A video recommendation method, apparatus, electronic device, and storage medium

Abstract

This invention provides a video recommendation method, apparatus, electronic device, and storage medium. When determining the correspondence between videos, for each first preset tag, based on the number of clicks on each video corresponding to the first preset tag and the search frequency of all first preset tags, the weight of each video corresponding to the first preset tag is determined. Based on the weights of each video corresponding to each first preset tag, a first similarity is calculated between the videos. The correspondence between the videos is determined based on the first similarity. Since the number of clicks can reflect user preferences for videos and tags, and tags can objectively reflect the similarity between videos, this invention, by applying the embodiments of the invention, allows the target related videos for the primary demand video to be obtained based on the similarity of the videos themselves and a fitting of real user behavior. Therefore, it can better reflect the correlation between videos and improve the accuracy of video recommendations.

Description

Technical Field

[0001] This invention relates to the field of video recommendation technology, and in particular to a video recommendation method, apparatus, electronic device, and storage medium. Background Technology

[0002] In audio and video platforms, users typically obtain desired audio and video content by entering search terms. Based on the user's search terms, the platform recommends related videos to stimulate potential viewing demand, increase average viewing time per user, and improve user experience.

[0003] Taking video platforms as an example, currently, video platforms typically recommend videos related to the user's search target video based on the similarity between the tags contained in the video. A video usually includes multiple tags, but the relevance of each tag to the video content is not the same. For example, a video contains the tags "suspense" and "comedy," and the main content of the video is suspense and mystery, while comedic elements only account for a small part of the video content. Therefore, for this video, the tag "suspense" is more relevant to the video content, while the tag "comedy" is less relevant.

[0004] In related technologies, the relevance of the tags contained in the video to the video is not considered when recalling related videos based on tags. As a result, when recalling related videos of the target video based on tags, videos that contain the same tags as the target video but have low relevance to those tags are recalled, resulting in low accuracy of recommended videos. Summary of the Invention

[0005] The purpose of this invention is to provide a video recommendation method, apparatus, device, and storage medium to improve the accuracy of video recommendations. The specific technical solution is as follows:

[0006] In a first aspect of this invention, a video recommendation method is provided, the method comprising:

[0007] Based on the target search term, the video with the highest similarity to the target search term is selected as the primary demand video;

[0008] Based on the correspondence between the videos, the videos corresponding to the main demand video are obtained as target-related videos and recommended.

[0009] The correspondence between the videos is determined in advance through the following steps:

[0010] For each first preset tag, the weight of each video corresponding to the first preset tag is determined based on the number of clicks of each video corresponding to the first preset tag and the search frequency of all first preset tags, wherein the number of clicks of the video corresponding to the first preset tag is the number of times the video is clicked in the search results obtained when searching with the first preset tag;

[0011] Based on the weights of each video corresponding to each of the first preset tags, a first similarity is calculated between each video; the first similarity is positively correlated with the weights of each video corresponding to the same first preset tag among the first preset tags;

[0012] The correspondence between the videos is determined based on the first similarity between them.

[0013] In one possible embodiment, calculating the first similarity between the videos based on the weights corresponding to the first preset tags for each video includes:

[0014] Each of the aforementioned videos is taken as a target video, and for each of the aforementioned target videos, all the first preset tags corresponding to the target video are determined as target tags;

[0015] For each target tag, the videos included in the search results obtained when searching with the target tag are determined as candidate videos corresponding to the target video for the target tag.

[0016] For each target tag, according to a preset algorithm, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag;

[0017] A first similarity is determined between the target video and each candidate video corresponding to each target tag.

[0018] In one possible embodiment, the first preset tag includes: a video type tag and an actor tag;

[0019] For each target tag, according to a preset algorithm, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag, including:

[0020] For each target video type tag, the similarity between the target video and each of the candidate videos is calculated based on the weight of the target video corresponding to the target video type tag, the weight of each candidate video corresponding to the target video type tag, and the inverse document frequency of the target video type tag.

[0021] For each target actor tag, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target actor tag and the weight of each candidate video corresponding to the target actor tag.

[0022] In one possible embodiment, the method further includes:

[0023] Based on the preset weights corresponding to the second preset tags of each video, the second similarity between each video is calculated;

[0024] The step of determining the video corresponding to each video based on the first similarity between the videos includes:

[0025] Based on the first similarity and the second similarity between the videos, the corresponding video for each video is determined.

[0026] In one possible embodiment, the weight of the video corresponding to the first preset tag is positively correlated with the number of clicks on the video corresponding to the first preset tag, and negatively correlated with the search frequency of all first preset tags.

[0027] In one possible embodiment, determining the video with the highest similarity to the target search term as the primary demand video based on the target search term includes:

[0028] Video search results are obtained based on the target search terms;

[0029] Sort the search results;

[0030] Select the first video in the search results as the primary video in your search query; or...

[0031] The video with the most clicks in the historical search records of the target search term is selected as the primary video.

[0032] In one possible embodiment, the step of obtaining the video corresponding to the main demand video as the target related video and recommending it includes:

[0033] Obtain the video corresponding to the main demand video as the target related video;

[0034] Sort the videos related to the target;

[0035] Displays the target-related videos in sorted order.

[0036] In a second aspect of the invention, a video recommendation device is also provided, the device comprising:

[0037] The main demand video determination module is used to determine the video with the highest similarity to the target search term as the main demand video based on the target search term;

[0038] The target-related video recommendation module is used to obtain videos corresponding to the main demand video as target-related videos and recommend them based on the correspondence between the videos; wherein, the correspondence between the videos is determined in advance through the following steps:

[0039] For each first preset tag, the weight of each video corresponding to the first preset tag is determined based on the number of clicks of each video corresponding to the first preset tag and the search frequency of all first preset tags, wherein the number of clicks of the video corresponding to the first preset tag is the number of times the video is clicked in the search results obtained when searching with the first preset tag;

[0040] Based on the weights of each video corresponding to each target first preset label, a first similarity is calculated between each video; the first similarity is positively correlated with the weights of each video corresponding to the same first preset label among the first preset labels;

[0041] The correspondence between the videos is determined based on the first similarity between them.

[0042] In one possible embodiment, calculating the first similarity between the videos based on the weights corresponding to the first preset tags for each video includes:

[0043] Each of the aforementioned videos is taken as a target video, and for each of the aforementioned target videos, all the first preset tags corresponding to the target video are determined as target tags;

[0044] For each target tag, the videos included in the search results obtained when searching with the target tag are determined as candidate videos corresponding to the target video for the target tag.

[0045] For each target tag, according to a preset algorithm, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag;

[0046] A first similarity is determined between the target video and each candidate video corresponding to each target tag.

[0047] In one possible embodiment, the first preset tag includes: a video type tag and an actor tag;

[0048] For each target tag, according to a preset algorithm, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag, including:

[0049] For each target video type tag, the similarity between the target video and each of the candidate videos is calculated based on the weight of the target video corresponding to the target video type tag, the weight of each candidate video corresponding to the target video type tag, and the inverse document frequency of the target video type tag.

[0050] For each target actor tag, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target actor tag and the weight of each candidate video corresponding to the target actor tag.

[0051] In one possible embodiment, the device further includes:

[0052] The second similarity calculation module is used to calculate the second similarity between the videos based on the preset weights of the videos corresponding to the second preset tags.

[0053] Determining the correspondence between the videos based on a first similarity among them includes:

[0054] The correspondence between the videos is determined based on the first similarity and the second similarity between the videos.

[0055] In one possible embodiment, the weight of the video corresponding to the first preset tag is positively correlated with the number of clicks on the video corresponding to the first preset tag, and negatively correlated with the search frequency of all first preset tags.

[0056] In one possible embodiment, the main demand video determination module is specifically used to perform video search based on target search terms to obtain search results;

[0057] Sort the search results;

[0058] Select the first video in the search results as the primary video in your search query; or...

[0059] The video with the most clicks in the historical search records of the target search term is selected as the primary video.

[0060] In one possible embodiment, the target-related video recommendation module is specifically used to obtain the video corresponding to the main demand video as the target-related video;

[0061] Sort the videos related to the target.

[0062] Displays the target-related videos in sorted order.

[0063] In another aspect of the present invention, an electronic device is also provided, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0064] Memory, used to store computer programs;

[0065] When a processor executes a program stored in memory, it implements the video recommendation method steps described in any of the first aspects above.

[0066] In another aspect of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, and when executed by a processor, the computer program implements the video recommendation method described in any of the first aspects above.

[0067] In another aspect of the invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the video recommendation methods described in the first aspect above.

[0068] The video recommendation method, apparatus, electronic device, and storage medium provided in this invention determine the primary demand video through target search terms, and then, based on the correspondence between videos, obtain the videos corresponding to the primary demand video as target-related videos and recommend them. When determining the correspondence between videos, for each first preset tag, based on the number of clicks on each video corresponding to the first preset tag and the search frequency of all first preset tags, determine the weight of each video corresponding to the first preset tag. Based on the weight of each video corresponding to each first preset tag, calculate the first similarity between videos. Based on the first similarity between videos, determine the correspondence between videos. Since the number of clicks can reflect user preferences for videos and tags, and tags can objectively reflect the similarity between videos, the target-related videos for the primary demand video, obtained by applying this invention, are based on the similarity of the videos themselves and fitting and synthesizing user real behavior. Therefore, it can better reflect the correlation between videos and improve the accuracy of video recommendations. Attached Figure Description

[0069] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the accompanying drawings used in the description of the embodiments or the prior art will be briefly introduced below.

[0070] Figure 1 This is a flowchart illustrating a video recommendation method provided in an embodiment of the present invention;

[0071] Figure 2 This is a schematic diagram of a process for establishing a correspondence in an embodiment of the present invention;

[0072] Figure 3 This is another flowchart illustrating the establishment of a correspondence in an embodiment of the present invention;

[0073] Figure 4 This is a schematic diagram illustrating a specific example of the video recommendation method provided in this embodiment of the invention;

[0074] Figure 5 This is a schematic diagram of a video recommendation device provided in an embodiment of the present invention;

[0075] Figure 6 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0076] The technical solutions of the present invention will now be described with reference to the accompanying drawings in the embodiments of the present invention.

[0077] In related technologies, two common methods are used to retrieve related videos for a target video and recommend them to users: One method is to calculate the similarity between each video and the target video based on video tags or actors involved, thus retrieving related videos. These video tags include video type tags, subject matter tags, etc. The other method is to retrieve related videos based on user behavior. For example, collaborative filtering algorithms can be used to retrieve related audio and video content. Collaborative filtering algorithms involve pre-calculating the similarity between users and recommending videos viewed and saved by users similar to the target; or pre-calculating the similarity between items based on all users' historical preference data and then recommending items similar to the user's favorite items.

[0078] However, in both of the above methods, since each tag in a video and the weight of each actor in the video are different, recalling videos solely based on the similarity between video tags will result in the recall of videos with low relevance, leading to poor relevance between the recommended videos and the target videos, and thus low video recommendation accuracy. Furthermore, user behavior has a certain degree of randomness; therefore, recalling videos solely based on user behavior will also result in poor relevance between the recommended videos and the target videos, leading to low video recommendation accuracy.

[0079] To improve the accuracy of video recommendations, embodiments of the present invention provide a video recommendation method, apparatus, electronic device, and storage medium. The video recommendation method provided by the embodiments of the present invention will first be described by way of example.

[0080] The video recommendation method provided in this invention can be applied to any electronic device with video recommendation functionality. This electronic device can be a computer, server, mobile terminal, etc.

[0081] like Figure 1 As shown, Figure 1 A flowchart illustrating a video recommendation method provided in an embodiment of the present invention may specifically include the following steps:

[0082] Step S101: Based on the target search term, determine the video with the highest similarity to the target search term as the primary demand video.

[0083] Step S102: Based on the correspondence between the videos, obtain the videos corresponding to the main demand video as target-related videos and recommend them.

[0084] Among them, such as Figure 2 As shown, the correspondence between the above videos is established in advance through the following steps:

[0085] Step S201: For each first preset tag, based on the number of clicks of each video corresponding to the first preset tag and the search frequency of all first preset tags, determine the weight of each video corresponding to the first preset tag, wherein the number of clicks of a video corresponding to the first preset tag is the number of times the video is clicked in the search results obtained when searching with the first preset tag;

[0086] Step S202: Calculate the first similarity between videos based on the weights of each video corresponding to each first preset label; the first similarity is positively correlated with the weights of the same first preset labels in each video corresponding to each first preset label.

[0087] Step S203: Determine the correspondence between videos based on the first similarity between them.

[0088] In this embodiment of the invention, the primary demand video is determined through the target search term. Then, based on the correspondence between the videos, the videos corresponding to the primary demand video are obtained as target-related videos and recommended. When determining the correspondence between videos, for each first preset tag, the weight of each video corresponding to the first preset tag is determined based on the number of clicks on each video corresponding to the first preset tag and the search frequency of all first preset tags. Based on the weights of each video corresponding to each first preset tag, the first similarity between the videos is calculated. Based on the first similarity between the videos, the correspondence between the videos is determined. Since the number of clicks can reflect the user's preference for videos and tags, and tags can objectively reflect the similarity between videos, in this embodiment of the invention, the target-related videos for the primary demand video are obtained by combining the similarity of the videos themselves with fitting the user's actual behavior. Therefore, it can better reflect the correlation between videos and improve the accuracy of video recommendations.

[0089] The following is an exemplary description of steps S101-S102 and S201-S203:

[0090] In step S101, the target search term can be a precise search term entered by the user in the preset search box, or a broad search term. A precise search term (precise search query) refers to a user's main demand pointing to a specific video, such as "video A name", "video B name", etc. A broad search term (broad search query) refers to a user's main demand not for a specific video, but for a category of videos, such as: rural TV dramas, classic spy dramas, suspenseful crime movies, action movies featuring actor A, etc.

[0091] The aforementioned videos can be single videos or video albums. A video album may include multiple videos on the same theme, such as videos belonging to the same TV series, videos belonging to the same variety show, etc. This invention does not impose specific limitations on this.

[0092] As a specific implementation method, when determining the main demand video based on the above-mentioned target search terms, the main demand video can be obtained through the following steps:

[0093] Step S111: Perform a video search based on the target search term to obtain the search results;

[0094] Step S112: Sort the search results;

[0095] Step S113: Select the first video in the search results as the primary video requested, or...

[0096] Select the video with the most clicks in the historical search records of the target search term as the primary video.

[0097] In steps S111-S112, video searching and sorting can be performed using any feasible method. For example, in step S111, video searches can be performed based on the similarity between the video information of each video and the target search term to obtain search results. The aforementioned video information specifically refers to the video's descriptive information, which may include the video type, actors, main content keywords, etc. In step S112, the videos can be sorted from high to low based on their similarity to the target search term. The first video in the sorted results is then selected as the primary video requested.

[0098] As another specific implementation, the electronic device can maintain historical search records, recording each search term, the videos included in the search results for each search term, and the number of clicks on each video. For example, a historical search record table of "search terms, videos, and click counts" can be created, where the search terms can include both precise and general search terms. Upon receiving the target search term, the device can search for the corresponding videos in the table according to the target search term, and select the video with the most clicks in the historical search record of the target search term as the primary video requested.

[0099] After obtaining the aforementioned main requirement video, the target related videos can be retrieved from the correspondence between the videos, and then displayed. For example, the target related videos can be displayed in a preset related video display bar below the main requirement video. Alternatively, the target related videos can be displayed on the main requirement video playback page. This invention does not impose specific limitations on this approach.

[0100] Alternatively, a ranking model can be used to sort the videos related to the target, and then the sorted videos can be displayed. This ranking model could be a Deep FM model, a LambdaMart model, etc., and the specific model can be chosen based on actual needs.

[0101] The correspondence between the above videos was pre-built offline, including the correspondence between each video and its related videos.

[0102] The electronic device maintains a historical search record table, which may include the correspondence between a first preset tag and the number of clicks for each video. Therefore, in step S201, the first preset tag can be extracted from the historical search record table, and the number of clicks for each video corresponding to the first preset tag can be obtained. The number of clicks for each video corresponding to the first preset tag is the number of times the video was clicked in the search results obtained when searching with the first preset tag.

[0103] For example, assuming the first preset tag is "Actor A", the search results obtained by searching using the first preset tag include: "Video 1", "Video 2", and "Video 3". "Video 1" is a TV series starring "Actor A", "Video 2" is a movie starring "Actor A", and "Actor A" makes a cameo appearance in "Video 3". Within a historical time period, the user clicked "Video 1" 13 times, "Video 2" 11 times, and "Video 3" 1 time. Therefore, the number of clicks for "Actor A" in "Video 1" is 13, the number of clicks for "Actor A" in "Video 2" is 11, and the number of clicks for "Actor A" in "Video 3" is 1.

[0104] It is understood that the number of videos corresponding to the first preset tag in this embodiment is variable, specifically determined by the search results obtained when searching with the first preset tag. That is, the number of videos corresponding to the first preset tag is equal to the number of videos included in the search results. Consequently, the number of clicks corresponding to the first preset tag is also variable. For example, assuming the first preset tag is "Actor A," and a search for "Actor A" yields "Video 1," "Video 2," and "Video 3," the number of clicks corresponding to the first preset tag can be one of three things: "Video 1" corresponding to "Actor A," "Video 2" corresponding to "Actor A," and "Video 3" corresponding to "Actor A." Similarly, assuming the first preset tag is "Director B," and a search for "Director B" yields "Video 4," "Video 5," and "Video 6," the number of clicks can be one of three things. 、 For videos “7”, “8”, and “9”, the number of clicks corresponding to the first preset tag can be one of six possibilities: “Video 4” corresponds to the number of clicks for “Director B”, “Video 5” corresponds to the number of clicks for “Director B”, “Video 6” corresponds to the number of clicks for “Director B”, “Video 7” corresponds to the number of clicks for “Director B”, “Video 8” corresponds to the number of clicks for “Director B”, and “Video 9” corresponds to the number of clicks for “Director B”.

[0105] In step S202, the first preset tag points to the video category tag. For example, the first preset tag for a rural TV series is "rural," the first preset tag for a classic spy drama is "spy drama," the first preset tag for a suspenseful crime-solving film is "suspense" and "crime-solving," and the first preset tag for an action movie featuring actor A is "action" and "actor A." As a specific implementation, video categories can be preset, and the first preset tag can be matched with each video category. The aforementioned first preset tag may include a type tag, actor tag, era tag, character tag, and region tag, etc. The era tag indicates the era in which the video content is set, such as ancient times, the Spring and Autumn and Warring States periods, the Three Kingdoms period, the Northern and Southern Dynasties, the Five Dynasties and Ten Kingdoms period, the Yuan Dynasty, etc.; the character tag is the character included in the video; and the region tag indicates the region where the video is set.

[0106] A video typically includes multiple primary pre-defined tags, but the relevance of each tag to the video varies. In other words, the weight of each primary pre-defined tag within the video is not the same. For example, while a suspenseful crime drama might contain primary pre-defined tags such as suspense, crime-solving, and romance, the romance-related content constitutes a very small portion of the drama. Similarly, the weight of the lead and supporting characters in a drama may differ. Current techniques typically rely on human experience to pre-determine the weight of each primary pre-defined tag within the video. However, human experience is often inaccurate, and the process is inefficient and costly due to the need for manual annotation.

[0107] In this embodiment, the weight of each video corresponding to the first preset tag can be determined based on the number of clicks for each video corresponding to the first preset tag, as well as the search frequency of all first preset tags. Since no manual intervention or labeling is required, significant labor costs are saved, and the efficiency of weight setting is improved.

[0108] In one implementation method, the weight of a video corresponding to a first preset tag is positively correlated with the number of clicks on the video corresponding to the first preset tag, and negatively correlated with the search frequency of the first preset tag. The calculation method for determining the weight of each video corresponding to the first preset tag includes:

[0109] Step S221: When searching based on the first preset tag, obtain the number of times the video is clicked in the search results and the number of clicks for each first preset tag contained in each video;

[0110] Step S222: Based on the search frequency of all first preset tags, normalize the number of clicks of each video corresponding to the first preset tag, and determine the weight of each video corresponding to the first preset tag.

[0111] The following is an exemplary description of steps S221-S222 above:

[0112] In step S221, the number of times each video was clicked in the search results obtained when searching with the first preset tag can be obtained from the above historical search record table.

[0113] For example, for video A related to doctors, there may be many clicks on the video using the first preset tags. Specifically, suppose the first preset tags included in the video are: Actor A, Actor B, Romance, Doctor. Then the click situation of the video when searching for each of the first preset tags is as follows: "All TV series starring Actor A": 780; "TV series starring Actor A": 1307; "TV series starring Actor B": 708; "TV series about doctors": 53; "Romance TV series": 106; "TV series about doctors": 3371.

[0114] Then, the number of clicks for each video corresponding to the first preset tag can be counted, and the number of clicks for the same tag in the first preset tag can be merged.

[0115] Based on the above examples, after merging the identical tags and their corresponding click frequencies in the first preset tags, we can obtain the number of clicks for video A in each of the first preset tags as follows: Actor A: 780 + 1307 = 2087, Actor B: 708; Romance: 106; Doctor: 53 + 3371 = 3424. Taking actor A as an example, 2087 is the number of clicks for actor A in video A, and it is also the number of clicks for actor A in the first preset tag corresponding to video A.

[0116] In step S222, the sum of the search frequencies of all first preset tags can be used as the denominator to normalize the click count corresponding to each first preset tag. Specifically, the click count can be normalized to a floating-point number.

[0117] Based on the above example, if the search frequency for "Actor A" is 2535, the search frequency for "Actor B" is 2082, the search frequency for "romance" is 3533, and the search frequency for "doctor" is 3804, then in TV series A, the weights of each first preset tag are: Actor A -> 0.823, Actor B -> 0.34, Romance -> 0.03, Doctor -> 0.9.

[0118] Obtain the weight of each first preset tag contained in each video within that video, that is, obtain the weight of each video corresponding to the first preset tag. Based on the above example, the weight of TV series A corresponding to the first preset tag actor A is 0.823.

[0119] As one specific implementation, an inverted index can be established for each first preset tag to record the weight of each video corresponding to the first preset tag. The inverted index is a mapping from the first preset tag to video identifiers. These video identifiers can be video names, encodings, etc.

[0120] For example, suppose the first preset tags are: Actor A, Doctor. Then the inverted index for the first preset tag: Actor A is: Video A->0.823, Video B->0.793, ... The inverted index for the first preset tag: Doctor is: Video A->0.9, Video C->0.85, Video D->0.64, ...

[0121] As mentioned above, the first preset tags can include type tags, actor tags, and tags such as role, year, and region. For the role, year, and region tags, since the weights of these tags do not change with the number of searches, their weights can be set to default weights, such as 0.1 for each.

[0122] The aforementioned character, era, and region information are all built-in information in the video descriptions; therefore, they can be directly extracted from each video description. Inverted indexes can also be created for these character, era, and region tags. For example, the inverted index for the character A tag is: Video A->0.1, Video B->0.1, Video C->0.1… The inverted index for the Spring and Autumn and Warring States periods tag is: Video D->0.1, Video E->0.1, Video F->0.1. The inverted index for the country A tag is: Video G->0.1, Video H->0.1, Video I->0.1.

[0123] By obtaining the weight of each tag contained in each of the above videos in the video, and the weight of each video corresponding to each tag in the tag, the relevant videos of each video can be determined based on the above weights.

[0124] As one specific implementation method, such as Figure 3 As shown, the first similarity between videos can be calculated through the following steps:

[0125] Step S301: Take each video as the target video, and for each target video, determine all the first preset tags corresponding to the target video as target tags;

[0126] Step S302: For each target tag, determine the videos included in the search results obtained when searching with the target tag, and use them as candidate videos corresponding to the target tags.

[0127] Step S303: For each target tag, according to the preset algorithm, calculate the similarity between the target video and each candidate video based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag.

[0128] Step S304: Determine the first similarity between each video based on the similarity between the target video and each candidate video corresponding to each target tag.

[0129] The following is an exemplary description of steps S301-S304 above:

[0130] In step S301, each video can be a video from an established correspondence, where the video information of each video is known, and each video is identified as a target video. A first preset tag can be extracted from the video information as the target tag, which may include: type tag, actor tag, role, year, and region tags, etc. Then, an inverted index can be obtained for each target tag, and the similarity between the target video and each candidate video in each inverted index can be calculated. Specifically, the similarity can be calculated between the target tag of the target video and the same tags in the candidate videos.

[0131] In step S302, the videos included in the search results obtained by searching with the target tag are used as candidate videos corresponding to the target tag.

[0132] For example, suppose target video 1 includes target tags such as: Actor A, Actor B. Search results for Actor A include: Video 1, Video 2, Video 3. Search results for Actor B include: Video 2, Video 4, Video 5. Then, the candidate videos for target video 1 corresponding to the target tag: Actor A are: Video 1, Video 2, Video 3; and the candidate videos for target video 1 corresponding to the target tag: Actor B are: Video 2, Video 4, Video 5.

[0133] In step S303, for each target label, the product of the weight of the target video corresponding to the target label and the weight of each candidate video corresponding to the target label can be used as the similarity between the label of the target video and each candidate video.

[0134] For example, taking the aforementioned target video 1 as an example, for the target tag: actor A, assuming that the weight of target video 1 corresponding to actor A is 0.5, the weight of video 1 corresponding to actor A is 0.3, the weight of video 2 corresponding to actor A is 0.2, and the weight of video 3 corresponding to actor A is 0.3, then the similarity between target video 1 and video 1 is: 0.5 × 0.3 = 0.15, the similarity between target video 1 and video 2 is: 0.5 × 0.2 = 0.1, and the similarity between target video 1 and video 2 is: 0.5 × 0.3 = 0.15. For the target tag: Actor B, assuming that the weight of target video 1 corresponding to actor B is 0.8, the weight of video 2 corresponding to actor A is 0.2, the weight of video 4 corresponding to actor A is 0.2, and the weight of video 5 corresponding to actor A is 0.1, then the similarity between target video 1 and video 2 is: 0.8 × 0.2 = 0.16, the similarity between target video 1 and video 2 is: 0.8 × 0.2 = 0.16, and the similarity between target video 1 and video 2 is: 0.8 × 0.1 = 0.08.

[0135] In step S304, the first similarity between each video is the weighted sum of the similarities between the target video and each candidate video corresponding to each target tag.

[0136] To facilitate understanding of the calculation process of the first similarity between the videos in the embodiments of this application, the first target video and the second target video are used as examples for illustration.

[0137] Assume the target tags in the first target video are: Tag 1, Tag 2, and Tag 3. The target tags in the second target video are: Tag 2, Tag 3, and Tag 4. The weight of the first target video corresponding to Tag 1 is 0.5, the weight corresponding to Tag 2 is 0.6, and the weight corresponding to Tag 3 is 0.8. The weight of the second target video corresponding to Tag 2 is 0.8, the weight corresponding to Tag 3 is 0.4, and the weight corresponding to Tag 4 is 0.5.

[0138] The first target video and the second target video have the same target labels: label 2 and label 3. Therefore, the similarity between the first target video and the second target video is: 0.6×0.8+0.8×0.4=0.8.

[0139] In one possible embodiment, different algorithms can be used to calculate similarity for different target tags. The first preset tags include: video type tags and actor tags. Therefore, step S303 can specifically be:

[0140] S3031. For each target video type tag, calculate the similarity between the target video and each of the candidate videos based on the weight of the target video corresponding to the target video type tag, the weight of each candidate video corresponding to the target video type tag, and the inverse document frequency of the target video type tag.

[0141] IDF (Inverse Document Frequency) refers to the logarithm of the number of documents in a corpus containing a particular word. Specifically, IDF = log(total number of documents in the corpus / number of distinct documents containing the word in the corpus). In this embodiment, the IDF for each type of tag is log(total occurrences of the generic term / number of occurrences of the generic term containing that tag).

[0142] As one specific implementation, the similarity of the above-mentioned type tags is: the weight of the type tag in the target video * the weight of the type tag in the candidate video * the IDF value of the type tag. Of course, it can also be multiplied by a preset coefficient, such as 0.2, 0.3, etc., and this invention does not specifically limit it.

[0143] S3032. For each target actor tag, calculate the similarity between the target video and each candidate video based on the weight of the target video corresponding to the target actor tag and the weight of each candidate video corresponding to the target actor tag.

[0144] As one specific implementation method, the actor similarity between the actor tag in the target video and the actor tag in the second candidate video can be calculated using the following formula: actor tag weight in the target video * actor tag weight in the second candidate video. Alternatively, this can be multiplied by a preset coefficient to obtain the actor similarity.

[0145] Then, the similarity between the video and each of the first candidate videos can be obtained based on the type similarity between the video and each of the first candidate videos, and the actor similarity between the video and each of the second candidate videos.

[0146] As one implementation method, the aforementioned type similarity or actor similarity can be directly used as the similarity between the candidate video and the target video.

[0147] As another implementation, since the first and second candidate videos may contain identical candidate videos, there is type similarity and actor similarity between the identical candidate videos and the target video. In this case, the type similarity and actor similarity between the video and the identical candidate videos can be weighted and summed according to the preset weights of each target tag to obtain the similarity between the video and the identical candidate videos.

[0148] The weights of the target tags mentioned above can be preset according to actual needs. For example, the weight of the type tag can be 0.6, the weight of the actor tag can be 0.3, and so on. That is, the similarity of the same candidate videos is the sum of the type tag weight * type similarity + actor tag weight * actor similarity.

[0149] As mentioned above, the target tags also include character, era, and region tags. The similarity between the candidate video and the target video corresponding to each character, era, and region tag is a default value, such as 0.1, 0.2, etc. Alternatively, weights can be pre-set for the character, era, and region tags, such as 0.1, 0.3, etc., and the similarity of the type tag, actor tag, and character, era, and region tags can be weighted to obtain the similarity between the video and each candidate video.

[0150] In one possible embodiment, the method further includes:

[0151] S103. Calculate the second similarity between videos based on the preset weights corresponding to the second preset labels for each video;

[0152] The above-mentioned determination of the correspondence between videos based on the first similarity between them includes:

[0153] The correspondence between videos is determined based on the first and second similarity scores.

[0154] In S103, as described above, some tags, such as tags for characters, eras, and regions in a video, do not have their weights change with the number of searches. Therefore, tags whose weights do not change with the number of searches are called second preset tags. The weights of the second preset tags are pre-set, and the preset weights of the second preset tags are the default weights set by those skilled in the art.

[0155] The second similarity between videos is the weighted sum of the weighted products of the same second preset labels in each video.

[0156] For example, suppose the second preset tags included in the first target video are: tag a, tag b, and tag c, and the second preset tags included in the second target video are: tag b and tag c. The preset weight of each of the second preset tags is 0.1.

[0157] Among them, the second preset labels that are the same in the first target video and the second target video are: label b and label c. Then the second similarity between the first target video and the second target video is: 0.1×0.1+0.1×0.1=0.02.

[0158] Based on the first and second similarities between the videos, the correspondence between the videos can be determined. Specifically, the first and second similarities between the videos can be weighted and summed to determine the correspondence between the videos.

[0159] For example, assuming the first similarity between the first target video and the second target video is 0.8 and the second similarity is 0.02, the sum of the first similarity and the second similarity is 0.82. Based on the sum of the first similarity and the second similarity, the correspondence between the videos is determined.

[0160] In one possible embodiment, to further improve the accuracy of video recommendations, S102 specifically includes:

[0161] S1021. Obtain the video corresponding to the main requirement video as the target related video;

[0162] In this step, a similarity threshold can be preset, and videos with a similarity greater than the target search term can be selected as the primary videos. Alternatively, the videos in the search results obtained by searching for the target search term can be sorted in descending order of similarity to the target search term, and a preset number of the top-ranked videos can be selected as the primary videos.

[0163] S1022. Sort the videos related to the target;

[0164] In this step, the videos related to the target search terms are sorted in descending order of similarity to the target search terms.

[0165] S1023. Display the sorted target-related videos.

[0166] In this step, all sorted target-related videos can be displayed on the screen in chronological order for the user to select. Alternatively, only one target-related video can be displayed at a time, and the remaining videos can be displayed sequentially after the user has viewed that video. Another option is to display a preset number of target-related videos at a time, in the order they are sorted.

[0167] like Figure 4 As shown, Figure 4 This is a schematic diagram illustrating a specific example of the video recommendation method provided in this embodiment of the invention. It can be divided into an online component and an offline component. The offline component may include the following steps:

[0168] Step S401: Extract the general search query and the click data of each video to obtain the set of [general search query, clicked video, number of clicks].

[0169] Step S402: Extract type tags and actor tags from the general search query.

[0170] Step S403: Merge type tags / actor tags for the same video and normalize the weights to obtain a set of [video, [type tag / actor tag], weight].

[0171] This involves merging the click counts corresponding to the same video type tag and actor tag, and then normalizing the click counts to obtain the video type tag weights and actor weights for each video.

[0172] Step S404: Merge type tags / actor tags for the same video, construct an inverted index of type tags and actor tags with the video, and obtain a set of [type tags / actor tags, [video, weight]].

[0173] Step S405: Extract the character, year, and region from the video metadata, and construct an inverted index from the character, year, and region to the video. This results in a set of [character, year, region, [video, default weight]].

[0174] In this embodiment, the execution order of steps S401-S404 and step S405 is not specifically limited.

[0175] Step S406: For each video, extract the inverted index containing the type tag, actor tag, role, year, and region. The videos contained in the inverted index are the candidate videos for the aforementioned videos.

[0176] Step S407: Calculate the relevance weight of the label of each candidate video in each inverted index with the same label of the above video, and perform weighted merging to obtain the relevance weight of the candidate video with the above video.

[0177] Step S408: Sort by relevance weight, take TOPN, create an index, and obtain a set of [video, [related video, weight]].

[0178] like Figure 4 As shown, the online portion may include the following steps:

[0179] Step S411: Obtain the refined search query input by the user.

[0180] Step S412: Perform a video search based on the refined search query, and determine the video of primary demand from the search results.

[0181] Step S413: Obtain the set of related videos and weights corresponding to the main required video from the offline set of [video, related video, weight].

[0182] Step S414: Sort the relevant videos and return them to the user.

[0183] Traditional video recommendation methods use information such as type tags / actor tags for videos based on primary demand to recall videos with similar tags / actor tags. These type tags / actor tags lack weight, leading to the recall of candidate videos with low relevance. As mentioned above, related technologies using user behavior or collaborative methods also cannot guarantee the relevance between videos. This invention utilizes a broad search query and the click count of the corresponding video to construct a mapping relationship between videos and tags / actors, mapping click counts to tag weights. These weights reflect, to some extent, the user's viewing interest in that tag, resulting in more accurate relevance calculations. Furthermore, information such as character, era, and region is used to assist in relevance calculation, further improving the accuracy of video recommendations.

[0184] Based on the same inventive concept, embodiments of the present invention also provide a video recommendation device, such as... Figure 5 As shown, the device includes:

[0185] The main demand video determination module 501 is used to determine the video with the highest similarity to the target search term as the main demand video based on the target search term;

[0186] The target-related video recommendation module 502 is used to obtain the video corresponding to the main demand video as the target-related video and recommend it based on the correspondence between the videos; wherein, the correspondence between the videos is determined in advance through the following steps:

[0187] For each first preset tag, the weight of each video corresponding to the first preset tag is determined based on the number of clicks of each video corresponding to the first preset tag and the search frequency of all first preset tags, wherein the number of clicks of the video corresponding to the first preset tag is the number of times the video is clicked in the search results obtained when searching with the first preset tag;

[0188] Based on the weights of each video corresponding to each target first preset label, a first similarity is calculated between each video; the first similarity is positively correlated with the weights of each video corresponding to the same first preset label among the first preset labels;

[0189] The correspondence between the videos is determined based on the first similarity between them.

[0190] In one possible embodiment, calculating the first similarity between the videos based on the weights corresponding to the first preset tags for each video includes:

[0191] Each of the aforementioned videos is taken as a target video, and for each of the aforementioned target videos, all the first preset tags corresponding to the target video are determined as target tags;

[0192] For each target tag, the videos included in the search results obtained when searching with the target tag are determined as candidate videos corresponding to the target video for the target tag.

[0193] For each target tag, according to a preset algorithm, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag;

[0194] A first similarity is determined between the target video and each candidate video corresponding to each target tag.

[0195] In one possible embodiment, the first preset tag includes: a video type tag and an actor tag;

[0196] For each target tag, according to a preset algorithm, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag, including:

[0197] For each target video type tag, the similarity between the target video and each of the candidate videos is calculated based on the weight of the target video corresponding to the target video type tag, the weight of each candidate video corresponding to the target video type tag, and the inverse document frequency of the target video type tag.

[0198] For each target actor tag, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target actor tag and the weight of each candidate video corresponding to the target actor tag.

[0199] In one possible embodiment, the device further includes:

[0200] The second similarity calculation module 503 is used to calculate the second similarity between the videos based on the preset weights of the videos corresponding to the second preset tags.

[0201] Determining the correspondence between the videos based on a first similarity among them includes:

[0202] The correspondence between the videos is determined based on the first similarity and the second similarity between the videos.

[0203] In one possible embodiment, the weight of the video corresponding to the first preset tag is positively correlated with the number of clicks on the video corresponding to the first preset tag, and negatively correlated with the search frequency of all first preset tags.

[0204] In one possible embodiment, the main demand video determination module 501 is specifically used to perform video search based on target search terms to obtain search results;

[0205] Sort the search results;

[0206] Select the first video in the search results as the primary video in your search query; or...

[0207] The video with the most clicks in the historical search records of the target search term is selected as the primary video.

[0208] In one possible embodiment, the target-related video recommendation module 502 is specifically used to obtain the video corresponding to the main demand video as the target-related video;

[0209] Sort the videos related to the target.

[0210] Displays the target-related videos in sorted order.

[0211] This invention also provides an electronic device, such as... Figure 6 As shown, it includes a processor 601, a communication interface 602, a memory 603, and a communication bus 604, wherein the processor 601, the communication interface 602, and the memory 603 communicate with each other through the communication bus 604.

[0212] Memory 603 is used to store computer programs;

[0213] When processor 601 executes a program stored in memory 603, it performs the following steps:

[0214] Based on the target search term, the video with the highest similarity to the target search term is selected as the primary demand video;

[0215] Based on the correspondence between the videos, the videos corresponding to the main demand video are obtained as target-related videos and recommended.

[0216] The correspondence between the videos is determined in advance through the following steps:

[0217] For each first preset tag, the weight of each video corresponding to the first preset tag is determined based on the number of clicks of each video corresponding to the first preset tag and the search frequency of all first preset tags, wherein the number of clicks of the video corresponding to the first preset tag is the number of times the video is clicked in the search results obtained when searching with the first preset tag;

[0218] Based on the weights of each video corresponding to each of the first preset tags, a first similarity is calculated between each video; the first similarity is positively correlated with the weights of each video corresponding to the same first preset tag among the first preset tags;

[0219] Based on the first similarity between the videos, the correspondence between the videos is determined. The communication bus mentioned in the terminal above can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc. For ease of representation, only one thick line is used in the figure, but this does not indicate that there is only one bus or one type of bus.

[0220] The communication interface is used for communication between the aforementioned terminal and other devices.

[0221] The memory may include random access memory (RAM) or non-volatile memory, such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0222] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0223] In another embodiment of the present invention, a computer-readable storage medium is also provided, wherein a computer program is stored therein, and when the computer program is executed by a processor, it implements any of the video recommendation methods described in the above embodiments.

[0224] In another embodiment of the present invention, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to execute any of the video recommendation methods described in the above embodiments.

[0225] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of the present invention are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0226] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply the existence of any such actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article, or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or elements inherent to such process, method, article, or device. In the absence of further limitations, an element defined by the phrase "comprising a ..." does not exclude the presence of other identical elements in the process, method, article, or device comprising the element.

[0227] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the embodiments of apparatus, electronic devices, storage media, and program products are basically similar to the method embodiments, so the descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0228] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of protection of the present invention.

Claims

1. A video recommendation method, characterized in that, The method includes: Based on the target search term, the video with the highest similarity to the target search term is selected as the primary demand video; Based on the correspondence between the videos, the videos corresponding to the main demand video are obtained as target-related videos and recommended. The correspondence between the videos is determined in advance through the following steps: For each first preset tag, the weight of each video corresponding to the first preset tag is determined based on the number of clicks of each video corresponding to the first preset tag and the search frequency of all first preset tags. The number of clicks of a video corresponding to the first preset tag is the number of times the video is clicked in the search results obtained when searching with the first preset tag. A video includes multiple first preset tags, and the weight of the video corresponding to the first preset tag is the weight of each first preset tag contained in the video within the video. This weight represents the relevance of each first preset tag contained in the video to the video, and the relevance of each first preset tag to the video is different. The weight of a video corresponding to the first preset tag is positively correlated with the number of clicks of the video corresponding to the first preset tag and negatively correlated with the search frequency of all first preset tags. Based on the weights of each video corresponding to each of the first preset tags, a first similarity is calculated between each video; the first similarity is positively correlated with the weights of each video corresponding to the same first preset tag among the first preset tags; Based on the first similarity between the videos, the correspondence between the videos is determined; The method further includes: calculating a second similarity between the videos based on a preset weight corresponding to a second preset tag for each video, wherein the second preset tag is a tag whose weight value in the video does not change with the number of searches; Determining the correspondence between the videos based on a first similarity among them includes: The correspondence between the videos is determined based on the first similarity and the second similarity between the videos.

2. The method according to claim 1, characterized in that, The step of calculating the first similarity between the videos based on the weights corresponding to the first preset tags of each video includes: Each of the aforementioned videos is taken as a target video, and for each of the aforementioned target videos, all the first preset tags corresponding to the target video are determined as target tags; For each target tag, determine each video included in the search results obtained when searching with the target tag, as a candidate video corresponding to the target tag; For each target tag, according to a preset algorithm, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag; A first similarity is determined between the target video and each candidate video corresponding to each target tag.

3. The method according to claim 2, characterized in that, The first preset tags include: video type tags and actor tags; For each target tag, according to a preset algorithm, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target tag and the weight of each candidate video corresponding to the target tag, including: For each target video type tag, the similarity between the target video and each of the candidate videos is calculated based on the weight of the target video corresponding to the target video type tag, the weight of each candidate video corresponding to the target video type tag, and the inverse document frequency of the target video type tag. For each target actor tag, the similarity between the target video and each candidate video is calculated based on the weight of the target video corresponding to the target actor tag and the weight of each candidate video corresponding to the target actor tag.

4. The method according to claim 1, characterized in that, The step of determining the video with the highest similarity to the target search term as the primary demand video includes: Video search results are obtained based on the target search terms; Sort the search results; Select the first video in the search results as the primary video in your search query; or... The video with the most clicks in the historical search records of the target search term is selected as the primary video.

5. The method according to claim 1, characterized in that, The step of obtaining and recommending videos corresponding to the main demand video as target-related videos includes: Obtain the video corresponding to the main demand video as the target related video; Sort the videos related to the target. Displays the target-related videos in sorted order.

6. A video recommendation device, characterized in that, The device includes: The main demand video determination module is used to determine the video with the highest similarity to the target search term as the main demand video based on the target search term; The target-related video recommendation module is used to obtain videos corresponding to the main demand video as target-related videos and recommend them based on the correspondence between the videos; wherein, the correspondence between the videos is determined in advance through the following steps: For each first preset tag, the weight of each video corresponding to the first preset tag is determined based on the number of clicks of each video corresponding to the first preset tag and the search frequency of all first preset tags. The number of clicks of a video corresponding to the first preset tag is the number of times the video is clicked in the search results obtained when searching with the first preset tag. A video includes multiple first preset tags, and the weight of the video corresponding to the first preset tag is the weight of each first preset tag contained in the video within the video. This weight represents the relevance of each first preset tag contained in the video to the video, and the relevance of each first preset tag to the video is different. The weight of a video corresponding to the first preset tag is positively correlated with the number of clicks of the video corresponding to the first preset tag and negatively correlated with the search frequency of all first preset tags. Based on the weights of each video corresponding to each of the first preset tags, a first similarity is calculated between each video; the first similarity is positively correlated with the weights of each video corresponding to the same first preset tag among the first preset tags; Based on the first similarity between the videos, the correspondence between the videos is determined; The device further includes: a second similarity calculation module, used to calculate the second similarity between the videos based on the preset weights of the videos corresponding to the second preset tags, wherein the second preset tag is a tag whose weight value in the video does not change with the number of searches; The target-related video recommendation module is specifically used to determine the correspondence between the videos based on the first similarity and the second similarity between the videos.

7. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method described in any one of claims 1-5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method described in any one of claims 1-5.

Images