scoring sub-unit, configured to use a prediction model to predict a recommended score of a target object for the content to be detected based on an interest popularity curve;

[0031] A sorting sub-unit, configured to select recommended content from a candidate set by sorting the recommended scores of all the content to be detected in the candidate set.

[0032] In some embodiments, the behavior items include content features and behavior features, the content to be detected includes domain labels, and the scoring sub-unit includes:

[0033] A splicing sub-module, configured to splice the interest popularity curve, the content features and behavior features of each behavior item in the historical behavior sequence to obtain the personal features of the target object;

[0034] A behavior metric sub-module, configured to predict the behavior metric of the target object for the content to be detected based on the personal features and the domain labels of the content to be detected;

[0035] A scoring sub-module, configured to determine the recommended score of the content to be detected based on the behavior metric.

[0036] In some embodiments, the prediction model includes a click type prediction network and a reading duration prediction network, and the behavior metric sub-module is configured to:

[0037] Use the click type prediction network to predict the click type of the target object for the content to be detected based on the personal features and the domain labels of the content to be detected;

[0038] Use the reading duration prediction network to predict the reading duration of the target object for the content to be detected based on the personal features and the domain labels of the content to be detected;

[0039] Calculate the behavior metric according to the click type and the reading duration.

[0040] In some embodiments, the recommendation unit further includes:

[0041] An object information sub-unit, configured to obtain the object information of the target object;

[0042] A rough sorting sub-unit, configured to screen multiple candidate contents from a content pool based on the object information to obtain a candidate set;

[0043] The recommendation unit further includes:

[0044] A re-sorting sub-unit, configured to re-sort all the recommended content and send the re-sorted recommended content to the target object.

[0045] In some embodiments, the re-sorting sub-unit is configured to:

[0046] Determine the interest breadth type of the target object based on the interest breadth curve;

[0047] If the target object is of the first breadth type, increase the sorting priority weight of the recommended content that conforms to the first breadth type;

[0048] If the target object is of the second breadth type, increase the sorting priority weight of the recommended content that conforms to the second breadth type;

[0049] Re - sort all the recommended content according to the sorting priority weight of the recommended content.

[0050] In some embodiments, the recommendation unit further includes:

[0051] The non - interaction subunit is used to determine the domain label that the target object has never had an interaction behavior with according to the historical behavior sequence of the target object, and the domain label represents the content domain to which the content belongs;

[0052] The type subunit is used to determine the interest breadth type of the target object in different time periods based on the interest breadth curve;

[0053] The selection subunit is used to select unexplored content from the content pool if the target object is of the first breadth type in the target time period, and the unexplored content is the content with the domain label that the target object has never had an interaction behavior with;

[0054] The simultaneous sending subunit is used to simultaneously send the unexplored content when sending the recommended content to the target object in the target time period.

[0055] In some embodiments, the simultaneous sending subunit is used for:

[0056] Determine the mixing ratio weight according to the interest breadth curve, and the mixing ratio weight is used to adjust the ratio of the recommended content to the unexplored content when sending;

[0057] According to the mixing ratio weight, simultaneously send the unexplored content when sending the recommended content to the target object.

[0058] The embodiment of the present application also provides an electronic device, including a memory storing multiple instructions; the processor loads the instructions from the memory to execute the steps in any one of the content recommendation methods provided by the embodiment of the present application.

[0059] The embodiment of the present application also provides a computer - readable storage medium, the computer - readable storage medium stores multiple instructions, and the instructions are suitable for being loaded by a processor to execute the steps in any one of the content recommendation methods provided by the embodiment of the present application.

[0060] Embodiments of the present application can obtain the historical behavior sequence of a target object. The historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to content. Determine the similarity between two behavior items separated by a fixed interval in the historical behavior sequence. According to the similarity, construct an interest breadth curve, which reflects the change of interest breadth over time, and the interest breadth represents the degree of breadth of the target object's interests. Based on the interest breadth curve, determine the recommended content suitable for the target object.

[0061] By deeply mining the historical behavior sequence of an object, embodiments of the present application can comprehensively capture the behavior patterns of the object, including the content it is interested in, interaction behavior habits, etc., and obtain and characterize the interest breadth curve of the object. This interest breadth curve can vividly show the change of the degree of breadth of the object's interests over time in a way similar to a heartbeat curve, helping to distinguish whether the object has broad or concentrated interests, and using this interest breadth curve to more accurately match the recommended content for the object. For example, for an object with broad interests, the recommended content is more diverse, avoiding the problem of information cocoons and enhancing the exploration and freshness of the object; while for an object with concentrated interests, the recommended content is more accurate, improving the interest matching degree and object satisfaction. Thus, embodiments of the present application can broaden the scope of recommended content, not only meeting the interest matching of the object but also avoiding the defect of convergent recommended content, effectively improving the quality of content recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0062] To more clearly illustrate the technical solutions in the embodiments of the present application, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the drawings in the following description are only some embodiments of the present application. For those skilled in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0063] Figure 1a is a schematic diagram of the scenario of the content recommendation method provided by the embodiments of the present application;

[0064] Figure 1b is a schematic flowchart of the content recommendation method provided by the embodiments of the present application;

[0065] Figure 1c is a schematic diagram of the similarity between behavior items of the content recommendation method provided by the embodiments of the present application;

[0066] Figure 2a is a schematic diagram of the content recommendation system of the content recommendation method provided by the embodiments of the present application;

[0067] Figure 2b is a schematic diagram of the interest breadth curves of Object 1 and Object 2 in the news scenario when the content recommendation method provided by the embodiments of the present application is applied;

[0068] Figure 3a It is a schematic structural diagram of a content recommendation device provided by an embodiment of the present application;

[0069] Figure 3b It is a schematic structural diagram of a content recommendation device provided by an embodiment of the present application;

[0070] Figure 3c It is a schematic structural diagram of a content recommendation device provided by an embodiment of the present application;

[0071] Figure 3d It is a schematic structural diagram of a content recommendation device provided by an embodiment of the present application;

[0072] Figure 4 It is a schematic structural diagram of an electronic device provided by an embodiment of the present application. Specific embodiments

[0073] Next, the technical solutions in the embodiments of the present application will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts shall fall within the protection scope of the present application.

[0074] The embodiments of the present application provide a content recommendation method, device, electronic device, and storage medium.

[0075] Among them, the content recommendation device can be specifically integrated in an electronic device, and the electronic device can be a device such as a terminal or a server. Among them, the terminal can be a device such as a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, or a personal computer (PC); the server can be a single server or a server cluster composed of multiple servers.

[0076] In some embodiments, the content recommendation device can also be integrated in multiple electronic devices. For example, the content recommendation device can be integrated in multiple servers, and the content recommendation method of the present application can be implemented by multiple servers.

[0077] In some embodiments, the terminal can also be used as a server to implement some or all of the functions of the server.

[0078] For example, refer to Figure 1a, the electronic device can be a server, which can obtain the historical behavior sequence of the target object from the object information library. The historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content. The server can determine the similarity between two behavior items separated by a fixed interval in the historical behavior sequence. Based on the similarity, an interest breadth curve is constructed. The interest breadth curve reflects the change of interest breadth over time, and the interest breadth represents the extent of the target object's broad interests. Based on the interest breadth curve, recommended content suitable for the target object is determined in the content pool, and finally the recommended content is sent to the mobile terminal of the target object.

[0079] For example, in some embodiments, a news recommendation server can generate personalized recommendations for mobile terminal objects. The news recommendation server can retrieve the object's most recent 40 news click records from the object information library and form a historical behavior sequence [behavior item 1, behavior item 2,..., behavior item 40] in chronological order. Each behavior item can include interaction behavior characteristics such as the article ID, category label, and reading duration of the news. Then, the similarity between two adjacent behavior items in the historical behavior sequence is calculated, and the 39 obtained similarity values are connected in chronological order to form an interest breadth curve.

[0080] If the interest breadth curve of object A is a high-frequency fluctuating interest breadth curve, it indicates that object A has broader interests. Therefore, when the news recommendation server pushes news to object A, 30% of the information breaking through the information cocoon of non-historical preference categories can be inserted.

[0081] If the interest breadth curve of object B is a low-frequency fluctuating interest breadth curve, it indicates that object B's interests are more concentrated. Therefore, when the news recommendation server pushes news to object B, in-depth mining can be performed on the fields that object B is interested in, and long-form content can be appended based on the fields with a high proportion of behaviors such as "collection" and "like" in the behavior items to enhance the object's stickiness.

[0082] The following will be described in detail respectively. It should be noted that the serial numbers of the following embodiments do not limit the preferred order of the embodiments.

[0083] In this embodiment, a content recommendation method is provided. As Figure 1b shown, the specific process of this content recommendation method can be as follows:

[0084] 110. Obtain the historical behavior sequence of the target object. The historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content.

[0085] Among them, the historical behavior sequence refers to the set of behavior records of the target object interacting with the content within a period of time, arranged in chronological order.

[0086] Among them, the information carriers whose content is interacted with can include articles, music, videos, commodities, advertisements, images, etc.

[0087] Among them, the behavior item is the smallest recording unit of a single content interaction, which can record the behavior of an object towards a certain content and can include content features and behavior features. The content features describe the attributes of the content itself, and the behavior features describe the interaction method between the object and the content.

[0088] The content features are used to depict the inherent attributes of the interacted content, helping to understand the classification, theme or value of the content. In some embodiments, the content features may include:

[0089] Category: such as article type, first-level category, second-level category, etc.;

[0090] Theme: such as content ID, content keywords, tags, etc.;

[0091] Source: such as article author, publishing medium, etc.;

[0092] Publication time: such as the freshness of the content, etc.;

[0093] Popularity: such as the click volume, repost volume, comment volume of the content, etc.

[0094] The behavior features are used to depict the behavior patterns of the interaction between the object and the content, reflecting the interest intensity, preference and intention of the object. In some embodiments, the behavior features may include:

[0095] Object identity: such as object account, object identity information, etc.;

[0096] Interaction type: such as click, like, favorite, repost, comment, etc.;

[0097] Interaction duration: such as browsing duration, page sliding speed, number of repeated visits, etc.;

[0098] Interaction depth: such as whether to browse completely, whether to view the comment area, whether to click on relevant recommendations, etc.

[0099] In some embodiments, the historical behavior sequence of the target object can be obtained in various ways. For example, the behavior items of the target object can be collected from different data sources and sorted into a historical behavior sequence according to the time order. Among them, the data sources can include the client of the target object, server logs, object information databases, third-party data, and so on.

[0100] For example, in some embodiments, the client can record the interaction behavior between the object and the content, while the backend service can record the log data of the object requests, which includes the interaction behavior between the object and the content. After obtaining the initial data, the data can be cleaned, such as denoising, complementing, de-duplicating, etc.

[0101] 120. Determine the similarity between two behavioral items in a historical behavior sequence that are separated by a fixed interval.

[0102] Among them, there are multiple possibilities for the calculation method of similarity, not a single fixed pattern. For example, the similarity between adjacent behavioral items in the historical behavior sequence can be calculated to capture the interest drift of the target object. For example, the similarity between each behavioral item and the global mean of all behavioral items in the historical behavior sequence can be calculated to establish a baseline for the long-term interest of the object and filter out the interference of accidental behaviors; for example, the aggregated value of the behavioral item similarity can be calculated within a sliding time window to balance the contradictory relationship between short-term behavior fluctuations and long-term stable preferences.

[0103] The fixed interval can include a position interval, a time interval, etc. The position interval means that there is only a fixed position interval in the sequence between behavioral items, such as a pure sequential relationship separated by 1 node, 2 nodes, etc.; the time interval means a fixed time interval between behavioral items, such as occurring once every 5 minutes. Among them, the value of the fixed interval has a significant impact on the interest generalization modeling. In some embodiments, the value of the fixed interval can be adjusted based on the application scenario. For example, a small fixed interval value of 1 to 3 is suitable for capturing immediate interest drift; a medium fixed interval value of 4 to 10 is suitable for identifying interest migration patterns; a large fixed interval value greater than 10 is suitable for reflecting long-term interest evolution.

[0104] The similarity between behavioral items is the similarity between the multi-dimensional features included in the behavioral items, and the multi-dimensional features can include content features, behavioral features, etc. For example, the multi-dimensional features of a behavioral item can be expressed as [click, reading duration, article ID, category, first-level category, second-level category, etc.], and the similarity between two behavioral items is obtained by calculating the similarity of the multi-dimensional features.

[0105] There are various calculation methods for similarity. For example, cosine similarity, Euclidean distance, Manhattan distance, Jaccard coefficient, Dice coefficient, edit distance, dynamic time warping distance, KL divergence, Pearson correlation coefficient, deep learning-based semantic similarity, similarity calculation combined with time decay, and combinations of the above multiple methods, etc., can be calculated.

[0106] For example, in some embodiments, for the historical behavior sequence [behavior item 1, behavior item 2,..., behavior item 40], assuming a fixed interval of 1, the cosine similarity between adjacent behavior items can be calculated to obtain multiple similarities: [cos(behavior item 1, behavior item 2), cos(behavior item 2, behavior item 3), cos(behavior item 3, behavior item 4), …, cos(behavior item 39, behavior item 40)]. The interest generality is obtained by subtracting the cosine similarity from 1. Finally, these interest generalities are connected into a curve to obtain the interest generality curve [1 - cos(behavior item 1, behavior item 2), 1 - cos(behavior item 2, behavior item 3), 1 - cos(behavior item 3, behavior item 4), …, 1 - cos(behavior item 39, behavior item 40)].

[0107] In some embodiments, two sequences can be intercepted from the historical behavior sequence through a sliding window. Therefore, determining the similarity between two behavior items separated by a fixed interval in the historical behavior sequence may include:

[0108] Extracting a first subsequence and a second subsequence from the historical behavior sequence, where the first subsequence includes the i-th to the j-th behavior items in the historical behavior sequence, and the second subsequence includes the (i + k)-th to the (j + k)-th behavior items in the historical behavior sequence, both i and j are positive integers, and k is the fixed interval;

[0109] Determining the similarity between the corresponding item pairs in the first subsequence and the second subsequence, where the corresponding item pair is two behavior items with corresponding positions in the first subsequence and the second subsequence.

[0110] For example, referring to Figure 1c , the first behavior item in the historical behavior sequence is the behavior item most recently generated by the target object, the fixed interval k is 1, i = 1, j = 39, then the first subsequence S 1 includes the 1st to the 39th behavior items in the historical behavior sequence, that is, S 1 = [x 1 1 , x 1 2 , x 1 3 , …, x 1 39 , and the second subsequence S 2 includes the 2nd to the 40th behavior items in the historical behavior sequence, that is, S 2 = [x 2 2 , x 2 3 , x 2 4 , …, x 2 40 .

[0111] Among them, the corresponding item pair includes the first corresponding item in the first subsequence and the second corresponding item in the second subsequence corresponding to the first subsequence. Then S 1 's x 1 1 and S 2 's x 2 2 are a corresponding item pair, the first corresponding item is x 1 1 , and the second corresponding item is x 2 2 ; S 1 's x 1 2 and S 2 's x 2 3 are a corresponding item pair, the first corresponding item is x 1 2 , and the second corresponding item is x 2 3 , and so on. The similarity is calculated pairwise between corresponding items.

[0112] In some embodiments, when calculating the similarity, context information can be considered to balance the contradictory relationship between short-term behavior fluctuations and long-term stable preferences. Therefore, determining the similarity between corresponding item pairs in the first subsequence and the second subsequence includes:

[0113] Determine the first set, where the first set includes the first corresponding item in the first subsequence and its adjacent behavior items, that is, the first set is the context set of the first corresponding item;

[0114] Determine the second set, where the second set includes the second corresponding item in the second subsequence and its adjacent behavior items, that is, the second set is the context set of the second corresponding item;

[0115] Obtain the similarity between the first corresponding item and the second corresponding item according to the mean of the first set and the mean of the second set.

[0116] For example, there are accidental actions in user behavior. In this embodiment, by aggregating adjacent items, the single-point behavior is extended to a local context window to reduce the influence of short-term noise, so as to smooth the interest fluctuation caused by such accidental operations and capture more stable information. For example, in a recommendation system, a user may suddenly click on a content that they don't usually view one day, but in the long run, their interests may be more stable, so it is necessary to combine the context to reduce such noise.

[0117] Among them, calculating the mean of the set can be to average the values of these items, or other aggregation methods, such as taking the average vector. The mean is used to synthesize the information of multiple points, reduce the influence of random fluctuations of a single point, and thus more stably reflect the trend at this position.

[0118] In some embodiments, to improve the accuracy of analysis and avoid excessive noise caused by too small a window when the data is sparse or losing details due to too large a window when the data is dense, the sequence lengths of the first subsequence and the second subsequence obtained can be dynamically adjusted by analyzing the activity of the object in real time. Therefore, extracting the first subsequence and the second subsequence from the historical behavior sequence may include:

[0119] Obtain the behavior activity of the target object;

[0120] Based on the behavior activity, determine the window size of the sliding window;

[0121] Place the sliding window in the historical behavior sequence to obtain the first subsequence, where the first subsequence includes the i-th to the j-th behavior items in the historical behavior sequence, and the difference between i and j is one window size;

[0122] Move the sliding window in the historical behavior sequence with a step size of k to obtain the second subsequence.

[0123] Among them, the behavior activity refers to the activity frequency or intensity degree of the object.

[0124] In some embodiments, a smaller window can be used when the object's behavior is dense, and a larger window can be used when it is sparse. For example, for a highly active object, the window can be reduced to the hourly level to capture the subtle interest drift in the dense behavior; for a low-active object, the window can be extended to the weekly / monthly level to integrate the discrete behavior of the object.

[0125] Among them, there are various ways to obtain the behavior activity of the target object. For example, it can be obtained through analysis and statistics based on the daily login times, usage duration, interaction behaviors, etc. of the target object within a period of time.

[0126] For example, core events such as object login, click, interaction, consumption, etc. can be collected, and interaction types such as object ID, behavior type, timestamp, operation duration, page path, etc. can be recorded. Outliers in the data can be cleaned, and missing values can be filled, etc.; indicators such as login times, behavior frequency, behavior diversity, etc. can be calculated based on these data; finally, various analysis methods such as weighted summation, exponential decay calculation, etc. can be performed on these indicators to finally obtain the activity score of the object.

[0127] 130. Construct an interest breadth curve according to similarity. The interest breadth curve reflects the change of interest breadth over time, and the interest breadth characterizes the degree of interest breadth of the target object.

[0128] The interest breadth curve proposed in the embodiments of the present application describes the change of the degree of interest breadth over time, providing a decision-making basis for personalized services.

[0129] After obtaining the similarity through step 120, the similarity needs to be further processed to obtain the interest generality, so that the interest generality can accurately reflect the interest change of the object. For example, multiple processes such as normalization, linear transformation, exponential transformation, logarithmic transformation, etc. can be performed on the similarity to obtain the interest generality.

[0130] For example:

[0131] Interest generality = (1 - similarity) k

[0132] Among them, when k > 1, the interest generality is more sensitive to low similarity and changes more steeply; when 0 < k < 1, the change of the interest generality is more gentle.

[0133] For example:

[0134]

[0135] Among them, the decay rate is controlled by adjusting k. The larger k is, the faster the interest generality in the high similarity region drops; after normalization, the interest generality is 1 when the similarity = 0 and 0 when the similarity = 1.

[0136] For example:

[0137]

[0138] Among them, the larger k is, the more significant the increase in the interest generality in the low similarity region; when the similarity approaches 1, the interest generality approaches 0.

[0139] For example:

[0140] If the similarity ≤ t, then

[0141] If the similarity > t, then

[0142] Among them, when t = 0.5, the interest generality drops faster in the low similarity region.

[0143] For example:

[0144]

[0145] Among them, the threshold for the sudden drop of the interest generality is set by m, and the steepness of the curve is controlled by k. For example, when m = 0.5, the interest generality drops rapidly near the similarity = 0.5.

[0146] For example:

[0147] Interest generality = 1 - similarity 2

[0148] Among them, when the similarity is relatively high, the interest breadth decreases faster, highlighting the low interest breadth of highly similar behaviors.

[0149] For example:

[0150]

[0151] Among them, the interest breadth changes relatively gently in the low similarity region and decreases faster in the high similarity region.

[0152] For example:

[0153] Interest breadth = -similarity · log(similarity) - (1 - similarity) · log(1 - similarity)

[0154] Among them, when the similarity is close to 0 or 1, the interest breadth is relatively low; it is the highest in the middle region (similarity ≈ 0.5), reflecting the maximization of uncertainty.

[0155] Among them, only when the similarity is lower than the threshold θ, the interest breadth is considered to be 1, otherwise it is 0.

[0156] In summary, in this embodiment, the interest breadth needs to decrease as the similarity increases to ensure that low similarity corresponds to high interest breadth. In some embodiments, the interest breadth needs to be normalized so that it falls within a reasonable range (such as [0, 1]) for subsequent analysis.

[0157] For example, the interest breadth can be set to 1 - similarity. For example, the interest breadth between the 2nd and 3rd behavior items in the sequence can be set to 1 - cos(behavior item 2, behavior item 3).

[0158] 140. Based on the interest breadth curve, determine the recommended content suitable for the target object.

[0159] In the embodiment of the present application, by analyzing the interest breadth curve of the target object, the changing trend of the breadth of its interests over time is understood, so as to infer the interest preferences of the target object at different time periods, and based on this, recommend content that is more in line with what the target object is currently interested in.

[0160] For example, if the curve is relatively high, it indicates that the target object has a wide range of interests and may be interested in various types of content. The target object may be more willing to accept diverse content, so various types of content can be recommended. If the curve is relatively low, it means that the target object's interests are more concentrated and may be more interested in specific types of content. The target object may be more inclined to specific types of content, so content highly relevant to its current interests can be recommended. If the curve shows an upward or downward trend, it indicates that the target object's interest range is expanding or shrinking. For example, if the interest breadth curve shows that the target object's interest range is expanding, some content that is related to its existing interests but slightly different can be recommended to explore its potential interests.

[0161] Based on the interest breadth curve, there are various schemes for determining the recommended content suitable for the target object. For example, machine learning models, clustering, and classification algorithms can be used to determine the recommended content.

[0162] For example, a content scoring model based on machine learning can be adopted to score different content according to the interest breadth curve, and the content with high scores can be pushed to the object as recommended content. For example, a behavior prediction model based on machine learning can be used to predict the fields that the target object will be interested in within a certain period in the future according to the interest breadth curve, and the content in this field can be pushed to the object as recommended content.

[0163] For example, the target object can be classified according to the interest breadth curve, and according to the classification type, the content that meets the classification type can be screened from the content pool and pushed to the object as recommended content.

[0164] For example, the target object can be classified according to the interest breadth curve. When making regular pushes to the object, according to the classification type, different proportions of content in the fields that the target object has not explored can be incorporated into the regular push stream, and so on.

[0165] For example, the timeliness and diversity of the recommended content can be dynamically adjusted by combining the fluctuations of the interest breadth curve and the time decay factor. For example, the interest breadth curve can be divided into different time windows to analyze the stability of the object's interest breadth. If the object's interest breadth curve fluctuates violently, it indicates that the object's short-term interests change rapidly, and content with strong timeliness, such as hot news and short-term trend content, can be recommended. If the curve fluctuates gently, it indicates that the object's long-term interests are stable, and in-depth content, such as feature articles and series of tutorials, can be recommended.

[0166] For example, according to the characteristics of the interest breadth curve, the weights of collaborative filtering and content filtering can be dynamically switched. For instance, when the peak value of the interest breadth curve is greater than the preset peak value, the weight of content filtering is strengthened to recommend diverse content; when the trough value of the interest breadth curve is less than the preset trough value, the weight of collaborative filtering is strengthened to recommend content based on the preferences of similar objects. And the rate of interest change is judged by the curve slope to quickly switch the strategy.

[0167] For example, an object interest graph can be constructed using the interest breadth curve, and cross-domain content can be recommended through graph diffusion. For example, the historical behavior sequence of the object is mapped to interest nodes, and an interest graph is constructed according to behavior similarity; when the interest breadth curve rises, content in adjacent interest fields is recommended by expanding from the graph; when the curve drops, only the content of the core nodes of the graph is focused on without excessive expansion.

[0168] For example, when the end of the object interest breadth curve is rising, the probability of exploring new domain content is increased; when the end of the curve is dropping, exploration is reduced and content of verified interests is recommended preferentially.

[0169] Therefore, in some embodiments, a preliminary candidate set can be screened out from the massive content, and a prediction model is used to predict the score of the object for the candidate content in combination with the interest breadth curve of the object. After sorting by the score, high-score content is selected and pushed to the object. Therefore, determining the recommended content suitable for the target object based on the interest breadth curve can include:

[0170] Obtain the content to be detected in the candidate set;

[0171] Adopt a prediction model to predict the recommended score of the target object for the content to be detected based on the interest breadth curve;

[0172] Select the recommended content from the candidate set by sorting the recommended scores of all the content to be detected in the candidate set.

[0173] Among them, in this embodiment, a preliminary candidate set is screened out from the massive content through coarse-grained filtering, reducing the complexity of subsequent calculations and ensuring the basic relevance of the candidate content. In some embodiments, for new objects lacking behavior data, the candidate set can rely on default strategies such as popular content and regional preferences to provide basic recommendations.

[0174] Among them, the prediction model can be a deep learning model, such as a time series model, a multi-task model, etc. The time series model can include Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Transformer, etc., which can directly model the time series changes of the interest popularity curve and predict future interests. The multi-task model can embed the behavior data and the content to be detected of the object respectively and calculate the matching score, and can include Deep Structured Semantic Model (DSSM), Sentence-BERT (SBERT), User-Item Twin Towers multi-task recommendation model, etc.

[0175] Among them, the method of sorting and selecting recommended content in the candidate set can include directly sorting in descending order of the prediction score and selecting the top top-K content, etc.

[0176] The following is a detailed introduction:

[0177] In some embodiments, the behavior item can include content features and behavior features, the content to be detected can include domain labels, and a prediction model is used to predict the recommendation score of the target object for the content to be detected based on the interest popularity curve, which can include:

[0178] Concatenate the interest popularity curve, the content features and behavior features of each behavior item in the historical behavior sequence to obtain the personal features of the target object;

[0179] Based on the personal features and the domain labels of the content to be detected, predict the behavior metrics of the target object for the content to be detected;

[0180] Based on the behavior metrics, determine the recommendation score of the content to be detected.

[0181] In the embodiment of the present application, the personal features of the target object are used to predict whether the target object is interested in the domain labels of the content to be detected, that is, the behavior metrics. The behavior metrics refer to the quantitative evaluation values of the specific interaction behaviors that the target object may generate for the content to be detected predicted by the model. It is the core intermediate variable connecting the object features and the final recommendation score, and can include interaction probability, interaction duration, interest matching degree, negative feedback risk, etc.

[0182] For example, the interaction probability can include click-through rate (CTR), play rate, purchase rate, download rate, etc.; the interaction duration can include stay duration, number of re-watches, interaction comment probability, sharing willingness, etc.; the negative feedback risk can include skip rate, negative review rate, report probability, etc.; the interest matching degree can include the degree of fit between the content and the object's interests, etc.

[0183] Among them, content features are features that describe the attributes of the content itself, which can include basic attribute features, semantic features, timeliness features, and so on. The basic attribute features of the content can include the unique identifier (ID), type, classification label, keyword, theme, etc. of the content; the semantic features of the content can include title embedding, global summary features, sentiment tendency, etc.; the timeliness features of the content can include the release timestamp, relevance to hot events, timeliness decay coefficient, etc.

[0184] For example, the content features can be (primary category cat1 = technology, secondary category cat2 = artificial intelligence, keywords = ["GPT-4", "large model"], theme = [0.7, 0.2, 0.1], sentiment tendency = 0.8).

[0185] Among them, behavior features are dynamic behavior features generated during the interaction between the object and the content, which can include explicit behavior features, implicit behavior features, sequence features, and so on. Explicit behavior features can include whether there is interaction, interaction duration, interaction type, interaction speed, etc.; implicit behavior features can include attention heat map, interaction completion rate, interaction time distribution, etc.; sequence features can include interest transfer matrix, behavior cycle pattern, time interval of consecutive behaviors, etc.

[0186] For example, the behavior features can be (whether clicked = 1, reading duration = 45s, time distribution = [09:30, 12:15, 20:40]).

[0187] Among them, the domain label describes the domain to which the content belongs, which can include domain type, primary category, secondary category, tertiary category, social propagation coefficient, etc.

[0188] In some embodiments, the recommendation score can be obtained by weighted summation of behavior metrics, where the weight distribution of the interaction probability and the interaction duration is dynamically adjusted according to the business objective. For example, in scenarios where the interaction probability is emphasized, such as in scenarios of information flow advertising, e-commerce promotions, news headlines, application download pages, etc., the weight of the interaction probability can be increased; in scenarios where the interaction duration is regarded, such as in online education courses, long video platforms, music streaming media, reading platforms, the weight of the interaction duration can be increased.

[0189] For example, the prediction model can be a multi-task model. The input of the multi-task model is the interest generality curve, the content features and behavior features of each behavior item in the historical behavior sequence, and the domain label of the content to be detected. The output of one tower in the multi-task model is the interaction probability of the target object for the content to be detected, and the output of the other tower is the interaction duration of the target object for the content to be detected; finally, the interaction probability and the interaction duration are weighted and summed to obtain the recommendation score of the content to be detected.

[0190] Thus, in some embodiments, the prediction model may include a click type prediction network and a reading duration prediction network. Based on the personal characteristics and the domain label of the content to be detected, the behavior metrics of the target object for the content to be detected may include:

[0191] Using the click type prediction network, based on the personal characteristics and the domain label of the content to be detected, predict the click type of the target object for the content to be detected;

[0192] Using the reading duration prediction network, based on the personal characteristics and the domain label of the content to be detected, predict the reading duration of the target object for the content to be detected;

[0193] Calculate the behavior metrics according to the click type and the reading duration.

[0194] Among them, the click type prediction network is used to predict the click probability or click type (i.e., whether to click) of the object on the content, and may include a Deep Neural Network (DNN), a Deep Factorization Machine (DeepFM), a Transformer-Encoder network, a Gradient Boosting Decision Tree (GBDT), and so on.

[0195] Among them, the reading duration prediction network is used to predict the stay duration of the object on the content, and may include a Deep Regression Network, a Deep Survival Network, a Temporal Convolutional Network (TCN), a Multi-Task Learning network, and so on.

[0196] In this embodiment, by integrating the interest generality curve of the object, the content features and behavior features in the historical behavior sequence, a personal characteristic that comprehensively depicts the object's interest preference is constructed. Based on this, combined with the domain label of the content to be detected, the behavior metrics of the object for this content are accurately predicted, and finally the recommendation score is calculated. The effect is to improve the personalization degree and accuracy of the recommendation system, achieve more efficient content matching and improve the object satisfaction, thereby optimizing the business conversion rate and the object retention rate.

[0197] In the embodiments of the present application, the content recommendation method can obtain the object information of the target object, such as the interest breadth curve of the target object, the historical behavior sequence, etc., and determine the recommended content suitable for the target object from the content pool. In some embodiments, the content recommendation method may include three steps: rough sorting, scoring, and re-ranking. Among them, the scoring step is the step of predicting the recommendation score of the target object for the content to be detected mentioned in the above embodiments of the present application. Before the scoring step, a rough sorting step may be included:

[0198] Obtain the object information of the target object;

[0199] Based on the object information, screen multiple candidate contents in the content pool to obtain a candidate set.

[0200] Among them, the object information may also include various data such as the fields that the object is known to be interested in, the fields that the object is known not to be interested in, the fields that the object may be interested in, the fields that the object may not be interested in, etc. Rough sorting can quickly screen out a candidate set related to the target object from the massive content pool and provide high-quality input for the follow-up.

[0201] After determining the recommended content suitable for the target object, a re-ranking step may also be included:

[0202] Re-rank all the recommended contents and send the re-ranked recommended contents to the target object.

[0203] Re-ranking can perform refined sorting on the recommended contents and optimize the personalization and business value of the recommendation results. For example, in video recommendation, if the initially generated recommendation list consists of movies of the same type, the re-ranking stage can introduce a diversity strategy and insert movies of other types such as comedy and romance into the list; for example, in news recommendation, the re-ranking stage can give priority to displaying the latest news and move the outdated news down.

[0204] In some embodiments, the re-ranking stage can dynamically adjust the sorting priority of the recommended contents according to the interest breadth level of the target object, so as to more accurately meet the needs of the target object and improve the recommendation effect. Therefore, re-ranking all the recommended contents may include:

[0205] Based on the interest breadth curve, determine the interest breadth type of the target object;

[0206] If the target object is of the first breadth type, increase the sorting priority weight of the recommended contents that conform to the first breadth type;

[0207] If the target object is of the second breadth type, increase the sorting priority weight of the recommended contents that conform to the second breadth type;

[0208] Re-rank all the recommended contents according to the sorting priority weights of the recommended contents.

[0209] Among them, the interest breadth type of the target object can be determined according to the shape of the interest breadth curve. For example, it can be judged whether the object's interest is gradually dispersed or concentrated according to the overall slope of the curve; for example, it can be judged whether the object's interest is stable according to the variance of the curve; for example, it can be judged whether there are periodic or sudden interest changes according to the local shape similarity of the curve.

[0210] Among them, the interest breadth types can include the first breadth type, the second breadth type, and can also include the third breadth type, the fourth breadth type, and so on. For example, the interest breadth types can be divided into the first breadth type, the second breadth type, and the third breadth type according to high, medium, and low. Objects with a high breadth type have a wide range of interests and strong exploration ability. The recommended content needs to take into account diversity, exploration ability, and interest matching degree. Objects with a low breadth type have concentrated interests, and the recommended content needs to more accurately match their core interests and reduce unnecessary exploration.

[0211] For example, if the target object is of the high breadth type, the content in the un-explored fields of the target object can be ranked in the front part, and the content in the fields that the target object is already interested in can be ranked in the back part; if the target object is of the medium breadth type, the content in the un-explored fields of the target object and the content in the fields that the target object is already interested in can be evenly arranged; if the target object is of the low breadth type, the content in the un-explored fields of the target object can be ranked in the back part, and the content in the fields that the target object is already interested in can be ranked in the front part.

[0212] In some embodiments, the un-explored content can be recommended dynamically by combining the object's historical behavior sequence, interest breadth curve, and by time period. For example, when the object is in a time period with wide interests and strong exploration ability, more content in the un-explored fields can be recommended to it, such as articles in the "travel" or "art" categories that the object has never clicked on; when the object is in a time period with concentrated interests and weak exploration ability, less content in the un-explored fields can be recommended to it.

[0213] Therefore, after determining the recommended content suitable for the target object, it can also include:

[0214] According to the target object's historical behavior sequence, determine the domain labels of the fields in which the target object has never had an interaction behavior. The domain labels represent the content fields to which the content belongs;

[0215] Based on the interest breadth curve, determine the interest breadth types of the target object in different time periods;

[0216] If the target object is of the first breadth type in the target time period, select un-explored content from the content pool. The un-explored content is the content with the domain labels of the fields in which the target object has never had an interaction behavior;

[0217] When sending recommended content to a target object during a target time period, unexplored content is sent simultaneously.

[0218] The proportion of unexplored content sent can be dynamically adjusted according to the interest breadth curve. Therefore, in some embodiments, when sending recommended content to a target object during a target time period, simultaneously sending unexplored content may include:

[0219] Determine a mixing ratio weight according to the interest breadth curve, and the mixing ratio weight can be used to adjust the ratio of recommended content to unexplored content when sending.

[0220] According to the mixing ratio weight, when sending recommended content to the target object, send unexplored content simultaneously.

[0221] For example, through the interest breadth curve, the interest breadth types of the object at different time periods such as morning, noon, and evening are identified in real time, and the recommendation strategy is dynamically adjusted. For example, during the morning commute period, the target object is of a high breadth type, and this target object may be in an open state of information reception, and diverse content such as international news and unexplored scientific and technological trivia can be recommended; during the lunch break, the target object is of a medium breadth type, and moderately diverse content such as light entertainment, food, and short readings can be recommended; during the evening rest period, the target object is of a low breadth type, and precise content such as the target object's frequently watched movies and TV shows and in-depth long articles can be recommended.

[0222] For example, through the interest breadth curve, the interest breadth types of the object at different time periods on weekdays and non-working days are identified in real time, and the recommendation strategy is dynamically adjusted. For example, from Monday to Friday, the target object is in a working state, with fragmented time, and tends to efficiently obtain core interest content related to work and study. Therefore, it is mainly based on precise matching, and content in the target object's historical high-frequency interest fields is recommended for it, restricting the proportion of content in unexplored fields; for example, on Saturday and Sunday, the target object has ample time, is more open to interests, and tends to explore content in the fields of entertainment, leisure, and niche areas. Cross-domain content is recommended for it, and the proportion of content in unexplored fields is greatly increased.

[0223] As can be seen from the above, the embodiments of the present application can obtain the historical behavior sequence of the target object. The historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content; determine the similarity between two behavior items separated by a fixed interval in the historical behavior sequence; according to the similarity, construct an interest breadth curve, and the interest breadth curve reflects the change of interest breadth over time, and the interest breadth represents the degree of the target object's interest breadth; based on the interest breadth curve, determine the recommended content suitable for the target object.

[0224] In the embodiments of the present application, by deeply mining the historical behavior sequence of an object, comprehensively capturing its behavior pattern, depicting an interest breadth curve similar to a heartbeat curve, and intuitively reflecting the change of the object's interest breadth over time, the objects with broad or concentrated interests can be accurately distinguished. Based on this, the system recommends diverse content for objects with broad interests, avoids information cocoons and improves exploration; recommends accurately matched content for objects with concentrated interests to improve satisfaction. This solution broadens the scope of recommended content, satisfies both interest matching and avoids content convergence, thereby improving the quality of content recommendation.

[0225] The method described in the above embodiments will be further described in detail below.

[0226] In this embodiment, taking news push as an example, the method of the embodiments of the present application will be described in detail.

[0227] As Figure 2a shown, the embodiments of the present application propose a content recommendation system, including an object information module, a content pool module, a rough sorting module, a scoring module, and a re-ranking module. Among them, the object information module integrates the static features and dynamic behavior features of the object; the content pool module integrates a large amount of recommendable content data.

[0228] The embodiments of the present application propose a deep information mining method based on the behavior sequence of the target object, which realizes the concrete representation of the dynamic change of the object's interest by constructing an interest breadth curve. This curve adopts a time series similarity calculation method to intuitively reflect the evolution law of the object's interest breadth in the form of an electrocardiogram signal pattern, effectively improving the intention recognition accuracy and business prediction effect of the recommendation system. When recommending for the target object, the predicted recommended content comprehensively considers the interest breadth and potential interest prediction, rather than the content highly similar to historical clicks, thus broadening the scope of recommended content.

[0229] The specific process is as follows:

[0230] (1) Rough sorting stage.

[0231] In the rough sorting stage, a candidate set can be roughly screened from the content pool according to the object information of the target object. Among them, the object information can include multi-dimensional data such as the social attributes, consumption habits, interest preferences, and behavior preferences of the target object.

[0232] Among them, the rough sorting stage can include the steps of:

[0233] Obtain the object information of the target object;

[0234] Based on the object information, screen multiple candidate contents in the content pool to obtain a candidate set;

[0235] Among them, the object information of the target object can come from the object information module, and can include the social attribute characteristics, historical behavior data, interest map, etc. of the target object.

[0236] Among them, in addition to including the content that exactly matches the interests of the target object, the candidate content also needs to include the content that the target object may be interested in but has not come into contact with yet. Therefore, after directly and precisely matching the classification labels of the content with the interest labels in the object information, it is also necessary to add the content that the target object may be interested in but has not come into contact with yet.

[0237] Among them, various methods can be used to find the content that the target object may be interested in but has not come into contact with yet (i.e., unexplored content) in the content library:

[0238] Exclude the content that the target object has come into contact with in the content library;

[0239] Screen in the content library for the content that the target object may be interested in but has not come into contact with yet.

[0240] Among them, the database can be used to record all the object information of the user, and the content that has been contacted is automatically excluded before recommendation.

[0241] Among them, according to the object information, relevant uncontacted content in the content library can be matched as unexplored content.

[0242] In some embodiments, similar objects of the target object can be found through user collaborative filtering, so that the content that the similar objects are interested in in the content library is recommended to the target object as unexplored content.

[0243] In some embodiments, a neural network model can be used to predict the content that the target object may be interested in but has not come into contact with yet in the content library.

[0244] (2) Scoring stage.

[0245] In the scoring stage, the historical behavior sequence of the target object can be obtained. The historical behavior sequence can include behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content; based on the similarity between the behavior items in the historical behavior sequence, an interest breadth curve is constructed, and the interest breadth curve reflects the change in the interest breadth of the target object over time; based on the interest breadth curve, the recommended content suitable for the target object is determined.

[0246] Specifically, the scoring stage can be divided into the following processes:

[0247] (1) Preprocessing of the historical behavior sequence.

[0248] Obtain the most recent n click sequences S' = [s 1 , s 2 , s3 …s n . Among them, s is the behavior item, including the content features and behavior features of object A. The content features include the unique ID identifier of the content read by the target object, the content type type, the first-level content category cat1, and the second-level content category cat2; the behavior features include the access time time of the content, the interaction duration duration, the behavior type click, etc.

[0249] Then, a series of preprocessing is performed on the sequence S' to obtain the historical behavior sequence S.

[0250] In some embodiments, data cleaning can be performed on the sequence S', such as deduplication processing, missing value processing, time sorting, etc.

[0251] Among them, the deduplication processing can retain the last operation among multiple identical operations within a short period of time, avoiding interference caused by repeated clicks on the same content within a short period of time.

[0252] Among them, the missing value processing can fill the missing fields with default values, such as marking the empty category as unknown, removing incomplete records, etc.

[0253] Among them, the time sorting can be arranged in ascending order according to the access time time field to ensure that the behavior sequence is strictly arranged in the order of time occurrence.

[0254] In some embodiments, feature engineering can be performed on the sequence S', such as performing feature engineering processing on the content features and behavior features respectively.

[0255] Among them, the content feature processing includes directly encoding the high-frequency content IDs and merging the low-frequency content IDs into <unk>Category; alternatively, the content features can also be mapped to low-dimensional dense vectors through an embedding layer.

[0256] The content feature processing also includes encoding categorical features (type / cat1 / cat2), such as using one-hot encoding or word embedding for encoding.

[0257] In some embodiments, if there is a hierarchical relationship, such as cat1→cat2, they can be directly concatenated into one dimension.

[0258] Among them, the behavior feature processing includes converting the time feature (time) into a timestamp, calculating the time gap time_gap between adjacent behaviors, and also includes extracting periodic features; normalizing the interaction duration (duration) and classifying it as short / medium / long duration according to business logic; and directly encoding the behavior type (click) into a category label (such as click = 0, favorite = 1, share = 2).

[0259] In some embodiments, sequence construction can be performed on the sequence S'.

[0260] For example, keep the last n behaviors, truncate the excess, and use <pad>Filling;

[0261] For example, splice the multi-dimensional features of each behavior item into a unified vector, such as: [content ID_emb, type_emb, cat1_emb, cat2_emb, duration_norm, time_gap, click_type];

[0262] For example, add behavioral position encoding (such as the position embedding of Transformer);

[0263] In some embodiments, the sequence S' can be normalized and stored.

[0264] For example, normalize continuous values such as duration and time_gap.

[0265] (2) Construction of the interest generality curve.

[0266] 1. Obtain the behavioral activity of the target object;

[0267] 2. Based on the behavioral activity, determine that the window size of the sliding window = 40;

[0268] 3. Place the sliding window in the historical behavior sequence to obtain the first subsequence S 1 =[s 1 , s 2 , s 3 …s 39 ;

[0269] 4. Move the sliding window in the historical behavior sequence with a step size of 1 to obtain the second subsequence S 2 =[s 2 , s 3 , s 4 …s 40 ;

[0270] 5. Determine the similarity between the corresponding term pairs in the first subsequence S 1 and the second subsequence S 2 as [cos(s 1 , s 2 ), cos(s 2 , s 3 ), cos(s 3 , s 4 ), …, cos(s 39 , s 40 ;

[0271] 6. Generate the interest generality curve [1 - cos(s 1 , s 2 ), 1 - cos(s 2 , s 3 ),1 - cos(s 3 ,s 4 ),…,1 - cos(s 39 ,s 40 )], the interest breadth curve includes the interest breadth, and the interest breadth = 1 - similarity.

[0272] Among them, there are various methods for calculating the behavior activity of the target object. For example, in some embodiments, the weighted sum of the login frequency, online duration, and interaction times can be calculated to obtain the behavior activity of the target object:

[0273] Behavior activity = k1 (number of logins / time period) + k2 (total online duration / time period) + k3 (w1 number of likes + w2 number of comments + w3 number of shares +...)

[0274] Among them, k1 to k3, w1 to w3 are all weights and can be set according to actual needs.

[0275] Among them, the window size of the sliding window can be determined based on the relationship between the activity and the activity threshold. For example, if the activity ≥ 5, the window size is set to the first size; if 5 > activity ≥ 10, the window size is set to the second size; if 10 > activity ≥ 15, the window size is set to the third size; if 15 > activity ≥ 20, the window size is set to the fourth size, and so on.

[0276] In some embodiments, the behavior activity of the target object can be updated in real time to obtain the latest behavior activity, so the window size can also be adjusted in real time.

[0277] For example, if object A clicks 20 times per minute for 3 consecutive minutes, it is detected that the activity is continuously higher than the threshold, and the window is reduced to 40; subsequently, if object A only clicks 3 times in the past 5 minutes, it is detected that the activity is continuously lower than the threshold, and the window is expanded to 200 to accumulate more behavior data.

[0278] (3) The prediction model determines the recommended content suitable for the target object according to the interest breadth curve.

[0279] 1. Obtain the content to be detected in the candidate set, and the content to be detected may include the domain label D;

[0280] 2. The interest breadth curve [1 - cos(s 1 ,s 2 ),1 - cos(s 2 ,s 3 ),1 - cos(s 3 ,s 4 ),…,1 - cos(s 39 ,s 40 )], the historical behavior sequence S = [s 1 , s 2 , s 3 … s n , splice the content features and behavior features of each behavior item to obtain the personal feature U of the target object;

[0281] 3. Use the click type prediction network in the prediction model to predict the click type of the target object for the content to be detected based on the personal feature U and the domain label D of the content to be detected;

[0282] 4. Use the reading duration prediction network in the prediction model to predict the reading duration of the target object for the content to be detected based on the personal feature and the domain label of the content to be detected;

[0283] 5. Calculate the behavior metrics according to the click type and the reading duration;

[0284] 6. Determine the recommendation score of the content to be detected based on the behavior metrics;

[0285] 7. Sort the recommendation scores of all the content to be detected in the candidate set, and select the recommended content from the candidate set.

[0286] Among them, the domain label can include the domain to which the content belongs, and can include various characteristic data such as its first-level category, second-level category, type, theme, etc. that can indicate the domain to which it belongs.

[0287] Among them, there are various ways of feature splicing, which can include horizontal splicing, vertical splicing, feature crossing, embedding splicing, sequence splicing, etc.

[0288] In some embodiments, the prediction model can construct a click type prediction training dataset based on the personal feature vector extracted from the historical behavior data, the domain label of the content to be detected, and its corresponding click type label; design a deep neural network structure composed of a fully connected layer and a softmax classifier, input the personal feature and the domain label after feature cross-fusion into the network; introduce a cross-entropy loss function to measure the difference between the predicted click type distribution and the true label, use the Adam optimization algorithm to iteratively update the network parameters, and dynamically adjust the learning rate through an exponential decay strategy to improve the model convergence effect.

[0289] In some embodiments, a training sample set containing personal characteristics, domain labels, and the correlation coefficient between the reading duration after standardization processing can be constructed based on the user's historical reading duration log data; a regression neural network architecture with a non-linear activation function can be built, and the user attributes and content domain features can be fused through a feature splicing layer; the mean squared error loss function is used to measure the deviation between the predicted duration and the actual duration, and the gradient descent algorithm with L2 regularization is used to optimize the network weight parameters, combined with an early stopping mechanism to prevent overfitting.

[0290] In some embodiments, the click type prediction network and the reading duration prediction network can adopt independent training pipelines, and the model iteration optimization can be completed through mini-batch data parallel computing under a distributed training framework, and finally a dual-task prediction model that can synchronously output the click type probability distribution and the reading duration prediction value can be generated.

[0291] Among them, the click type classification differentiates different interaction behaviors of users, and each type represents different user intentions. For example: for active clicks, a higher weight can be set; for passive clicks, such as accidental touches, a lower or negative weight can be set; for in-depth interactions, such as clicking on comment or share buttons, a higher weight can be set.

[0292] In some embodiments, the behavior metric can be set as:

[0293] Behavior metric = ∑(click type weight × number of clicks) + α × log(read duration + 1)

[0294] Among them, α is the adjustment coefficient of the reading duration, which is used to balance the influence of clicks and duration.

[0295] In some embodiments, the behavior metric can be converted into a standardized recommendation score from 0 - 1 or 1 - 5 points.

[0296] (III) Re-ranking stage.

[0297] In the re-ranking stage, the recommended content can be re-ranked to achieve the diversity control and strategy intervention of the recommended content, and finally the re-ranked recommended content is sent to the target object.

[0298] For example, in order to avoid the recommended content being too similar and improve diversity, the similarity between the contents can be calculated, and a diversity constraint can be introduced during sorting. For example, the maximum marginal relevance (MMR) can be introduced to comprehensively consider the relevance and diversity of the content during sorting. For example, clustering re-ranking can be introduced to cluster the recommended content, and then content is selected from different categories for re-ranking. For example, random sampling can be introduced, and on the basis of relevance, some contents are randomly shuffled to increase diversity.

[0299] For example, the recommendation order can be adjusted according to business requirements or strategies, and the sorting can be adjusted according to predefined rules. For example, weights can be assigned to different contents or features, and re-ranking can be performed by combining the original scores. For example, time decay processing can be performed on content with strong timeliness, and new content can be recommended preferentially.

[0300] For example, the recommendation order can be optimized according to the user's historical behavior or real-time feedback. For example, a machine learning model can be used to predict the user's click probability, and re-ranking can be performed according to the predicted value. For example, the order can be adjusted according to the matching degree between the object portrait of the target object and the content. For example, the recommendation order can be dynamically adjusted according to the user's real-time behavior.

[0301] For example, in order to balance multiple metrics such as click-through rate, diversity, and user satisfaction, a multi-objective optimization algorithm is used for re-ranking, and the sorting strategy is dynamically adjusted through reinforcement learning to optimize the long-term benefits.

[0302] For example, the recommendation order can be adjusted according to the user's current application scenario, such as time, location, device, etc.

[0303] By analyzing the shape of the interest curve in the embodiments of the present application, the preference characteristics of the target user can be deeply understood, so as to formulate a more accurate content recommendation strategy. When it is observed that the curve shows a relatively high average peak value, this indicates that the user shows diversified interest tendencies, and their content consumption behavior has a wide coverage. For such users, it is appropriate to adopt a diversified content recommendation strategy to meet their wide range of interest needs by providing high-quality content across different fields. On the contrary, when the curve shows a relatively low average trough value, this often means that the user has a more focused interest area and shows a stronger preference for specific types of content. In this case, the recommendation system should focus on exploring the user's core interest points and providing vertical content that highly matches their preferences. In addition, the trend changes shown by the curve also contain important information: when the interest range shows an expanding trend, new content related to their existing interests but with exploratory nature can be appropriately introduced to guide the user to discover potential interests; while when the interest range shows a shrinking trend, more attention should be paid to their core interest areas and more accurate vertical content should be provided to improve user stickiness. This recommendation strategy based on the dynamic changes of the interest curve can effectively improve the accuracy of content recommendation and user satisfaction.

[0304] Therefore, in some embodiments, the steps performed in the re-ranking stage may include:

[0305] 1. Based on the interest breadth curve, determine the interest breadth type of the target object;

[0306] 2. If the target object is of the high-breadth type, increase the sorting priority weight of the recommended content that conforms to the high-breadth type;

[0307] 3. If the target object is of a low interest breadth type, increase the sorting priority weight of the recommended content that conforms to the low interest breadth type;

[0308] 4. Re-sort all the recommended content according to the sorting priority weights of the recommended content to obtain the re-sorted recommended content;

[0309] 5. Determine the domain labels for which the target object has never had an interaction behavior based on the historical behavior sequence of the target object. The domain labels represent the content domains to which the content belongs;

[0310] 6. Based on the interest breadth curve, determine the interest breadth types of the target object at different time periods;

[0311] 7. If the target object is of a high interest breadth type during the target time period, select un-explored content from the content pool. The un-explored content is the content with the domain labels for which the target object has never had an interaction behavior;

[0312] 8. Determine the mixing ratio weight according to the interest breadth curve. The mixing ratio weight can be used to adjust the ratio of the recommended content to the un-explored content when sending;

[0313] 9. Send the un-explored content simultaneously when sending the recommended content to the target object according to the mixing ratio weight.

[0314] In some embodiments, when the interest breadth curve shows that the click-through rates of the target object in multiple domains such as politics, technology, and entertainment all remain above 0.8, and the curve variance is less than 0.1, it indicates that the user has a stable and broad news reading interest. At this time, a cross-domain mixed recommendation strategy can be adopted, combining diversified content such as international current politics, AI technology breakthroughs, and star dynamics in a single push card, and interspersing local people's livelihood reports and sports event bulletins in the information stream. Each single push covers 3-5 different domains.

[0315] In some embodiments, when the interest breadth curve fluctuates by more than 3 standard deviations within 15 minutes in the morning, the news hot spot tracking system can be accessed in real time to preferentially push instant content such as sudden financial data and major social events, while reducing the recommendation weight of in-depth reports. If the curve shows a continuous stable state (fluctuation amplitude <0.2) for 0.5 hours in the evening on weekdays, it will automatically switch to the special planning recommendation mode and push continuous in-depth reports such as "Series of Observations on the Global Climate Summit" and "Ten-Day Talks on AI Ethics".

[0316] In some embodiments, by analyzing the spatial feature vectors of the interest popularity curve, when it is detected that the user's attention duration to the "regional conflict" type label has increased by 120% compared to the previous month, a dynamic geographical recommendation matrix can be constructed: prioritize pushing the content of on-the-spot reports by war correspondents in the Middle East region to account for 40% of the total, simultaneously reduce the pushing frequency of local municipal news, and recommend associated energy market analysis content through the diffusion of the interest graph.

[0317] In some embodiments, in the news hot spot exploration mechanism, when the slope at the end of the curve exceeds a threshold (such as Δ>0.7 / 10min), trigger the hot spot prediction engine: extract potential interest points based on the user's real-time reading trajectory (such as detecting 3 consecutive clicks on reports about quantum computing), and immediately call the news semantic graph to generate a cross-dimensional content package, including an associated news combination such as "Latest Breakthroughs in Quantum Computing", "Stock Price Fluctuations of Related Enterprises", and "Exclusive Interviews with Research Teams" for pushing, with the exploration depth increased by 2.3 times compared to the baseline.

[0318] In some embodiments, in the news push reordering strategy, when it is detected that the information entropy value of the interest popularity curve is higher than the historical 80th percentile, adopt a diversity enhancement algorithm: mix and sort the three types of content, namely breaking news, background interpretation, and expert comments, in a ratio of 4:3:3, ensuring that each single feed stream contains at least 2 news labels that the user has not encountered before. At the same time, introduce a time decay factor to reduce the sorting weight of hot news 30 minutes ago by 50%, and give priority to displaying real-time updated live on-site text and pictures.

[0319] In some embodiments, for the first-week push to new users, dynamically adjust the geographical coverage range through the initial slope of the popularity curve: when the rising slope on the first day of the curve > 45°, adopt a 5:5 push ratio of "global hot spots + local important news"; if the slope < 30°, focus on local news to account for 70% of the total, and at the same time embed a small amount of national news with high interaction rates as interest probes.

[0320] In some embodiments, in the news interest decay warning, when the persistence index of the curve in a specific field (such as sports events) has decreased by more than 20% for 5 consecutive days, automatically trigger the content preservation mechanism: suspend the push of regular event reports, and instead recommend associated derivative content such as "Cross-border Movements of Sports Stars" and "Release of New Sports Technology Products", and gradually resume the original content supply after the user interaction rate rebounds to the threshold.

[0321] The design of the recommendation strategy based on the interest popularity curve has diverse implementation paths, and personalized content push can be carried out through various technical means and algorithm models. Specifically, the following multiple solutions can be adopted:

[0322] For example, in some embodiments, a machine learning model can be used to construct a content scoring system, quantitatively evaluate various types of content based on the interest breadth curve, and preferentially recommend the content with higher scores to the target users. At the same time, a behavior prediction model can also be used to predict the target user's future interest areas based on the interest breadth curve and push relevant topic content accordingly.

[0323] For example, in some embodiments, through clustering and classification algorithms, the user group can be segmented according to the interest breadth curve, and content that matches its category characteristics can be screened from the content library for precise push. In addition, during the regular content push process, a certain proportion of content in new fields can be appropriately incorporated according to the user classification results to expand the user's interest boundary.

[0324] For example, in some embodiments, by combining the fluctuation characteristics of the interest breadth curve with the time decay effect, the timeliness and diversity of the recommended content can be dynamically optimized. Specifically, by dividing time windows to analyze the stability of user interests: when the curve fluctuates violently, focus on pushing time-sensitive content such as hot news; when the curve is stable, recommend in-depth content such as feature articles.

[0325] For example, in some embodiments, the weight ratio of collaborative filtering and content filtering can be flexibly adjusted according to the morphological characteristics of the interest breadth curve. When the peak of the curve exceeds the preset threshold, enhance the content filtering weight to enrich the recommendation diversity; when the trough of the curve is lower than the preset value, strengthen the collaborative filtering weight and focus on pushing the preferred content of similar users. At the same time, the rapid switching of the recommendation strategy is achieved through the change speed of the curve slope.

[0326] For example, in some embodiments, a user interest map can be constructed based on the interest breadth curve to achieve cross-domain content recommendation. The specific method is to map the user's historical behavior sequence to interest nodes and construct a relevance map: when the interest breadth curve shows an upward trend, recommend content in adjacent interest fields; when the curve drops, focus on the core interest nodes to avoid over-expansion.

[0327] In some embodiments, when the end of the object's interest breadth curve is rising, increase the probability of exploring new field content; when the end of the curve is falling, reduce exploration and preferentially recommend the verified interest content.

[0328] In some embodiments, if the target object is of a high breadth type, the content in the unexplored fields of the target object can be ranked in the front part, and the content in the fields it is known to be interested in can be ranked in the back part; if the target object is of a medium breadth type, the content in the unexplored fields of the target object and the content in the fields it is known to be interested in can be evenly arranged; if the target object is of a low breadth type, the content in the unexplored fields of the target object can be ranked in the back part, and the content in the fields it is known to be interested in can be ranked in the front part.

[0329] In some embodiments, un-explored content can be recommended dynamically by combining the historical behavior sequence of the object, the interest breadth curve, and by time period. For example, when the object is in a time period with broad interests and strong exploration tendency, more content in un-explored fields can be recommended to it, such as articles in the "travel" or "art" categories that the object has never clicked on; when the object is in a time period with concentrated interests and weak exploration tendency, less content in un-explored fields can be recommended to it.

[0330] In some embodiments, if the target object is of the first breadth type in the target time period, un-explored content is selected from the content pool. The un-explored content is content with domain labels in fields where the target object has never had an interaction behavior; when sending recommended content to the target object in the target time period, the un-explored content is sent simultaneously.

[0331] The proportion of un-explored content sent can be dynamically adjusted according to the interest breadth curve. Therefore, in some embodiments, the mixing ratio weight can be determined according to the interest breadth curve, and the mixing ratio weight can be used to adjust the ratio of recommended content to un-explored content when sending; according to the mixing ratio weight, when sending recommended content to the target object, the un-explored content is sent simultaneously.

[0332] For example, through the interest breadth curve, the interest breadth type of the object in different time periods such as morning, noon, and evening can be identified in real time, and the recommendation strategy can be dynamically adjusted. For example, during the morning commute period, the target object is of the high breadth type, and this target object may be in an open state of information reception, and diverse content such as international news and un-explored scientific and technological trivia can be recommended; during the lunch break, the target object is of the medium breadth type, and moderately diverse content such as light entertainment, food, and short readings can be recommended; during the evening rest period, the target object is of the low breadth type, and precise content such as the target object's frequently watched movies and TV shows and in-depth long articles can be recommended.

[0333] For example, through the interest breadth curve, the interest breadth type of the object in different time periods on weekdays and non-working days can be identified in real time, and the recommendation strategy can be dynamically adjusted. For example, from Monday to Friday, the target object is in a working state, with fragmented time, and tends to efficiently obtain core interest content related to work and study. Therefore, it is mainly based on precise matching, and content in the target object's historical high-frequency interest fields is recommended to it, restricting the proportion of content in un-explored fields; for example, on Saturday and Sunday, the target object has ample time, more open interests, and tends to explore content in the fields of entertainment, leisure, and niche areas. Cross-domain content is recommended to it, and the proportion of content in un-explored fields is increased significantly.

[0334] These solutions together constitute a multi-level and multi-dimensional recommendation strategy system based on the interest breadth curve, which can effectively improve the accuracy of content recommendation and user satisfaction.

[0335] Reference Figure 2b , the ordinate is the value of the interest breadth, the abscissa is the time, the farther the curve is from the 0-axis, the more extensive the interest of the target object is. The closer the curve is to the 0-axis, the more concentrated the interest of the target object is. For the interest breadth curve of object A, the average similarity is 0.695 and the average interest breadth is 0.305. The average interest breadth of object A is relatively larger than that of object B. Therefore, when recommending news to object A, not only the interest matching degree should be considered, but also the diversity and exploration degree of interest should be improved simultaneously. For the interest breadth curve of object B, the average similarity is 0.798 and the average interest breadth is 0.202. The average interest breadth of object B is relatively smaller than that of object A. Therefore, when recommending news to object B, more emphasis should be placed on the interest matching degree, and the diversity and exploration degree of interest should be reduced.

[0336] The embodiments of the present application deeply mine the deep correlation information hidden in the behavior sequence of the target object, and significantly enhance the interest prediction ability of the recommendation model by constructing the interest breadth curve. Experimental data show that this method realizes an offline AUC improvement of 1.2% and an online CTR increase of 0.8% in the news recommendation system.

[0337] As can be seen from the above, the embodiments of the present application propose an interest breadth curve, which quantifies and represents the variation law of the object's interest breadth over time by constructing a similarity fluctuation curve of the time-series behavior sequence; the embodiments of the present application propose a sliding window mechanism to adaptively adjust the sequence analysis window size according to the object's behavior activity; the dual-network prediction model architecture designed in the embodiments of the present application integrates a click type prediction network and a reading duration prediction network for comprehensive scoring; the embodiments of the present application establish a diversity recommendation mechanism driven by interest breadth, and dynamically adjust the proportion of exploration content through a mixed ratio weight.

[0338] Compared with the existing content recommendation methods, this solution broadens the recommendation diversity through interest breadth analysis, overcomes the information cocoon effect caused by traditional methods; deeply mines the time-series correlation patterns of behavior sequences and extracts deep features not utilized by traditional methods; constructs a highly interpretable visualization analysis tool to improve the interpretability of the recommendation system. Therefore, the embodiments of the present application can improve the quality of content recommendation.

[0339] It can be understood that in the specific implementation of the present application, relevant data such as the historical behavior sequence of the target object and object information are involved. When the following embodiments of the present application are applied to specific products or technologies, permission or consent is required, and the collection, use, and processing of relevant data need to comply with the relevant laws, regulations, and standards of relevant countries and regions.

[0340] To better implement the above method, an embodiment of the present application further provides a content recommendation device, which can be specifically integrated in an electronic device, and the electronic device can be a device such as a terminal or a server. Among them, the terminal can be a device such as a mobile phone, a tablet computer, a smart Bluetooth device, a notebook computer, or a personal computer; the server can be a single server or a server cluster composed of multiple servers.

[0341] For example, in this embodiment, taking the content recommendation device specifically integrated in the server as an example, the method of the embodiment of the present application will be described in detail.

[0342] For example, as Figure 3a shown, the content recommendation device may include a history unit 310, a similarity unit 320, a construction unit 330, and a recommendation unit 340, as follows:

[0343] (1) History unit 310.

[0344] The history unit 310 can be used to obtain the historical behavior sequence of the target object. The historical behavior sequence may include behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content.

[0345] (2) Similarity unit 320.

[0346] The similarity unit 320 can be used to determine the similarity between two behavior items separated by a fixed interval in the historical behavior sequence.

[0347] For example, as Figure 3b shown, in some embodiments, the similarity unit 320 may include a sequence subunit 3201 and a similarity subunit 3202, where:

[0348] The sequence subunit 3201 can be used to extract a first subsequence and a second subsequence from the historical behavior sequence. The first subsequence includes the i-th to j-th behavior items in the historical behavior sequence, and the second subsequence includes the (i + k)-th to (j + k)-th behavior items in the historical behavior sequence, where i and j are positive integers, and k is a fixed interval.

[0349] Among them, in some embodiments, the sequence subunit 3201 can be used to:

[0350] Obtain the behavior activity of the target object;

[0351] Based on the behavior activity, determine the window size of the sliding window;

[0352] Place the sliding window in the historical behavior sequence to obtain a first subsequence, which includes the i-th to j-th behavior items in the historical behavior sequence, and the difference between i and j is one window size;

[0353] Move the sliding window in the historical behavior sequence with a step size of k to obtain a second subsequence.

[0354] The similarity sub-unit 3202 can be used to determine the similarity between the corresponding item pairs in the first subsequence and the second subsequence. The corresponding item pairs are two behavior items with corresponding positions in the first subsequence and the second subsequence.

[0355] Among them, in some embodiments, the similarity sub-unit 3202 is used for:

[0356] The corresponding item pairs include the first corresponding item in the first subsequence and the second corresponding item in the second subsequence. Determining the similarity between the corresponding item pairs in the first subsequence and the second subsequence includes:

[0357] Determine a first set, where the first set includes the first corresponding item in the first subsequence and its adjacent behavior items;

[0358] Determine a second set, where the second set includes the second corresponding item in the second subsequence and its adjacent behavior items;

[0359] According to the mean value of the first set and the mean value of the second set, obtain the similarity between the first corresponding item and the second corresponding item.

[0360] (III) Construction unit 330.

[0361] The construction unit 330 can be used to construct an interest popularity curve according to the similarity. The interest popularity curve reflects the change of interest popularity over time, and the interest popularity characterizes the degree of interest breadth of the target object.

[0362] (IV) Recommendation unit 340.

[0363] The recommendation unit 340 can be used to determine the recommended content suitable for the target object based on the interest popularity curve.

[0364] As Figure 3c shown, in some embodiments, the recommendation unit 340 may include a to-be-detected sub-unit 331, a scoring sub-unit 332, and a sorting sub-unit 333, where:

[0365] (1) To-be-detected sub-unit 331.

[0366] The to-be-detected sub-unit 331 can be used to obtain the content to be detected in the candidate set.

[0367] (2) Scoring sub-unit 332.

[0368] The scoring sub-unit 332 can be used to adopt a prediction model and predict the recommended score of the target object for the content to be detected based on the interest popularity curve.

[0369] In some embodiments, the behavior item may include content features and behavior features, the content to be detected may include domain labels, and the scoring sub-unit 332 may include a splicing sub-module 3321, a behavior index sub-module 3322, and a scoring sub-module 3323, where:

[0370] The splicing sub-module 3321 may be configured to splice the interest popularity curve, the content features and behavior features of each behavior item in the historical behavior sequence to obtain the personal features of the target object;

[0371] The behavior index sub-module 3322 may be configured to predict the behavior index of the target object for the content to be detected based on the personal features and the domain labels of the content to be detected;

[0372] The scoring sub-module 3323 may be configured to determine the recommended score of the content to be detected based on the behavior index.

[0373] In some embodiments, the prediction model may include a click type prediction network and a reading duration prediction network, and the behavior index sub-module 3322 may be configured to:

[0374] Adopt the click type prediction network to predict the click type of the target object for the content to be detected based on the personal features and the domain labels of the content to be detected;

[0375] Adopt the reading duration prediction network to predict the reading duration of the target object for the content to be detected based on the personal features and the domain labels of the content to be detected;

[0376] Calculate the behavior index according to the click type and the reading duration.

[0377] (3) Sorting sub-unit 333.

[0378] The sorting sub-unit 333 may be configured to select the recommended content from the candidate set by sorting the recommended scores of all the content to be detected in the candidate set.

[0379] As Figure 3d shown, in some embodiments, the recommendation unit 340 may further include an object information sub-unit 334, a rough sorting sub-unit 335, and a re-sorting sub-unit 336, as follows:

[0380] (4) Object information sub-unit 334, the object information sub-unit 334 may be configured to obtain the object information of the target object.

[0381] (5) Rough sorting sub-unit 335, which may be configured to screen multiple candidate contents from the content pool based on the object information to obtain a candidate set.

[0382] The recommendation unit 340 may further include:

[0383] (6) A reordering subunit 336, which can be used to reorder all recommended contents and send the reordered recommended contents to the target object.

[0384] In some embodiments, the reordering subunit 336 can be used to:

[0385] Determine the interest generality type of the target object based on the interest generality curve;

[0386] If the target object is of the first generality type, increase the sorting priority weight of the recommended contents that conform to the first generality type;

[0387] If the target object is of the second generality type, increase the sorting priority weight of the recommended contents that conform to the second generality type;

[0388] Reorder all recommended contents according to the sorting priority weights of the recommended contents.

[0389] In some embodiments, the recommendation unit 340 may further include:

[0390] An un-interacted subunit, which can be used to determine the domain labels that the target object has never had interaction behaviors with according to the historical behavior sequence of the target object, and the domain labels represent the content domains to which the contents belong;

[0391] A type subunit, which can be used to determine the interest generality types of the target object at different time periods based on the interest generality curve;

[0392] A selection subunit, which can be used to select un-explored contents in the content pool if the target object is of the first generality type in the target time period, and the un-explored contents are the contents with domain labels that the target object has never had interaction behaviors with;

[0393] A simultaneous sending subunit, which can be used to simultaneously send un-explored contents when sending recommended contents to the target object in the target time period.

[0394] In some embodiments, the simultaneous sending subunit can be used to:

[0395] Determine a mixing ratio weight according to the interest generality curve, and the mixing ratio weight can be used to adjust the ratio of the recommended contents to the un-explored contents when sending;

[0396] According to the mixing ratio weight, simultaneously send un-explored contents when sending recommended contents to the target object.

[0397] In specific implementation, the above-mentioned each unit can be implemented as an independent entity, or can be combined arbitrarily to be implemented as the same or several entities. For the specific implementation of the above-mentioned each unit, reference can be made to the foregoing method embodiments, which will not be elaborated herein.

[0398] In the embodiments of the present application, the term "module" or "unit" refers to a computer program with a predetermined function or a part of a computer program, which works together with other related parts to achieve a predetermined goal, and can be implemented in whole or in part by using software, hardware (such as a processing circuit or a memory), or a combination thereof. Similarly, one processor (or multiple processors or memories) can be used to implement one or more modules or units. In addition, each module or unit can be a part of an overall module or unit that includes the functions of that module or unit.

[0399] As can be seen from the above, the content recommendation device in this embodiment obtains the historical behavior sequence of the target object from the historical unit. The historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content. The similarity unit determines the similarity between two behavior items separated by a fixed interval in the historical behavior sequence. The construction unit constructs an interest popularity curve based on the similarity. The interest popularity curve reflects the change of interest popularity over time, and the interest popularity represents the extent of the target object's interest. The recommendation unit determines the recommended content suitable for the target object based on the interest popularity curve.

[0400] Thus, the embodiments of the present application can improve the quality of content recommendation.

[0401] The embodiments of the present application also provide an electronic device, which can be a device such as a terminal, a server, etc. Among them, the terminal can be a mobile phone, a tablet computer, a smart Bluetooth device, a laptop computer, a personal computer, etc.; the server can be a single server or a server cluster composed of multiple servers, etc.

[0402] In some embodiments, the content recommendation device can also be integrated in multiple electronic devices. For example, the content recommendation device can be integrated in multiple servers, and multiple servers are used to implement the content recommendation method of the present application.

[0403] In this embodiment, the electronic device in this embodiment will be described in detail by taking the example that the electronic device is a server. For example, as Figure 4 shown, it shows a schematic structural diagram of the electronic device involved in the embodiments of the present application. Specifically:

[0404] The electronic device may include a processor 410 with one or more processing cores, a memory 420 with one or more computer-readable storage media, a power supply 430, an input module 440, a communication module 450, and other components. Those skilled in the art can understand that Figure 4 the structural diagram of the electronic device shown in does not constitute a limitation on the electronic device, and it may include more or fewer components than shown, or combine certain components, or have different component arrangements. Among them:

[0405] The processor 410 is the control center of the electronic device, connecting various parts of the entire electronic device through various interfaces and circuits. By running or executing software programs and / or modules stored in the memory 420, and by invoking the data stored in the memory 420, it performs various functions of the electronic device and processes data, thereby conducting an overall inspection of the electronic device. In some embodiments, the processor 410 may include one or more processing cores; in some embodiments, the processor 410 may integrate an application processor and a modem processor. Among them, the application processor mainly processes the operating system, object interfaces, application programs, etc., and the modem processor mainly processes wireless communications. It can be understood that the above-mentioned modem processor may not be integrated into the processor 410 either.

[0406] The memory 420 can be used to store software programs and modules. The processor 410 executes various functional applications and data processing by running the software programs and modules stored in the memory 420. The memory 420 may mainly include a program storage area and a data storage area. Among them, the program storage area can store the operating system, application programs required for at least one function (such as the sound playback function, image playback function, etc.), etc.; the data storage area can store the data created according to the use of the electronic device. In addition, the memory 420 may include high-speed random access memory, and may also include non-volatile memory, such as at least one magnetic disk storage device, flash memory device, or other volatile solid-state storage devices. Correspondingly, the memory 420 may also include a memory controller to provide the processor 410 with access to the memory 420.

[0407] The electronic device further includes a power supply 430 that powers each component. In some embodiments, the power supply 430 may be logically connected to the processor 410 through a power management system, so as to implement functions such as management of charging, discharging, and power consumption management through the power management system. The power supply 430 may also include any components such as one or more DC or AC power supplies, a recharge system, a power failure detection circuit, a power converter or inverter, and a power status indicator.

[0408] The electronic device may further include an input module 440, which can be used to receive input digital or character information, and generate keyboard, mouse, joystick, optical or trackball signal inputs related to object settings and function controls.

[0409] The electronic device may further include a communication module 450. In some embodiments, the communication module 450 may include a wireless module. The electronic device can perform short-range wireless transmission through the wireless module of the communication module 450, thereby providing wireless broadband Internet access for the object. For example, the communication module 450 can be used to help the object send and receive emails, browse the web, and access streaming media, etc.

[0410] Although not shown, the electronic device may further include a display unit and the like, which will not be elaborated herein. Specifically, in this embodiment, the processor 410 in the electronic device will load the executable files corresponding to the processes of one or more application programs into the memory 420 according to the following instructions, and the processor 410 will run the application programs stored in the memory 420, thereby realizing various functions as follows:

[0411] Obtain the historical behavior sequence of the target object, where the historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content;

[0412] Determine the similarity between two behavior items separated by a fixed interval in the historical behavior sequence;

[0413] Construct an interest breadth curve according to the similarity, where the interest breadth curve reflects the change of interest breadth over time, and the interest breadth represents the degree of interest breadth of the target object;

[0414] Based on the interest breadth curve, determine the recommended content suitable for the target object.

[0415] For the specific implementation of each of the above operations, reference may be made to the previous embodiments, which will not be elaborated herein.

[0416] As can be seen from the above, the embodiments of the present application can improve the quality of content recommendation.

[0417] Those of ordinary skill in the art can understand that all or part of the steps in the various methods of the above embodiments can be completed by instructions, or by controlling relevant hardware through instructions. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0418] Therefore, the embodiments of the present application provide a computer-readable storage medium, in which multiple instructions are stored, and these instructions can be loaded by a processor to execute the steps in any content recommendation method provided by the embodiments of the present application. For example, the instructions can execute the following steps:

[0419] Obtain the historical behavior sequence of the target object, where the historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the interaction behaviors of the target object with respect to the content;

[0420] Determine the similarity between two behavior items separated by a fixed interval in the historical behavior sequence;

[0421] Construct an interest breadth curve according to the similarity, where the interest breadth curve reflects the change of interest breadth over time, and the interest breadth represents the degree of interest breadth of the target object;

[0422] Based on the interest breadth curve, determine the recommended content suitable for the target object.

[0423] Among them, the storage medium may include: read-only memory (ROM, Read Only Memory), random access memory (RAM, Random Access Memory), magnetic disk or optical disc, etc.

[0424] According to one aspect of the present application, a computer program product or a computer program is provided. The computer program product or the computer program includes computer instructions, and the computer instructions are stored in a computer-readable storage medium. The processor of the electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions, so that the electronic device executes the methods provided in the various optional implementation manners of the content push aspect or the message flow push aspect such as news, video, article, advertisement, etc. provided in the above embodiments.

[0425] Since the instructions stored in the storage medium can execute the steps in any one of the content recommendation methods provided in the embodiments of the present application, the beneficial effects that can be achieved by any one of the content recommendation methods provided in the embodiments of the present application can be realized. For details, see the previous embodiments and will not be repeated here.

[0426] The above has introduced in detail a content recommendation method, device, electronic device, and computer-readable storage medium provided by the embodiments of the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the method and its core idea of the present application; at the same time, for those skilled in the art, according to the idea of the present application, there will be changes in the specific implementation manner and application scope. In summary, the content of this specification should not be construed as a limitation to the present application.< / pad> < / unk>

Claims

1. A content recommendation method, characterized in that: include: Acquire a historical behavior sequence of a target object, wherein the historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent interactive behaviors of the target object with respect to content; Determining the similarity between two of the behavior items separated by a fixed interval in the historical behavior sequence; According to the similarity, construct an interest universality curve, wherein the interest universality curve reflects the change of interest universality over time, and the interest universality represents the extensiveness of the interest of the target object; Based on the interest universality curve, recommended content suitable for the target object is determined.

2. The content recommendation method according to claim 1, characterized in that: The determining the similarity between two behavior items separated by a fixed interval in the historical behavior sequence includes: Extracting a first subsequence and a second subsequence from the historical behavior sequence, wherein the first subsequence includes the i-th to j-th behavior items in the historical behavior sequence, and the second subsequence includes the i+k-th to j+k-th behavior items in the historical behavior sequence, wherein i and j are both positive integers, and k is a fixed interval; The similarity between the orthogonal item pairs in the first subsequence and the second subsequence is determined, where the orthogonal item pairs are two behavior items at corresponding positions in the first subsequence and the second subsequence.

3. The content recommendation method according to claim 2, characterized in that: The orthogonal item pair includes a first orthogonal item in the first subsequence and a second orthogonal item in the second subsequence, and determining the similarity between the orthogonal item pair in the first subsequence and the second subsequence includes: Determine a first set, the first set including the first homonymous item and its adjacent behavior items in the first subsequence; Determine a second set, the second set including the second homonymous item and its adjacent behavior items in the second subsequence; The similarity between the first homonymous item and the second homonymous item is obtained according to the mean of the first set and the mean of the second set.

4. The content recommendation method according to claim 2, characterized in that: The extracting the first subsequence and the second subsequence from the historical behavior sequence comprises: Obtaining the behavioral activity of the target object; Based on the activity level of the behavior, determining a window size of the sliding window; Placing the sliding window in the historical behavior sequence to obtain a first subsequence, wherein the first subsequence includes the i-th to j-th behavior items in the historical behavior sequence, and the difference between i and j is one of the window size; The sliding window is moved in the historical behavior sequence with a step length of k to obtain a second subsequence.

5. The content recommendation method according to claim 1, characterized in that: The determining, based on the interest universality curve, the recommended content adapted for the target object includes: Obtain the content to be detected in the candidate set; Using a prediction model, based on the interest prevalence curve, predict the recommendation score of the target object for the content to be detected; Recommended content is selected from the candidate set by sorting the recommendation scores of all the to-be-detected content in the candidate set.

6. The content recommendation method according to claim 5, characterized in that: The behavior item includes content features and behavior features, the content to be detected includes a domain label, and the prediction model is used to predict the recommendation score of the target object for the content to be detected based on the interest universality curve, including: The interest universality curve, the content features of each behavior item in the historical behavior sequence, and the behavior features are combined to obtain the personal features of the target object; Predicting the target object's behavior indicators for the content to be detected based on the personal characteristics and the domain label of the content to be detected; Based on the behavior indicator, a recommendation score for the content to be detected is determined.

7. The content recommendation method according to claim 6, characterized in that: The prediction model includes a click type prediction network and a reading time prediction network. The prediction of the target object's behavior indicators for the content to be detected based on the personal characteristics and the domain label of the content to be detected includes: Using the click type prediction network, based on the personal characteristics and the domain label of the content to be detected, predict the click type of the target object for the content to be detected; Using the reading time prediction network, based on the personal characteristics and the domain label of the content to be detected, predict the reading time of the target object for the content to be detected; Behavioral indicators are calculated based on the click type and reading time.

8. The content recommendation method according to claim 5, characterized in that: Also includes: Acquire object information of the target object; Based on the object information, multiple candidate contents are screened in the content pool to obtain a candidate set; After determining the recommended content suitable for the target object, the method further includes: All recommended contents are reordered, and the reordered recommended contents are sent to the target object.

9. The content recommendation method according to claim 8, characterized in that: The above method reorders all recommended contents, including: Based on the interest universality curve, determining the interest universality type of the target object; If the target object is of the first generality type, increasing the ranking priority weight of the recommended content that meets the first generality type; If the target object is of the second generality type, increasing the ranking priority weight of the recommended content that meets the second generality type; All the recommended contents are reordered according to the order priority weights of the recommended contents.

10. The content recommendation method according to claim 1, characterized in that: After determining the recommended content suitable for the target object, the method further includes: Determine, according to the historical behavior sequence of the target object, a domain label in which the target object has never generated any interactive behavior, wherein the domain label represents the content domain to which the content belongs; Based on the interest universality curve, determining the interest universality type of the target object in different time periods; If the target object is of the first extensiveness type in the target time period, unexplored content is selected from the content pool, where the unexplored content is content with a domain label with which the target object has never interacted; When the recommended content is sent to the target object in the target time period, the unexplored content is sent simultaneously.

11. The content recommendation method according to claim 10, characterized in that: When sending the recommended content to the target object in the target time period, simultaneously sending the unexplored content includes: Determine a mixing ratio weight according to the interest universality curve, wherein the mixing ratio weight is used to adjust the ratio of recommended content to unexplored content when it is sent; According to the mixed ratio weight, when the recommended content is sent to the target object, the unexplored content is also sent at the same time.

12. A content recommendation device, characterized in that: include: A history unit is used to obtain a historical behavior sequence of a target object, wherein the historical behavior sequence includes behavior items arranged in chronological order, and the behavior items represent the target object's interactive behavior with respect to the content; Similarity unit, used to determine the similarity between two behavior items separated by a fixed interval in the historical behavior sequence; A construction unit, constructing an interest universality curve according to the similarity, wherein the interest universality curve reflects the change of interest universality over time, and the interest universality represents the extensiveness of the interest of the target object; The recommendation unit is used to determine the recommended content suitable for the target object based on the interest universality curve.

13. An electronic device, characterized in that: The invention comprises a processor and a memory, wherein the memory stores a plurality of instructions; the processor loads instructions from the memory to execute the steps in the content recommendation method as claimed in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a plurality of instructions, and the instructions are suitable for being loaded by a processor to execute the steps in the content recommendation method according to any one of claims 1 to 11.

15. A computer program, comprising computer instructions, wherein the computer instructions are stored in a computer-readable storage medium; a processor of an electronic device reads the computer instructions from the computer-readable storage medium, and the processor executes the computer instructions to perform the steps in the content recommendation method as described in any one of claims 1 to 11.

Images