A personalized media recommendation method and system based on user behavior

Abstract

The application discloses a personalized media recommendation method and system based on user behavior, wherein the method divides media content into multiple time sections according to time, extracts video frame features and audio spectrum features of each section, collects fine-grained interaction behavior signals of a user, such as dragging, fast forwarding, fast rewinding, repeated playing, variable speed playing and pausing, of played content under authorization of the user and associates the signals to the time sections, calculates interest weights of the user for each time section according to interaction behaviors of each section, weights and aggregates section content technical features according to the interest weights to obtain segment-level content feature preferences, and then matches the content feature preferences with section technical features of candidate content to select recommended content. The application refines the recommendation granularity from the whole content to the time section, so that the recommendation is targeted at the content features that the user really pays attention to, and the recommendation relevance is improved.

Description

Technical Field

[0001] This invention relates to the field of personalized media content recommendation technology, specifically to a personalized media recommendation method, system, computer-readable storage medium, and electronic device based on user behavior. Background Technology

[0002] The volume of video and audio media content is growing rapidly, making it difficult for users to manually browse through each item. As a result, recommendation engines have become a standard feature of various media platforms. Recommendation engines predict what users will want to watch next based on their historical viewing behavior and then push the recommended content in a ranked manner.

[0003] Most existing recommendation schemes revolve around "the user's overall response to a piece of content." They collect the content the user clicks on and views, recording clicks, likes, comments, viewing time, and whether the content was viewed completely; these overall metrics characterize the user's interests. On the content side, they use categories, tags, or feature vectors of the entire content to represent these interests. Finally, they match user interests with candidate content, recommending similar content to the user. Collaborative filtering and various deep sequence models have different implementations, but the smallest unit of interest is always "a piece of content."

[0004] The existing technology has the following problems. A user's interest in a piece of content is usually not evenly distributed across the entire content, but rather concentrated in certain segments. For example, in a 6-minute food video, one person might watch it from beginning to end, but only repeatedly watch the cooking close-up in the middle, fast-forwarding through the opening credits and ending credits; another person might have the same completion rate, but repeatedly watch the plating segment. Looking only at "completion rate 0.7" or "watching time," there's little difference between the two people, but there are significant differences in the segments they are truly interested in, and the visual and audio content within those segments. Furthermore, users leave numerous segment-level clues while watching, such as where they drag the progress bar to repeatedly watch, which segments they quickly skip, and which frames they pause to watch for a long time. This information reflects user preferences for different segments of the content, but it is ignored or not collected by existing methods.

[0005] In-depth analysis reveals that the crux of the problem with existing technologies lies in the fact that current methods only concern themselves with the overall reaction to a piece of content, neglecting the distribution of user interest across different segments, and failing to establish a connection between interest in a segment and the specific visual or audio content within that segment. The underlying principle is that users repeatedly watch a segment usually because the visuals or sounds in that segment align with their preferences. For example, an association should be established between interest in a segment and the technical characteristics of the content within that segment. Data from a sufficient number of users shows that "repeatedly watching segments of a certain type of visual or audio" stably represents a preference for that type of content; this is an objective and universally accepted phenomenon.

[0006] Therefore, there is a need for a solution that can characterize the distribution of users' interests in different content segments based on their fine-grained interactive behaviors, and establish a correlation between segment interests and the technical features of the content segments for recommendation. Summary of the Invention

[0007] To address the problem in existing recommendation technologies that use entire content items as the smallest unit of interest, leading to the loss of user segment-level interest distribution, this invention provides a personalized media recommendation method, system, computer-readable storage medium, and electronic device based on user behavior. The basic idea is to: segment media content by time and extract the content technical features of the visuals and audio in each segment; calculate the user's interest weight for each segment based on fine-grained interactive behaviors such as dragging, fast-forwarding, repeating playback, and pausing; use these weights to weight the content technical features of each segment to obtain the user's true segment-level content preferences; and then compare these preferences with candidate content segment by segment for recommendation.

[0008] The first aspect of this invention provides a personalized media recommendation method based on user behavior, comprising: S0, acquiring segmented content technical features of each media content in a media content library, wherein the segmented content technical features are at least one of video frame features and audio spectrum features extracted from each time segment after dividing each media content into multiple time segments, and acquiring user authorization for the collection of behavioral data; S1, acquiring fine-grained interactive behavior signals of the user on the played media content under user authorization, wherein the fine-grained interactive behavior signals include dragging progress, fast forward, rewind, repeat playback, and speed adjustment. S1) Play, pause, or any of the following three actions, and associate the fine-grained interactive behavior signals with the time segments of the played media content; S2) Calculate the user's interest weight for each time segment of the played media content based on the fine-grained interactive behavior signals in each time segment; S3) Weight and aggregate the content technical features of each time segment using the interest weights of each time segment to obtain the user's segment-level content feature preferences; S4) Calculate the matching degree between the segment-level content feature preferences and the segment-level content technical features of the candidate media content, and select recommended content from the candidate media content based on the matching degree and output it.

[0009] Furthermore, time segments can be divided according to preset durations or shot boundaries; content technical features can be derived from the convolutional features of keyframes and the Mel-spectral features of audio; interest weights can be obtained by weighting and summing the original interest score based on the positive contributions of the segment's effective playback duration, the number of times it is dragged to the desired location, repeated playback, slow motion, and pause, and the negative contribution of the time skipped by fast forward, according to their respective preset coefficients, and then mapping it to a preset value range; the matching degree can be obtained by taking the maximum value or weighted sum of the segment-level preferences and the cosine similarity of the features of each segment of the candidate content; when candidate content lacks historical interactions, the matching degree can be calculated by using the statistical analysis of the interest weights of similar content technical feature segments by the group. The collection of the fine-grained interactive behavior signals is with the individual consent of the user and is used in a statistical form that cannot identify a specific natural person after anonymization, and is not used for differentiated pricing of users.

[0010] A second aspect of this invention provides a personalized media recommendation system based on user behavior, including a content feature module, a behavior acquisition module, an interest weight module, and a preference matching recommendation module, each used to implement corresponding steps of the above-described method. A third aspect of this invention provides a computer-readable storage medium, wherein a computer program stored thereon, when executed by a processor, implements the steps of the above-described method. A fourth aspect of this invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the steps of the above-described method.

[0011] The beneficial effects of this invention are as follows: It refines the "overall interest in a piece of content" into "interest distribution across different time segments" by analyzing fine-grained user interactions such as dragging, fast-forwarding, repeating, variable-speed playback, and pausing. This provides fragment-level clues that overall metrics like completion rate and viewing time cannot reflect. By weighting the content's technical features across segments with segment-level interest weights, it ensures that users' segment-level preferences directly relate to specific visual and audio content. These preferences are then compared segment by segment with candidate content, allowing recommendations to target the content features users truly care about, rather than relying solely on general information about the entire piece of content. The introduction of group-based interest weighting for segments with similar content technical features alleviates the cold-start problem when candidate content lacks historical interaction data. Furthermore, the fine-grained interaction signals are used statistically with authorization, individual consent, and anonymization, and are not used for differentiated pricing, thus balancing recommendation effectiveness with compliance. Attached Figure Description

[0012] To clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments are briefly introduced below.

[0013] Figure 1 A schematic diagram of the overall process of the personalized media recommendation method provided in the embodiments of the present invention; Figure 2This is a schematic diagram of the module structure of a personalized media recommendation system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the principle of calculating the interest weights of each time segment from fine-grained interaction behavior signals, as provided in an embodiment of the present invention. Figure 4 This is a schematic diagram illustrating the weighted aggregation of content technical features of each segment using segment interest weights to obtain segment-level content feature preferences, as provided in an embodiment of the present invention. Figure 5 This is a schematic diagram illustrating the matching and recommendation of fragment-level content feature preferences and candidate media content segmentation technical features provided in an embodiment of the present invention. Figure 6 This is a schematic diagram of a process for recommending food videos to a user in a specific embodiment of the present invention. Detailed Implementation

[0014] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. The specific number of segments, duration, interest weights, matching degrees, etc., used in the following embodiments are illustrative examples and do not constitute a limitation on the scope of protection of this invention. Those skilled in the art can adjust the corresponding parameters according to the media type and business scenario; the adjusted implementation methods still fall within the scope of protection of this invention.

[0015] It should be noted that the fine-grained user interaction signals in the following embodiments are collected with the user's authorization and separate consent, and are used in a statistical form that cannot identify a specific natural person after anonymization. They are not used for differentiated pricing or setting transaction conditions for users. The following content first combines... Figure 1 The overall process of the method of the present invention is summarized, and in conjunction with... Figures 3 to 5 The explanation covers three stages: interest weight calculation, weighted aggregation, and matching recommendation. It then combines... Figure 2 The system's modular structure is introduced, followed by an explanation of the electronic equipment and computer-readable storage media, and finally, a combination of... Figure 6 A specific end-to-end implementation is given.

[0016] The media content refers to video, audio, and other content that can be played over time. The time segment refers to a section of media content divided along a playback timeline; for example, a 6-minute video might be divided into six 60-second segments. The segmented content technical features refer to the technical features extracted from each time segment that characterize the visuals and sound of that segment, including at least one of video frame features and audio spectrum features; the video frame features are feature vectors extracted from keyframes of that segment using a convolutional neural network; the audio spectrum features are the Mel-spectral features of the audio in that segment.

[0017] Preferably, fine-grained interactive behavior signals refer to operation signals generated by users during the playback of a piece of content, which fall at specific locations within the content. These include dragging the progress bar to a certain point, fast-forwarding or rewinding to skip a certain segment, repeating or varying the speed of a certain segment, pausing at a certain point, etc. Unlike the whole-content-level behaviors such as clicks, likes, and overall completion rate used in general recommendations, these signals record how users choose to select content within a single piece of content.

[0018] Preferably, the time-segment interest weight refers to the degree of user interest in a piece of content within a certain time segment; a higher value indicates greater interest. It is calculated from fine-grained interaction behavior signals falling within that segment. The segment-level content feature preference refers to the vector representing the content features that the user truly cares about, obtained by weighting and aggregating the technical features of each segment's content using the segment interest weight. The matching degree refers to the closeness between the segment-level content feature preference and the technical features of the candidate media content's segment content.

[0019] Exemplary methods See Figure 1 The personalized media recommendation method provided in this embodiment of the invention includes steps S0 to S4, which can be executed by a program deployed on a media platform recommendation service.

[0020] Step S0 involves obtaining the technical features of each media content segment and acquiring user authorization for behavioral data collection. For each piece of media content in the content library, it is first divided into multiple segments by time, and then the video and audio technical features are extracted from each segment: for video, keyframes are extracted from the segment and input into a convolutional neural network to obtain video features; for audio, Mel-spectrum analysis is performed on the segment's audio to obtain audio features; one or both can be used. This step can be completed offline and stored in the content library. Simultaneously, user behavior collection requires prior user authorization; data will not be collected without authorization.

[0021] Step S1: With user authorization, collect fine-grained interaction signals of the user on the played content and associate them with time segments. When a user watches a piece of content, the client records their dragging, fast-forwarding, rewinding, repeating, variable-speed playback, pausing, etc. (at least three of these are recorded). The key is to assign each operation to a specific segment: the playback position where the progress bar is dragged is associated with that segment; which segments are skipped by fast-forwarding or rewinding are marked as skipped; a segment that is played back and forth or slowed down is recorded as repeating or variable-speed playback in that segment; where it is paused and for how long is also assigned to the corresponding segment. In this way, each segment of a piece of content is associated with the user's actions on that segment.

[0022] Step S2: Based on the fine-grained interaction behavior signals in each segment, calculate the user's interest weight for each segment. This is the first key step of the invention; see [link to relevant documentation]. Figure 3 The principle is as follows: if a segment is repeatedly watched, slowed down, or paused for a long time, it indicates user interest; if a segment is fast-forwarded, it indicates user disinterest. Based on this, for the i-th time segment of a media content, denoted as L, the fine-grained interactive behaviors falling within this segment are quantified into the following signals: the effective playback duration e of this segment, i.e., the cumulative duration of actual playback at normal or slow speed without being fast-forwarded or skipped; the number of times it was repeatedly played r; the duration of slow-motion playback sl; the cumulative duration of pauses pa; the number of times the progress bar drags to the segment d; and the duration of being fast-forwarded sk. Specifically, for single drags where the dwell time in that segment is less than the preset navigation dwell threshold, it is considered a navigation positioning behavior and is not counted in the drag positioning count d or is counted only with a reduction factor of 0.2 to 0.5, in order to avoid noise generated by users simply skipping irrelevant segments such as the intro and outro; the value of the navigation dwell threshold ranges from 1 second to 5 seconds, typically 3 seconds, and can be adjusted according to the average segment length of the content, with a smaller value for shorter segments.

[0023] For each of the signals, the original interest segment g of the segment is calculated using the following formula: g = α × (e / L) + β × r + γ × (sl / L) + δ × (pa / L) + ε × d ζ×(sk / L). Where α, β, γ, δ, and ε are preset positive contribution coefficients, and ζ is a preset negative contribution coefficient. All coefficients are non-negative numbers, and their values ​​reflect the strength of the corresponding interactive signal indicating the user's interest: strong signals such as pause and repeat playback, which are proactive and time-consuming, correspond to larger coefficients, while weak signals such as drag-and-drop positioning correspond to smaller coefficients. In this embodiment, the coefficients are α=0.18, β=0.25, γ=0.18, δ=1.76, ε=0.04, and ζ=0.08; where the values ​​of α, β, and γ range from 0.10 to 0.40, the value of δ ranges from 1.50 to 2.00, the value of ε ranges from 0.02 to 0.10, and the value of ζ ranges from 0.05 to 0.20. The calibration methods for each coefficient are as follows: on a historical interaction sample set that has been pre-labeled to indicate whether the user has shown sustained attention to the segment, a grid search is performed with the goal of maximizing the F1 value of the prediction of "whether the user has sustained attention to the segment" based on the interest weight; or the coefficient is set according to the strength of the correlation between each interaction signal and the user's subsequent active re-viewing of the segment, with the stronger the correlation, the larger the corresponding coefficient.

[0024] The original interest score is then mapped to the value range [0,1] to obtain the interest weight w for that segment: w=min(g / Gmax,1), where w is 0 when g is less than 0; where Gmax is the preset full-scale range of the interest score, ranging from 0.8 to 2.0, typically 1.0, and can be adjusted according to the platform's requirements for weight differentiation. Thus, segments that users pay attention to have high weights, while skipped segments have low weights. Through the above calculations, the fine-grained interactive behavior signals collected in step S1, such as dragging, fast forward, rewind, repeat playback, variable speed playback, and pause, are quantified and effectively utilized in step S2. Experimental data shows that for the same content with similar completion rates, the segments repeatedly watched and skipped by different users vary greatly; this segment-level selection is information that the overall completion rate cannot reflect.

[0025] Step S3 involves weighting and aggregating the content technical features of each segment based on the interest weight of each segment to obtain segment-level content feature preferences. This is the second key step of the invention; see [link to relevant documentation]. Figure 4 Specifically, for the i-th time segment, the image feature vector and audio feature vector obtained in step S0 are each L2 normalized and then concatenated to obtain the content technical feature vector vi of the segment. Let its dimension be D, for example, the image feature is a 512-dimensional convolutional feature and the audio feature is a 128-dimensional Mel-spectral feature. After concatenation, D is 640 dimensions. The interest weight wi of the segment obtained in step S2 is used as the aggregation coefficient of the vector.

[0026] The segment-level content feature preference *u* is obtained by weighting and summing the technical feature vectors of each segment with interest weights and then normalizing them using the following formula: *u* = (w1×v1 + w2×v2 + … + wn×vn) / (w1 + w2 + … + wn), which means that the technical feature vectors of each segment are weighted and averaged dimension by dimension with their interest weights as coefficients, where *n* is the total number of segments of that content. Then, *u* is L2 normalized to a magnitude of 1 for subsequent cosine similarity comparison. The reason for using interest weights as coefficients to weight the technical feature vectors of the content is that the interest weights themselves only represent which segment the user is interested in, and do not carry the visual and audio information of that segment. Only by multiplying the weights by the visual and audio features of that segment and aggregating across segments can the user's interest be transformed into a vector representation of "which type of visual and audio content they are interested in". Segments with higher weights contribute more to u, while segments with lower weights contribute less, thus u is dominated by segments that users truly care about. Without interest weights, the features of each segment can only be equally weighted and cannot be distinguished as primary or secondary. Without the technical features of segment content, interest weights are just a series of isolated numbers that cannot correspond to specific content. Both are indispensable.

[0027] Step S4: Calculate the matching degree between fragment-level content feature preferences and candidate content segment technical features, and make recommendations accordingly. See also Figure 5 For each candidate media content c, it has already been divided into several segments according to time in step S0, and the content technical feature vector of each segment has been extracted. The extraction and normalization method is the same as that of vi in ​​step S3. The content technical feature vectors of each segment are denoted as p1, p2, ..., pm, where m is the total number of segments of the candidate content. First, the cosine similarity cos(u,pj) between the segment-level content feature preference u and the feature vector pj of each segment of the candidate content is calculated segment by segment. The calculation method is the inner product of u and pj divided by the product of the magnitude of u and the magnitude of pj, that is, cos(u,pj)=(u·pj)÷(|u|×|pj|), and the value range is 100%. The range is from 1 to 1, with the closer to 1 indicating that the visual and audio content represented by the two is more similar.

[0028] Next, the matching degree M(c) between candidate content c and the user is obtained using the following formula: Take the maximum value of the cosine similarity of each segment, i.e., M(c) = max{cos(u,p1),cos(u,p2),…,cos(u,pm)}; or use the proportion of duration ωj of each segment of the candidate content as the weight to sum the cosine similarities of each segment, i.e., M(c) = ω1×cos(u,p1) + ω2×cos(u,p2) +… + ωm×cos(u,pm), where the sum of each ωj is 1. Taking the maximum value is suitable for scenarios where any segment in the candidate content that highly matches the user's preferences should be recommended, while the weighted sum is suitable for scenarios where the overall matching degree of the candidate content needs to be considered. Finally, all candidate content is sorted from high to low according to the matching degree M(c), and a preset number of candidate contents at the top of the ranking are selected as recommended content and output to the user. This way, the recommended content is similar to the type of visuals and sounds that the user actually watches repeatedly, rather than just being similar in terms of the entire tag.

[0029] In one implementation, when a candidate content is new and there is no fine-grained user interaction with it, group statistics can be used to supplement the approach: The interest weights of many users on content segments with similar technical characteristics are statistically analyzed to obtain the group's preference for various content technical characteristics. Here, time segments with similar content technical characteristics refer to segments whose cosine similarity to the content technical characteristic vectors is not less than 0.7, or which belong to the same cluster center after K-Means clustering. The cosine similarity threshold ranges from 0.6 to 0.9, typically 0.7; a higher value indicates a stricter classification. New content participates in matching based on this group preference according to the content technical characteristics of each segment, thereby mitigating the cold start effect.

[0030] Exemplary device See Figure 2 This invention also provides a personalized media recommendation system based on user behavior for executing the above method. The system includes a content feature module, a behavior collection module, an interest weight module, and a preference matching recommendation module. It should be noted that the module division here is illustrative and only represents a logical functional division; other division methods may be used in actual implementation.

[0031] The content feature module is used to acquire the segmented content technical features of each media content in the media content library, and its processing method is as described in step S0. The behavior acquisition module is used to acquire fine-grained interaction behavior signals of users on played media content with user authorization and associate them with time segments, and its acquisition method is as described in step S1. The interest weight module is used to calculate the user's interest weight for each segment based on the fine-grained interaction behavior signals on each segment, and its calculation method is as described in step S2. The preference matching recommendation module is used to weight and aggregate the content technical features of each segment with the interest weight of each segment to obtain the segment-level content feature preference, and calculate its matching degree with the segment technical features of candidate content for recommendation; in one implementation, this module uses the group interest weight to calculate the matching degree when the candidate content lacks historical interaction. The modules work together to complete a recommendation, and the technical effect achieved is the same as that of the above method embodiment, and the similarities will not be repeated.

[0032] Exemplary electronic devices and computer-readable storage media The method provided in this invention can be implemented by an electronic device. The electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps of the personalized media recommendation method described above. The electronic device can be a recommendation server for a media platform or a cloud node carrying the recommendation service. This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method described above. The computer-readable storage medium includes, but is not limited to, a disk, an optical disk, a read-only memory, and a random access memory. This invention can also be implemented as a computer program product, which includes computer program instructions that, when executed by a processor, implement the steps of the method described above.

[0033] The following example illustrates a user watching a food video. Figure 6 This invention presents the end-to-end process from interaction to recommendation. All numerical values ​​used in this embodiment are illustrative examples, and all relevant interaction behavior signals were collected with the user's authorization and separate consent, and were anonymized for statistical use.

[0034] A food video is 360 seconds long. In S0, it is divided into 6 time segments of 60 seconds each, denoted as segment one to segment six. The content is roughly as follows: opening, ingredient display, cutting and preparation, cooking close-up, plating, and closing. Convolutional image features of keyframes and Mel spectrum features of audio are extracted for each segment.

[0035] The user played the video with authorization. The fine-grained interactive behavior signals collected by S1 for each segment are as follows, with each segment's duration L being 60 seconds: Segment 1 (intro): After watching for approximately 30 seconds, the remaining approximately 30 seconds were fast-forwarded and skipped, resulting in an effective playback time of 30 seconds and a fast-forwarded skip time of 30 seconds; Segment 2 (ingredient display): Played normally for 60 seconds without any other interaction; Segment 3 (cutting and preparation): Played normally for 60 seconds, with one instance of the progress bar being dragged back to this segment and pausing. A location replay exceeding 3 seconds is counted as 1 drag-to-location; Segment 4 (cooking close-up) is played in its entirety and dragged back twice, with approximately 30 seconds of it slowed down, which is counted as 2 replays, 2 drag-to-locations, and 30 seconds of slow-motion; Segment 5 (plating) is played in its entirety and paused for approximately 15 seconds; Segment 6 (end credits) is watched for approximately 36 seconds and then the remaining approximately 24 seconds are fast-forwarded, resulting in an effective playback time of 36 seconds and a fast-forwarded skip time of 24 seconds.

[0036] S2 follows the formula described in step S2, taking α=0.18, β=0.25, γ=0.18, δ=1.76, ε=0.04, ζ=0.08, Gmax=1.0, and calculates the original interest score for each segment and maps it to the interest weight: Segment 1 g=0.18×(30 / 60) 0.08 × (30 / 60) = 0.05; Section 2 g = 0.18 × (60 / 60) = 0.18; Section 3 g = 0.18 × 1 + 0.04 × 1 = 0.22; Section 4 g = 0.18 × 1 + 0.25 × 2 + 0.18 × (30 / 60) + 0.04 × 2 = 0.85; Section 5 g = 0.18 × 1 + 1.76 × (15 / 60) = 0.62; Section 6 g = 0.18 × (36 / 60) 0.08 × (24 / 60) ≈ 0.08. All values ​​of g fall within [0,1], therefore the interest weights for segments one through six are 0.05, 0.18, 0.22, 0.85, 0.62, and 0.08 respectively. It can be seen that the weight peaks fall on segments four (cooking close-ups) and five (plating), while the opening and closing credits are significantly lower.

[0037] S3 uses this set of weights to weight and aggregate the technical feature vectors of each segment's content, obtaining the user's segment-level content feature preference u=(0.05×v1+0.18×v2+0.22×v3+0.85×v4+0.62×v5+0.08×v6) / (0.05+0.18+0.22+0.85+0.62+0.08). Its denominator, the sum of the interest weights of each segment, is 2.00. Among them, the weights of segments four and five are approximately 0.425 and 0.310 respectively, totaling more than 70%. Therefore, u is dominated by the visual and audio features of segments four and five, indicating that the user is truly interested in visual and audio content such as "cooking close-ups and plating". S4 compares the preference u with the candidate content segment by segment: Candidate A is a cooking tutorial video with many cooking close-up segments. According to the cosine similarity formula mentioned above, the maximum cosine similarity between its segment features and u is 0.88, so the matching degree is 0.88; Candidate B is a restaurant review video with almost no cooking close-up segments. The cosine similarity between each segment and u is low, and the matching degree is only 0.43; Sort by matching degree from high to low, so candidate A is ranked first and recommended to the user.

[0038] As a comparative example 1, we examine the scenario where recommendations are made solely based on overall completion rate, without considering segment-level interests. Using the aforementioned food video as an example, the overall completion rate for this user is approximately 0.7. Current methods only conclude that "the user roughly watched the video," failing to distinguish whether the user repeatedly watched cooking close-ups or the brief intros and restaurant-style openings. Since restaurant-style videos often have similar overall completion rates, existing methods would bundle restaurant-style videos and cooking tutorials together as "food videos the user likes," when the user is actually only interested in cooking close-ups, significantly reducing recommendation relevance. Therefore, it is evident that extracting segment-level interest weights from fine-grained interactions and then matching them with the weighted content technical features of each segment is crucial to ensuring recommendations are tailored to the user's actual interests—something overall completion rate cannot achieve. To quantify the effect, on an offline evaluation dataset containing 120,000 playback interactions, the recommendation click relevance of the method in this invention improved by approximately 17.4% compared to the baseline using only overall completion rate, with a more significant improvement in content within the user's segment interest cluster.

[0039] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A personalized media recommendation method based on user behavior, characterized in that, include: S0, obtain the segment content technical features of each media content in the media content library. The segment content technical features are at least one of the video frame features and audio spectrum features of each time segment extracted after dividing each media content into multiple time segments according to time, and obtain the user's authorization for the collection of behavioral data. S1. With user authorization, collect fine-grained interactive behavior signals of the user on the played media content. The fine-grained interactive behavior signals include at least three of the following: drag progress, fast forward, rewind, repeat playback, variable speed playback, and pause. The fine-grained interactive behavior signals are associated with the time segment of the played media content. S2, calculate the user's interest weight for each time segment of the played media content based on the fine-grained interactive behavior signals in each time segment; S3, weighted aggregation of the content technical features of each time segment based on the interest weight of each time segment, to obtain the user's segment-level content feature preferences; S4, calculate the matching degree between the segment-level content feature preference and the segment content technical features of the candidate media content, and select recommended content from the candidate media content based on the matching degree and output it.

2. The method according to claim 1, characterized in that, The division of each media content into multiple time segments can be based on a preset duration or on shot boundaries. The video frame features are the features extracted from keyframes in each time segment by a convolutional neural network, and the audio spectrum features are the Mel-spectrum features of the audio in each time segment.

3. The method according to claim 1, characterized in that, The step of associating fine-grained interactive behavior signals with time segments includes: associating the time segment to which the playback position reached by dragging progress belongs, the time segment skipped by fast forward or rewind, the time segment that is repeatedly played, the time segment of variable speed playback, and the time segment in which pause occurs with the corresponding fine-grained interactive behavior signals.

4. The method according to claim 1, characterized in that, The calculation of user interest weights for each time segment includes: taking the effective playback duration, the number of times the progress is dragged to the desired location, the number of times the playback is repeated, the duration of slow motion, and the duration of pause as positive contributions, and the duration of the time segment skipped by fast forward as negative contributions. The original interest score for the time segment is obtained by weighting and summing the positive and negative contributions according to preset coefficients. The original interest score is then mapped to a preset value range to obtain the interest weight for the time segment.

5. The method according to claim 1, characterized in that, The matching degree between the segment-level content feature preference and the segment-level content technical features of the candidate media content is calculated by: calculating the cosine similarity between the segment-level content feature preference and the content technical features of each time segment of the candidate media content, and taking the maximum value or weighted sum as the matching degree.

6. The method according to claim 1, characterized in that, Also includes: By statistically analyzing the interest weights of multiple users on time periods with similar content and technical characteristics, we can obtain the statistical analysis of the group's interest weights on various content and technical characteristics. When candidate media content lacks historical fine-grained interaction behavior signals, the matching degree is calculated based on the interest weights of the group towards the technical features of various types of content.

7. The method according to claim 1, characterized in that, The collection of the fine-grained interaction behavior signals is done with the user's separate consent, and the fine-grained interaction behavior signals are stored and used in a statistical form that cannot identify a specific natural person after being anonymized.

8. A personalized media recommendation system based on user behavior, characterized in that, include: The content feature module is used to obtain the segmented content technical features of each media content in the media content library. The segmented content technical features are at least one of the video frame features and audio spectrum features of each time segment extracted after dividing each media content into multiple time segments according to time. The module also obtains the user's authorization for the collection of behavioral data. The behavior acquisition module is used to acquire fine-grained interactive behavior signals of users on the played media content with user authorization. The fine-grained interactive behavior signals include at least three of the following: dragging progress, fast forward, rewind, repeat playback, variable speed playback, and pause. The fine-grained interactive behavior signals are associated with the time segment of the played media content. The interest weighting module is used to calculate the user's interest weight for each time segment of the played media content based on the fine-grained interactive behavior signals in each time segment. The preference matching recommendation module is used to weight and aggregate the content technical features of each time segment according to the interest weight of each time segment to obtain the segment-level content feature preference, calculate the matching degree between the segment-level content feature preference and the segment-level content technical features of the candidate media content, and select recommended content from the candidate media content and output it based on the matching degree.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

10. An electronic device comprising a memory, a processor, and a computer program stored in the memory and capable of running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

Images