Video recommendation

The explosive growth of video platforms can be attributed to their ability to recommend captivating content to users [22]. Within this context, accurately predicting watch time, a crucial metric that reflects user engagement, holds significant importance for recommender systems in various industries [23, 24]. While there has been considerable research focused on metrics like Click-Through-Rate (CTR) [25–29], the area of watch-time prediction has received comparatively less attention. Recent advancements in video recommendations have demonstrated remarkable performance by mining users’ preferences and leveraging historical preference data to forecast future videos, thereby assessing the effectiveness of recommender systems.

To address the challenge of predicting users’ video preferences, several innovative approaches have emerged [30, 31]. Dynamic Micro Video Recommendations (DMA) [32] propose an explicit modeling technique that captures the dynamic trends in users’ current preferences, incorporating both historical data and potential future trends. The task of recommending the next video to watch has long been a focal point in recommender system research. Markov Chain-based Transition Probability matrix [33] has been introduced as an efficient method to uncover individual behavior preferences. Additionally, the groundbreaking study of Multiscale Time Aware User Interest (MTIN) [34] introduces interest groups based on users’ interaction sequences, providing novel insights for video recommendation. In terms of video lifespan and streaming patterns, CONDE (Concept-Aware Graph Neural Network) [35] presents a concept-driven approach to representing user preferences in video recommendations. Going beyond click and rate-based approaches, Social4Rec [36] enhances the representation of user interests by incorporating social factors such as friendships, following bloggers, and interest groups. Finally, SEMI (Sequential Multi-model Information Transfer Network) [37] utilizes user behavior in e-commerce environments to enhance video recommendations, particularly in the context of purchasing interactions.

This section emphasizes the critical role of accurate watch-time prediction in video recommendations. The advancements in dynamic modeling, interest group analysis, concept-aware representations, and the integration of social aspects contribute to the ongoing efforts to improve the effectiveness and user experience of recommender systems in the domain of video recommendations.

Behavior modeling in RecSys

Behavior modeling in the realm of Recommender Systems (RS) encompasses the task of unveiling sequential patterns embedded within users’ historical sequences, enabling the prediction of subsequent items based on these patterns [38]. With a keen eye on optimizing long-term user satisfaction, Markov Decision Process (MDP) approaches [39] have emerged to encapsulate user satisfaction rewards, incorporating nuanced heuristics that consider both user stickiness and activeness. Retention, an invaluable metric reflecting prolonged user-system interactions, has come into sharp focus within the RS landscape. Remarkably, request-based MDP [40] has revolutionized reinforcement learning techniques, empowering the pursuit of maximal retention and long-term performance optimization.

Notably, RS research has experienced a paradigm shift from short-term engagement optimization to a resolute dedication to enhancing the long-term user experience [41, 42]. This transformation has witnessed the advent of surrogate selection techniques [43], forging connections between long-term outcomes and more immediate behavioral signals, thus enabling the fine-tuning of immediate-term behaviors. Another noteworthy study, PreRec [44], has laid the foundation for recommender systems to leverage insights gained from users’ historical behaviors, thereby optimizing long-term user engagement within the realm of recommendations.

Moreover, within the domain of online RS, a pioneering approach known as Long-Short term Temporal Meta-learning (LSTTM) has emerged [45], focused on unraveling users’ intricate internal and external behaviors and preferences, thus providing valuable insights for recommendations. Exploiting the power of long-term sequence models, revolutionary methods such as DREAM [46], SASRec [47], and BST [48] have entered the fray. Dynamic REcurrent BAsket Model (DREAM) astutely captures global sequential features that interlink items, while the self-Attention-based Sequential model (SASRec) harnesses the prowess of self-attention-based sequential models. Lastly. Behavior Sequence Transformer (BST) meticulously deciphers and harnesses users’ long-term interests through the Behavior Sequence Transformer.

Data collection and preprocessing

In this study, the dataset was collected from the Saudi Telecom Company (STC) [49] and comprises video-based data with N = 3, 598, 607 rows and d = 13 columns, including features such as user ID, watch-time duration, genres, date, etc. The STC’s new Open Data project for the general public, known as “IPTV,” and its streaming platform for movies and television shows, JAWWY, provided the sources for this dataset. The JAWWY contains over 3 million records of user behavior activity, making it a valuable resource for analyzing user behavior. Thus, the video-based data can be considered as an N × d matrix , where each row of X corresponds to a user and each column corresponds to a feature. In this study, X was used to analyze user behavior and identify patterns for recommendation systems.

From the perspective of user behavior analysis, this study used the STC dataset to identify long-term behavioral patterns for recommendation systems. To achieve this, we conducted statistical analysis and machine learning algorithms to model and predict user behavior based on various features. The video-streaming dataset used in this study is a collection of data points or observations that represent user behavior on a streaming platform. This dataset is composed of two main parts, as shown in Table 1 movies data and series data.

Download:

• PNG
  larger image
• TIFF
  original image

Table 1. Statistics of the datasets.

https://doi.org/10.1371/journal.pone.0312520.t001

Data cleaning.

Data cleaning is a fundamental process in data preprocessing [50–52], which involves the identification and management of incomplete, inaccurate, duplicated, null, or irrelevant data. Suppose we have an original dataset containing n observations and p features. The aim of data cleaning is to obtain a cleaned dataset , where n_c ≤ n and p_c ≤ p.

The preprocessing stage of data analysis is crucial for identifying and handling outliers, as it can yield valuable insights into the data. This research focused on analyzing the maximum number of users who were outliers in the dataset. The Jawwy service is similar to Netflix, and while customers watch a large number of videos for entertainment, a watch-time duration of 11944 is an abnormally high value. To account for this, the research considered two potential explanations for the outliers:

• The presence of duplicate records for the same user in the database.
• The absence of essential features in the service that help users limit their viewing time.

To identify and handle outliers, as shown in Fig 2, Inter-Quartile Range (IQR) was used as a preprocessing method to manage the outliers [53, 54]. We define the set of outliers as O = x ∈ X|x > Q3 + 1.5 × IQR or x < Q1 − 1.5 × IQR, where Q1 and Q3 are the first and third quartiles, respectively, and IQR = Q3 − Q1 is the interquartile range.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Logarithmic scale visualization and the impact of data preprocessing on watch-time duration in video-streaming data.

https://doi.org/10.1371/journal.pone.0312520.g002

Data transformation.

Following the completion of data cleaning, a range of data transformation methods were implemented to improve the quality and usability of the dataset [56].

Encoded transformation. In data processing, it is often necessary to convert categorical data into numerical data in order to use it in various analytical and modeling tasks [57]. One-hot encoding is a technique used to perform this conversion [58]. Categorical data represents qualitative variables with a finite number of categories, such as genres (animated, action, drama). However, many machine learning algorithms require numerical input data to operate on. One-hot encoding is a way to represent categorical data numerically without introducing any ordinal relationship between the categories.

Suppose we have a categorical variable C with k categories, represented by the set C = {c₁, c₂, …, c_k}. To transform this categorical variable into a numerical format, we can use a one-hot encoding. One-hot encoding creates k binary variables, one for each category, and assigns a value of 1 to the variable corresponding to the category and a value of 0 to all other variables. The resulting dataset has k columns, with each column representing one category.

The one-hot encoding of a categorical variable C can be defined as follows:

Let x be a categorical variable with k categories, represented by the set C = {c₁, c₂, …, c_k}. The hot-one encoding of x is a matrix X of size n × k, where n is the number of observations in the dataset, such that each element x_ij of the matrix is defined as: (2)

For example, suppose we have a dataset with a categorical variable “Genres” that can take on three values: “Action”, “Horror”, or “Drama”. The one-hot encoding of this variable would create three binary variables, one for each genre. The resulting dataset would have three columns, with each column representing one genre as shows in Table 2.

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Data transformation using one-hot encoder.

https://doi.org/10.1371/journal.pone.0312520.t002

Scaling transformation. In the preprocessing of a dataset, it is essential to perform scaling transformations, which involve applying scaling techniques to the features. The goal of scaling is to ensure that all the features are measured on the same scale to avoid issues of one feature dominating over another [59]. Min-max scaler is a data normalization technique used to transform features by scaling them to a specified range of values. Let’s consider a feature vector X = {x₁, x₂, ⋯, x_n} with n elements. The min-max scaler maps each element x_i to a new value in the range [a, b], where a and b are the minimum and maximum values of the range [0, 1], respectively.

The transformation is performed as follows:

The purpose of min-max scaling is to normalize the feature vector so that each element has equal weight and is on the same scale. It is commonly used in machine learning algorithms that require the features to be on a common scale, such as decision trees, random forests, etc. The example of min-max scaling is shown in Table 3.

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Example of min-max scaling on multiple features.

https://doi.org/10.1371/journal.pone.0312520.t003

Consider a dataset consisting of two features, denoted by X₁ and X₂, and assume that the dataset has been provided with values. The objective is to apply min-max scaling to the dataset to standardize the values of both features to the same scale. The first step involves computing the minimum and maximum values of each feature. Subsequently, the min-max scaling formula is applied to each feature for each sample. (3)

Analysis of user behavior patterns for data classification & clustering

In recommendation systems, user behavior analysis plays a vital role as it allows for a deeper understanding of user preferences, and interests. This information can be represented as a set of vectors denoted by X = {x₁, x₂, ⋯, x_n}, where each vector x_i represents the behavior pattern of user i. Classification and clustering are two common techniques used to analyze user behavior patterns in recommendation systems.

Exploring user behavior patterns for data classification and model implementation

The prediction of user behavior patterns constitutes a vital endeavor in which one endeavors to discern the proclivities and inclinations of individuals through an examination of their historical behavioral data [61]. This intricate procedure necessitates a thorough dissection of these behavior patterns, with the ultimate goal of offering tailor-made recommendations and enhancing user contentment.

The prediction of user behavior patterns can be formalized as follows. Let denote the user-behavior matrix, where each row corresponds to a user and each column corresponds to a behavior. Let denote the target variable, which is a binary indicator of whether a user is interested in a particular item or not. Let denote the feature matrix of the users, where each row corresponds to a user and each column corresponds to a user attribute, such as genres. The task is to learn a function that can predict the target variable y given the feature vector of a new user.

An approach for acquiring this function involves the utilization of a decision tree, as outlined in [62]. A decision tree partitions the feature space into a set of rectangular regions and assigns a label to each region. The construction of this tree is an iterative process, wherein, at each juncture, the system selects the feature that yields the highest information gain, proceeding until it fulfills a predefined termination criterion. The resultant decision tree can then be applied to classify novel users by tracing a trajectory from the tree’s root to a leaf node. This leaf node corresponds to the label of the segment encompassing the user’s feature vector.

The decision tree algorithm can be formalized as follows. Given a set of training data (xi, y_i){i = 1}^N, where and y_i ∈ {0, 1}, the algorithm learns a tree T that minimizes the following cost function: (9)

Common loss functions include the logistic loss for binary classification: (10) where is the predicted probability of class 1. Once the decision tree is trained, it can be used to predict the behavior patterns of new users by following the path from the root to a leaf node, which corresponds to the predicted label. The decision tree [63] can also be used to identify the most important features that contribute to the classification of users by computing the information gained at each split.

Exploring user behavior patterns for data clustering

In the realm of data clustering, forecasting user behavior patterns is a pivotal task, entailing the categorization of users into distinct clusters predicated on their historical behavioral data, as elucidated in [64]. This intricate procedure necessitates a meticulous examination of user behavior patterns, geared towards furnishing personalized recommendations and enhancing user contentment. A frequently employed clustering technique is the renowned k-means algorithm.

K-means is an unsupervised machine learning algorithm [65] employed to cluster similar data points together. In the context of user behavior prediction, K-means can be utilized to group users based on their behavior patterns. In the k-means algorithm, the distance criterion is typically set to the Euclidean distance between data points and cluster centroids [66]. The Euclidean distance [67] is a measure of the straight-line distance between two points in Euclidean space, which corresponds to the shortest path between them. (11)

In the assignment step, each data point is assigned to the nearest cluster centroid based on the Euclidean distance [68]. The distance between each data point and each cluster centroid is computed, and the data point is assigned to the cluster with the closest centroid.

In the update step, the cluster centroids are recalculated based on the mean of the data points assigned to each cluster. The centroid of each cluster is set to the mean of the data points in that cluster, which minimizes the total within-cluster sum of squares (WCSS). By using the Euclidean distance [69] as the distance criterion, the k-means algorithm seeks to minimize the total distance between data points and their respective cluster centroids, resulting in compact and well-separated clusters.

In k-means, K refers to the number of clusters that are desired. This parameter is usually determined by the analyst based on the characteristics of the data and the desired number of clusters. The choice of K can have a significant impact on the clustering results, as too few clusters may not capture all the relevant behavior patterns, while too many clusters may result in overfitting and poor generalization.

Algorithm 1: K-mean Clustering for User Behavior Prediction

Require: X ← the data points (matrix of size n × m)

Require: K ← the number of clusters

Ensure: C ← the final centroids (matrix of size K × m)

Ensure: idx ← a vector of size n containing the index of the cluster to which each data point belongs

1: C ← X[randomly select K rows]

2: old_idx ← None

3: while idx ≠ old_idx do

4:  old_idx ← idx

5:  for i ← 1 to n do

6:    idx[i] = arg min_k‖x_i − c_k‖²

7:   for k ∈ [1, K] do

8:    S_k = x_i: idx[i] = k

9:    if |S_k| > 0 then

10:

11:    else

12:     C_k ← X[randomly select one row]

13:    end if

14:   end for

15:  end for

16: end while

17: return C, idx

Moreover, the Within-Cluster Sum of Squares (WCSS) [70] serves as a metric employed to assess the quality of K-means clustering. WCSS is computed as the sum of the squared distances between each data point and its assigned centroid. The primary objective of K-means is to minimize WCSS, as it reflects the compactness of the clusters. A lower WCSS signifies that data points within each cluster are closer to each other, indicating the effectiveness of the clustering process. (12)

Hence, the optimal value of K can be determined using the elbow method, which involves plotting the WCSS for different values of K and selecting the value of K at the elbow point of the curve.

According to the results presented in Fig 3, the Elbow method was employed to ascertain the optimal number of clusters for the datasets analyzed in this study. The Elbow method entails plotting the within-cluster sum of squares (WCSS) against the number of clusters and identifying the elbow point on the resultant curve. The elbow point corresponds to the number of clusters at which the rate of decrease in WCSS begins to level off, beyond which the addition of more clusters does not yield a substantial reduction in WCSS.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 3. WCSS vs number of clusters plot generated using the elbow method.

The elbow point indicates the optimal number of clusters to use in the clustering algorithm, balancing cluster specificity and the within-cluster sum of squares (WCSS). (a) Movies data. (b) Series data. (c) Collective data.

https://doi.org/10.1371/journal.pone.0312520.g003

For the movie dataset, the curve reached the elbow point at a value of K = 3, indicating that three clusters would be optimal for this dataset. Conversely, for the series dataset and the collective dataset, the elbow points were observed at values of K = 5 and K = 6, respectively. Based on these results, the analysis will proceed using the optimal number of clusters identified for each dataset.

Principal component analysis.

Principal Component Analysis (PCA) is a technique employed to diminish the dimensionality of a dataset while preserving as much of the data’s variability as feasible [71]. This method finds extensive application in data analysis, machine learning, and statistics. PCA operates by identifying a fresh set of variables known as principal components, which are linear combinations of the original variables. (13)

These principal components are arranged in order of the variance they account for in the original dataset. The first principal component, PC₁, is the linear combination of X that has the largest variance. The weights must satisfy the constraint that . The second principal component, PC₂, is also the linear combination of X that has the second-largest variance, subject to the constraint that it is orthogonal to PC₁. Similarly, the k^th principal component, PC_k, is the linear combination of X that has the k^th-largest variance, subject to the constraints that it is orthogonal to PC₁, PC₂, ⋯, PC_(k−1). This can be expressed as: (14) where a_k1, a_k2, …, a_kp are the loadings or weights of the variables X₁, X₂, …, X_p in PC_k. The loadings must satisfy the constraints that for i = 1, 2, …, p and a_kia_kj = 0 for i ≠ j and i, j = 1, 2, …, k − 1. By choosing only the top few principal components, we can reduce the dimensionality of the dataset while still retaining a large amount of the variability in the original data. (15)

Let X be an n × p matrix representing a dataset with n observations and p variables. Assume that the variables have been centered to have a mean of 0. The goal of PCA is to find a new set of k ≤ p variables Z₁, Z₂, …, Z_k that are linear combinations of the original variables, such that the Z_i are orthogonal and explain as much of the variance in X as possible. Specifically, the Z_i are defined as: (16) where Xj is the jth variable of X, and aij are the loading coefficients that determine the weights of each variable in each Z_i. The loading coefficients are chosen to maximize the total variance explained by the Z_i, subject to the constraint that the Z_i are orthogonal and have unit length. (17)

Once the loading coefficients have been computed, each observation xi in X can be transformed into the lower-dimensional representation yi = (yi1, yi2, …, y_ik), where y_ij is the value of the jth component of Z_j for the ith observation. (18)

The resulting matrix Y, which has dimensions n × k, represents the lower-dimensional version of the original dataset. It can be used in place of X for visualization, clustering, classification, or other machine-learning tasks that are sensitive to the number of variables

In the context of k-means clustering (Fig 4), PCA can be used to visualize the clustering results when the dataset has many dimensions. Specifically, after performing k-means clustering on the high-dimensional data, we can use PCA to reduce the dimensionality of the data to two or three dimensions, which can be easily plotted on a graph.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 4. Visualization of K-means clustering results using PCA dimensionality reduction.

(a) Movies data. (b) Series data. (c) Collective data.

https://doi.org/10.1371/journal.pone.0312520.g004

Performance metrics & evaluation criteria

Performance metrics encompass quantitative measures employed to gauge the proficiency of a classification model. These metrics serve the purpose of evaluating the model’s precision in accurately predicting the class labels of test data. Within classification methodologies, performance metrics assume a pivotal role in appraising the caliber of classification outcomes [72]. They furnish insights into the model’s accuracy, dependability, and its capacity to generalize. Notably, among these performance measures, both the confusion matrix and classification report are employed to assess the effectiveness of a classification model.

A confusion matrix serves as a tabular tool employed to assess the performance of a classification model. It provides a breakdown of key metrics for each class in the dataset, including the count of true positives TP, false positives FP, true negatives TN, and false negatives FN. In the context of binary classification, the confusion matrix is structured with two rows and two columns, symbolizing the two classes under consideration. The matrix comprises four entries, which encompass:

• True positives TP: It represents the count of instances that are genuinely positive (true) and have been correctly classified as positive (true) by the model.
• False positives FP: It signifies the count of instances that are actually negative (false) but have been erroneously classified as positive (true) by the model.
• True negatives TN: It denotes the count of instances that are genuinely negative (false) and have been accurately classified as negative (false) by the model.
• False negatives FN: This corresponds to the count of instances that are truly positive (true) but have been incorrectly classified as negative (false) by the model.

These terms are used to calculate various performance measures such as precision, recall, F1-score, and accuracy.

A classification report serves as a concise summary of the performance metrics for a classification model. It encompasses precision, recall, and the F1-score for each class label, in addition to providing an overall accuracy assessment of the model.

• Precision measures the proportion of correctly classified instances in the positive class (i.e., true positives) out of all instances classified as positive by the model (true positives and false positives). (19)
• Recall (also known as sensitivity) measures the proportion of correctly classified instances in the positive class (true positives) out of all instances that are actually in the positive class (true positives and false negatives). (20)
• F1-score is a weighted average of precision and recall and provides a single value to summarize the performance of the model. (21)
• Accuracy measures proportion of correctly classified instances (both true positives and true negatives) out of all instances in the test data. (22)

Fundamentally, the confusion matrix offers a graphical portrayal of classification outcomes, whereas the classification report furnishes a more comprehensive and detailed analysis of the model’s performance. Together, these tools are invaluable for assessing the efficacy of a classification model.

Problem formulation

Let X = {x₁, x₂, x₃, …., x_m} denote the set of users, where |X| = m, and G = {g₁, g₂, g₃, …., g_n} denote the set of items or genres, where |G| = n. For a user, x ∈ X, let represent the sequence of genres that user x interacts with before time t, where . Let denote the set of historical sequences for all users.

The aim of historical sequence recommendation is to predict the next genre that a user x will interact with after time t, based on their historical sequence . This can be formally expressed as: (23) where f is a function that maps historical sequences to the next genre in the sequence. (24) Where T is the maximum time of the historical data. The goal is to learn this function f from the historical sequences in H, such that it accurately predicts the next genre in the sequence for any given user and time t.

The architecture of Duration-Count Matrix-based User Profiling (DCM-UP)

The architecture of the DMC-UP historical sequential model is depicted in Fig 5, consisting of four distinct components. Firstly, the Embedding State employs advanced embedding technology to model the distributed representation of both users and items. Secondly, the Dynamic Hierarchical Transformer State systematically merges each level from the top to the bottom layer by layer, updating each user’s interest or behavior through the count matrix and ultimately yielding the users’ long-term interest expression. Thirdly, the State Recommendation methodically models the relationship between user interest and their subsequent moment interest representation, resulting in a uniform and consistent expression of user interest. Lastly, the Fully Connected Control State accurately models the control or comparison between user interest and the implicit representation of the user based on items.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 5. The historical sequential architecture of DCM-UP.

It has four parts: embedding state, dynamic hierarchical updating state, state recommendation, and fully connected control state.

https://doi.org/10.1371/journal.pone.0312520.g005

Fully connected control state.

Let us define a binary labeling function that assigns the label T if the actual genre y matches the recommended genre , and F otherwise (Fig 6). In other words, (36)

Download:

• PNG
  larger image
• TIFF
  original image

Fig 6. The system keeps track of each user’s watch-time duration and genres based on user preferences.

The system updates the current watch-time duration by adding it to the previous duration values and recommends items (genres) to each user based on the highest weight of watch-time duration.

https://doi.org/10.1371/journal.pone.0312520.g006

Due to the “cold-start problem”, there are certain items that cannot be recommended to initial users. Our algorithm takes this into account and recommends only those items that can be recommended to the users.

Therefore, given a watch-time duration matrix M of size n × k where n is the number of users and k is the number of genres, our system recommends genres for each user by finding the genre with the highest watch-time duration for that user. The recommended genre is then compared with the actual genre watched by the user and a label is assigned using the function as described above.

Algorithm 4: Fully Connected Control State

Require: Watch-time duration matrix M of size n × k where n is the number of users and k is the number of genres

Ensure: Recommended genres for each user with assigned

1: n ← number of users;

2: k ← number of genres;

3: D ← empty dictionary;

4: for i from 1 to n do

5:  g_i ← genre with highest watch-time duration for user i

6:  if canRecommend(g_i) then

7:   D[i] ← g_i

8:  else

9:   D[i] ← defaultGenre

10:  end if

11: end for

12: return D

Duration-Count Matrix based User Similarity (DCM-US)

In this study, we propose a system that generates a matrix to measure user-to-user similarity by utilizing user-item generalization. The goal of this process is to predict the similarity between users by sequentially reading their historical data, based on specific items and their corresponding watch-time duration.

Algorithm 5: Generating User-to-User Similarity Matrix

Require: M (n × k matrix), T (integer), L (integer)

Ensure: S (n × n matrix)

1: X ← aggregate function(M)

2: for i ← 1 to n do

3:  for j ← 1 to k do

4:   X′ij ← 0

5:   for l ← 1 to L do

6:    if i − l ≥ 1 then

7:     X′ij ← X′ij + Xi − l, j

8:    end if

9:   end for

10:  end for

11: end for

12: for i ← 1 to n do

13:  for j = i to n do

14:   numerator ← 0

15:   for t = 1 to k do

16:    numerator ← numerator + (X′it − mean(X′i)) × (X′jt − mean(X′j))

17:   end for

18:   

19:   

20:   Sji ← S_ij

21:  end for

22: end for

23: return S

Let M be the watch-time duration matrix of size n × k, where n is the number of users and k is the number of items. Our system generates a user-item matrix X of size n × k, where X_ij denotes the duration of watch time for item j by user i. We use an aggregate function f to predict each user’s behavior based on the duration of watch time. This function computes the sum of the watch-time duration for each item and user, yielding a matrix X. (37)

To estimate the time a particular user spent watching a video for each specific item, the system analyzes the behavior of similar users. In the user’s sequential historical behavior, the system learns each user’s behavior by using the duration of watch time, sums it from the first level to the current level, and adds it into a matrix-specific chunk. Let S be a similarity matrix of size n × n, where S_ij represents the similarity between users i and j.

The estimated duration of watch time for item j by user i based on similar users’ behavior up to level T and depth L is denoted as X′ij. To compute X′ij, we sum the watch-time duration of similar users up to level T and depth L for item j and user i as follows: (38)

The process of adding this information into the matrix is repeated for all items, resulting in a matrix X′ that captures the estimated watch-time duration for each user-item pair. Finally, the system computes the similarity matrix S between users based on X′. The similarity between users i and j is computed as follows: (39) where and are the mean values of the estimated watch-time duration for each user-item pair for users i and j, respectively. This similarity measure allows us to compare the behavior of users and make recommendations based on their preferences.

Baseline method

To validate the effectiveness of DCM (Duration Count Matrix), we conduct a comparative analysis with a set of long-term user behavior modeling algorithms.

SEMI [37] a pioneering sequential multi-modal information transfer network, revolutionizes micro-video recommendations by harnessing users’ product domain behavior.

MMTHA [20] is a Multi-scale Modeling of Users’ Historical Behavior for Micro Video Recommendations that models users’ historical behavior across multiple scales. By capturing users’ short-term dynamic interests and incorporating long-term correlations, MMTHA effectively predicts user behavior on micro videos.

BAR [73] is a Behavior-Aware Recommendation that emphasizes the user’s sequential and heterogeneous one-class feedback. By incorporating behavior information into the input and output of a representation module, BAR effectively captures the item sequence and its relationship to the user’s real next behavior.

CT [74] revolutionizes the next video recommendation with its collaborative transformer architecture. This unified framework excels at capturing both micro video representation and sequential user-video historical interactions.

RLUR [40] is a Reinforcement Learning for User Retention to model long-term user feedback in the context of short-video recommendation systems that focus on predicting user retention. The objective is to minimize the time it takes for users to return to the system while maximizing long-term performance.

PreRec [44] is a Preference-based Recommender System that optimizes long-term user engagement in recommendations by leveraging preferences based on historical behavior, rather than relying solely on explicit behavior.

Experiments setup

Tables 4–6 presents a summary of the different hyperparameter values used in the proposed video recommendation models. To evaluate the performance of the proposed methodologies, we randomly assigned 70% of the data for each user as the training set and the remaining 30% as the testing set. Machine learning techniques, specifically the decision tree algorithm, were applied to assess the performance of the proposed methodologies. We comprehensively evaluated the proposed methodologies for video recommendations by selecting the parameter configuration. The hyperparameters, including the maximum depth, splitters, criterion, and maximum leaf nodes, were systematically configured for each dataset.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. Optimized hyperparameters for movies dataset.

https://doi.org/10.1371/journal.pone.0312520.t004

Download:

• PNG
  larger image
• TIFF
  original image

Table 5. Optimized hyperparameters for series dataset.

https://doi.org/10.1371/journal.pone.0312520.t005

Download:

• PNG
  larger image
• TIFF
  original image

Table 6. Optimized hyperparameters for video-streaming dataset.

https://doi.org/10.1371/journal.pone.0312520.t006

As we see in Table 4, we found that the optimal maximum depth for DCM-UP on the movies dataset was 5, while the optimal maximum depth for DCM-US was 4. Both techniques employed the entropy criterion and chose the best splitter. Furthermore, DCM-UP utilized 32 maximum leaf nodes, while DCM-US utilized 7 maximum leaf nodes. Therefore, we can express the hyperparameters selected for the movies dataset as:

Based on our analysis of the series dataset, we have determined that the ideal maximum depth for the DCM-UP model is 5, while the DCM-US model performs best with a maximum depth of 6. These findings are outlined in Table 5. For the DCM-UP model, we employed the gini criterion [75] to select the optimal splitter, while for the DCM-US model, the entropy criterion was utilized. Furthermore, the DCM-UP model was configured with a maximum of 32 leaf nodes, while the DCM-US model utilized a maximum of 12 leaf nodes. Therefore, we can summarize the hyperparameters chosen for the series dataset as follows:

Similarly, the hyperparameters chosen for DCM-UP and DCM-US on the collective video streaming dataset are presented in Table 6. In the case of DCM-UP, the optimal maximum depth was determined to be 7, whereas, for DCM-US, it was 5. Both models utilized the entropy criterion to select the most suitable splitter. Furthermore, DCM-UP was configured with a maximum of 100 leaf nodes, while DCM-US utilized only 12. In summary, the selected hyperparameters for the video streaming dataset can be summarized as follows:

Comparative analysis

To ascertain the efficacy of the suggested video recommendation techniques, an extensive comparative analysis was undertaken in this investigation. The comparison encompassed a wide range of machine-learning models for long-term behavioral prediction. The outcomes of this comparison, including the performance of the DCM-UP and DCM-US models in relation to other models, were presented in Tables 7–10, utilizing three distinct datasets. Essential evaluation metrics such as accuracy, precision, recall, and F1-score were employed to demonstrate the effectiveness of the techniques.

Download:

• PNG
  larger image
• TIFF
  original image

Table 7. A comparative analysis of model performance in Duration Count Matrix-based User Profiling (DCM-UP).

https://doi.org/10.1371/journal.pone.0312520.t007

Download:

• PNG
  larger image
• TIFF
  original image

Table 8. Comparative study: Analyzing user similarity with Duration Count Matrix (DCM-US) in movies dataset.

https://doi.org/10.1371/journal.pone.0312520.t008

Download:

• PNG
  larger image
• TIFF
  original image

Table 9. Comparative study: Analyzing user similarity with Duration Count Matrix (DCM-US) in series dataset.

https://doi.org/10.1371/journal.pone.0312520.t009

Download:

• PNG
  larger image
• TIFF
  original image

Table 10. Comparative study: Analyzing user similarity with Duration Count Matrix (DCM-US) in video streaming dataset.

https://doi.org/10.1371/journal.pone.0312520.t010

According to the results presented in Table 7, the DCM-UP model demonstrated superior performance compared to the Random Forest and Gradient Boosting models in the domain of decision trees. This can be attributed to several factors. Notably, the Decision Tree model exhibited remarkable accuracy in both the movies and series datasets, achieving training accuracies of 87% and 89%, respectively. In contrast, the Random Forest and Gradient Boosting models achieved lower accuracy scores of 75%, 74%, and 73% respectively. These findings highlight the efficacy of the DCM-UP model in comparison to the alternative models when considering decision tree-based approaches.

Based on the findings from DCM-US (Tables 8–10), an impressive 98% accuracy was attained in analyzing the movie dataset. Furthermore, when examining series data and video streaming datasets, the Decision tree model alone yielded a remarkable accuracy of 98%. Similarly, the random forest and gradient boosting approaches proved to be highly effective, achieving an accuracy of 97% in this particular context.

Evaluation results

The effectiveness of our innovative recommender model, based on the duration count matrix, is evaluated through a rigorous methodology. In order to achieve a comprehensive assessment, we partitioned our dataset into a 70-30 split, dedicating 70% of the watch-time duration for model training while reserving the remaining 30% for evaluation purposes. The model’s performance was subjected to a thorough analysis, employing fundamental metrics such as precision, recall, f1-score, and accuracy. The significance of achieving higher precision, recall, F1-score, and accuracy lies in the fact that they indicate the relevance and quality of the recommendations provided. An increase in these metrics corresponds directly to the enhancement of recommendation quality and precision, ultimately elevating the overall user experience.

In our evaluation, we conduct a thorough analysis of our proposed method against state-of-the-art approaches. This critical analysis serves as a benchmark, allowing us to explicate the advancements and contributions of our approach in relation to existing methodologies. Through a comprehensive review of relevant literature, we aim to identify key challenges and limitations encountered by previous methods. Moreover, this comparative analysis offers insights into the strengths and advantages of our methodology, Specifically in addressing key challenges within the video recommendation. Additionally, our evaluation highlights the robustness of our method across diverse datasets and scenarios, showcasing its adaptability and applicability in real-world settings.

Moreover, Table 11 offers an extensive evaluation of the proposed DCM-UP technique, considering a range of hyperparameter configurations and assessing multiple performance metrics. Remarkably, the analysis highlights the superiority of series data over movie data and video streaming data, achieving the highest level of accuracy. This notable difference in performance can be attributed to the significantly larger dataset size and the richer content available in the series data, which allows for more robust model training and better generalization.

Download:

• PNG
  larger image
• TIFF
  original image

Table 11. Optimizing optimizing DCM-UP model performance using tuned hyperparameters.

https://doi.org/10.1371/journal.pone.0312520.t011

In this study, when focusing on the analysis of series data, we observed that default hyperparameters (HPs) achieved a remarkable accuracy of 89%, while the tuned HPs achieved a slightly lower accuracy of 76%. Conversely, when examining both movie and series data together (as depicted in Fig 7), the utilization of tuned HPs resulted in an accuracy rate of 72%. In contrast, default HPs yielded an accuracy of 87% for movie data and 73% for video streaming data. These findings underscore the critical role of hyperparameter tuning in optimizing the model’s performance.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 7. The depicted figure illustrates the correlation between the maximum depth and the accuracy of the DCM-UP approach.

(a) Movies Dataset. (b) Series Dataset. (c) Video Streaming Dataset.

https://doi.org/10.1371/journal.pone.0312520.g007

In this study, we applied the innovative DCM-US technique to identify users sharing similar behaviors and preferences, thereby facilitating personalized recommendations. Table 12 shows the optimizing DCM-UP model performance Using default hyperparameters. The evaluation, as showcased in Table 13, was conducted across various datasets using finely-tuned hyperparameters (HPs), which led to exceptionally high accuracy outcomes. To be more specific, our analysis revealed that the movies data, series data, and video streaming data all achieved a remarkable accuracy rate of 98% each, as illustrated in (Fig 8).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 8. The depicted figure illustrates the correlation between the maximum depth and the accuracy of the DCM-US approach.

(a) Movies Dataset. (b) Series Dataset. (c) Video Streaming Dataset.

https://doi.org/10.1371/journal.pone.0312520.g008

Download:

• PNG
  larger image
• TIFF
  original image

Table 12. Optimizing DCM-UP model performance using default hyperparameters.

https://doi.org/10.1371/journal.pone.0312520.t012

Download:

• PNG
  larger image
• TIFF
  original image

Table 13. Fine-tuning the DCM-US model for maximum effectiveness.

https://doi.org/10.1371/journal.pone.0312520.t013

Delving into the intricacies of the model’s decision-making process, we directed our attention toward a vital hyperparameter known as “max depth.” Through thorough analysis and experimentation, we explored the impact of different “max depth” values on the model’s performance.

In our analysis of the movies dataset (Fig 8a), setting the “max depth” hyperparameter to 4 resulted in optimal performance. This balance between complexity and simplicity effectively captured inherent patterns, leading to highly accurate recommendations.

In analyzing series data (Fig 8b), we found that a “max depth” of 6 resulted in exceptional accuracy, allowing the model to capture intricate relationships effectively. Conversely, for the video streaming dataset (Fig 8c), including movies, series, and additional factors, a “max depth” of 5 achieved outstanding accuracy, successfully navigating complexities while avoiding overfitting.

After assessing the accuracy rates, the evaluation of model performance extends to error-based measures. [63] In this section, we examine key error-based metrics, including Gini impurity Entropy, and misclassification error, to provide a comprehensive understanding of our model’s performance.

In further detail, we analyze Gini impurity [75], a crucial metric that measures the impurity of a dataset by calculating the probability of misclassifying a randomly chosen instance. When assessing the DCM-UP, default hyperparameters produce an impurity score of 0.3, contrasting with the 0.1 impurity achieved with tuned hyperparameters. Similarly, leveraging the DCM-US technique results in a notable reduction in impurity, yielding a score of 0.08. (40)

Furthermore, Entropy [76] quantifies the uncertainty or disorder present in a dataset. It is computed as the negative sum of the probabilities of each class multiplied by the logarithm of the probability. In our evaluation of the DCM-UP, both default and tuned hyperparameters resulted in an impurity score of 0.3. On the contrary, when utilizing the DCM-US technique, we achieve a notable decrease in impurity, with a score of 0.03. (41)

Moreover, Misclassification error [77] evaluates the ratio of misclassified instances to the total number of instances. When evaluating the DCM-UP, we achieve a score of 0.1 with default HPs and 0.2 with tuned HPs. On the other hand, we achieve error score of 0.01, when employing the DCM-US technique. (42)

Hyperparameter senstivity analysis

In certain cases, default hyperparameters in decision tree models demonstrate superior performance compared to fine-tuned hyperparameters [78]. This can be attributed to several factors. Firstly, default hyperparameters are designed to provide a good balance between model complexity and generalizability. They are chosen based on heuristics and prior knowledge to work reasonably well across various datasets and scenarios. This inherent generalizability makes default hyperparameters robust and reliable. Secondly, default hyperparameters are less prone to overfitting, a phenomenon where the model becomes overly specialized to the training data and performs poorly on unseen data. The conservative nature of default hyperparameters helps prevent overfitting and encourages better generalization to new data. Additionally, in situations where the dataset is small or lacks diversity, fine-tuning hyperparameters may not yield significant improvements. The predefined values of default hyperparameters provide a sensible starting point that avoids over-optimization and ensures reasonable performance. Default hyperparameters offer a convenient and efficient solution in such scenarios. However, it is important to note that the performance of default hyperparameters can still vary depending on the specific dataset and problem domain. Therefore, it is advisable to explore and experiment with fine-tuning hyperparameters to uncover optimal configurations that suit the unique characteristics of the data. Ultimately, the choice between default and fine-tuned hyperparameters depends on the trade-off between performance and the available resources and constraints.