Note: Descriptions are shown in the official language in which they were submitted.
CA 03153598 2022-03-07
METHOD OF AND DEVICE FOR PREDICTING VIDEO PLAYBACK INTEGRITY
BACKGROUND OF THE INVENTION
Technical Field
[0001] The present invention relates to the fields of big data and deep
learning technologies, and
more particularly to a method of and a device for predicting video playback
integrity.
Description of Related Art
[0002] The video recommendation system is created by researching user
interests and
predilections on the basis of large number of users and large quantities of
videos in
dependence of big data analysis and artificial intelligence technology, the
system
recommends user-interested, high-quality videos to target users, solves the
problem of
information overload, achieves the effect of customized services, and enhances
both time
duration of stay and satisfaction of users. the video recommendation system
usually
includes two phases of recalling and sorting, of which the recalling phase is
to select
certain candidate sets from massive videos, and the sorting phase is to
perform more
precise and unified calculation on the candidate sets selected in the
recalling phase, and
to screen out few videos of excellent quality that are most interesting to
users from the
candidate sets.
[0003] Currently, the number of users registered in some video playback
platforms reaches
hundred millions, with daily UV (UniqueVisitor) access exceeding ten millions,
and the
number of plays per day is even higher in mobile ends. In order that users
find out contents
of interest to them from massive videos, a recommendation system is created by
collecting data of plural dimensions (including basic information of users,
playback
histories of users, video attributes and environment attributes, etc.) to
associate the users
1
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
with potentially preferred videos. There is fewer information usable for the
recommendation of short videos, only such information as titles and video
categories is
usable for the purpose, and the currently frequently used sorting model
employs the
method of CTR (Click-Through-Rate) prediction. Clickbaits might be encouraged
on the
basis of the click model, and this cannot bring about longer time duration of
stay of users,
and would adversely affect the watching time duration and satisfaction of
users. However,
watching time duration is an important optimization target in the information
flow, so it
is urgently needed to introduce playback integrity optimization in the short
video sorting
model, so as to enhance reality relevancy of recommendations, and to achieve
enhancement in the watching time duration and satisfaction of users.
SUMMARY OF THE INVENTION
[0004] In order to address problems pending in the state of the art,
embodiments of the present
invention provide a method of and a device for predicting video playback
integrity, by
introducing a playback integrity improving policy to predict user video
playback integrity,
and acquiring interest data of the user more approaching to reality in such
important
information flow aspect as the watching time duration, the present invention
enhances the
precision in recognizing user interests, hence enhances reality relevancy of
recommendations, and achieves relatively great enhancement in the watching
time
duration and satisfaction of users.
[0005] The technical solutions are as follows:
[0006] According to one aspect, there is provided a method of predicting video
playback integrity,
and the method comprises:
[0007] inputting to-be-tested data of a user video playback feature vector;
[0008] calculating through a preset video playback integrity prediction model;
and
[0009] outputting a video playback integrity value of the to-be-tested data;
wherein
2
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0010] the preset video playback integrity prediction model is obtained
through training by user
video playback training data, and the user video playback feature vector at
least includes
a user feature vector and a video feature vector.
[0011] Moreover, the method further comprises:
[0012] collecting user video playback information data;
[0013] screening the user video playback information data, and obtaining a
screening result; and
[0014] performing a feature extraction on the screening result, and generating
to-be-tested data
of the user video playback feature vector.
[0015] Further, the step of collecting user video playback information data
includes: obtaining
the user video playback information data containing user information, user
playback
historical information, video information and user client side information;
and/or
[0016] the step of screening the user video playback information data, and
obtaining a screening
result includes: screening the user video playback information data by
employing a multi-
channel recalling mode including user collaboration, user searching, a topic
model,
popular recommendation, a user portrait and a video tag, and obtaining a
screening result;
and/or
[0017] the step of performing a feature extraction on the screening result,
and generating to-be-
tested data of the user video playback feature vector includes: performing
word
segmentation on a video title and a video classification tag in the screening
result by
employing a word vector obtained by training a preset massive corpus through a
word2vec model and IDF weight training, generating a video word vector,
thereafter
performing word vector calculation according to the user playback historical
information
in conjunction with time decay, and generating a user word vector.
[0018] Further, the preset video playback integrity prediction model contains
DNNs of three
hidden layers.
3
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0019] Further, the preset video playback integrity prediction model is
obtained through training
by inputting the user video playback training data, wherein the user video
playback
training data is an independent variable, while a user watching history video
playback
integrity value is a dependent variable, and the user video playback training
data is a
feature vector combined by a historical user vector and a historical video
vector created
according to the user playback historical information.
[0020] Moreover, the method further comprises:
[0021] sorting video playback integrity values of the to-be-tested data in a
decreasing order,
obtaining topN video sorting results, and recommending the video sorting
results to a
corresponding user according to priority level, wherein N is an integer
greater than 1.
[0022] According to another aspect, there is provided a device for predicting
video playback
integrity, the device comprises a model calculating module, and the model
calculating
module is employed for:
[0023] inputting to-be-tested data of a user video playback feature vector,
calculating through a
preset video playback integrity prediction model, and outputting a video
playback
integrity value of the to-be-tested data, wherein the preset video playback
integrity
prediction model is obtained through training by user video playback training
data, and
the user video playback feature vector at least includes a user feature vector
and a video
feature vector.
[0024] Moreover, the device further comprises a data collecting module, a data
screening module,
and a vector generating module, of which the data collecting module collects
user video
playback information data, the data screening module screens the user video
playback
information data, and obtains a screening result, and the vector generating
module
performs a feature extraction on the screening result, and generates to-be-
tested data of
the user video playback feature vector.
4
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0025] Further, the data collecting module obtains the user video playback
information data
containing user information, user playback historical information, video
information and
user client side information; and/or
[0026] the data screening module screens the user video playback information
data by employing
a multi-channel recalling mode including user collaboration, user searching, a
topic
model, popular recommendation, a user poi _________________________________ ti
ait and a video tag, and obtains a screening
result; and/or
[0027] the vector generating module performs a feature extraction on the
screening result, and
generates to-be-tested data of the user video playback feature vector,
including:
performing word segmentation on a video title and a video classification tag
in the
screening result by employing a word vector obtained by training a preset
massive corpus
through a word2vec model and IDF weight training, generating a video word
vector,
thereafter performing word vector calculation according to the user playback
historical
information in conjunction with time decay, and generating a user word vector.
[0028] Moreover, the device further comprises a data recommending module for
sorting video
playback integrity values of the to-be-tested data in a decreasing order,
obtaining topN
video sorting results, and recommending the video sorting results to a
corresponding user
according to priority level, wherein N is an integer greater than 1.
[0029] The technical solutions provided by the embodiments of the present
invention bring about
the following advantageous effects.
[0030] 1. By modifying the traditional CTR prediction method altogether, a
video playback
integrity indicator is introduced, video playback integrities of different
users are predicted
through a well trained preset video playback integrity prediction model,
interest data of
users more approaching to reality is obtained in such important information
flow aspect
as the watching time duration through prediction results of the video playback
integrities,
the precision in recognizing user interests is enhanced, reality relevancy of
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
recommendations is hence enhanced, and relatively great enhancement is
achieved in the
watching time duration and satisfaction of users.
[0031] 2. Through vectorized representation of the user portrait, interest
transfer of the user is
reflected in combination with time decay of user behaviors, and hotspot videos
and
inadvertently clicked videos are filtered out in the process of user poll" __
ait, whereby
interference with actual interest of the user is avoided, and the user poi __
tiait is made more
precise.
[0032] 3. By collecting such relevant data as user behavior data, video
quality and video
information, etc., vectorized representation of user features and video
attributes is
effectively made, as well as proportions of videos played back at various time
periods,
proportions of various categories and other environment information, different
features
and different data sources are merged in the application of short video
recommendation
sorting model through deep learning modeling and prediction of potential
playback
integrities of videos not watched by users, excellent effect is achieved, and
average
watching time duration of users is enhanced.
[0033] 4. By creating such features as user features, video features,
contextual features and client
side classification, deep learning modeling is employed, the playback
integrity prediction
mode is applied in a group of 10% randomly selected users through the AB test,
and such
indicators as CTR, daily average playback volume and user average playback
integrity
are compared through the final report. In the end, user average playback
integrity and
daily average playback volume are enhanced to a greater extent with slight
decrease in
the CTR.
[0034] 5. A TF-IDF algorithm is employed in terms of video recommendation, and
key
information of videos is effectively emphasized through IDF values.
6
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0035] 6. Reality relevancy of recommendations is enhanced through prediction
of short video
playback integrities, and increase in the time duration of stay of users is
attempted.
BRIEF DESCRIPTION OF THE DRAWINGS
[0036] To more clearly describe the technical solutions in the embodiments of
the present
invention, drawings required to illustrate the embodiments are briefly
introduced below.
Apparently, the drawings introduced below are merely directed to some
embodiments of
the present invention, while persons ordinarily skilled in the art may further
acquire other
drawings on the basis of these drawings without spending creative effort in
the process.
[0037] Fig. 1 is a flowchart illustrating the method of predicting video
playback integrity
provided by an embodiment of the present invention;
[0038] Fig. 2 is a flowchart illustrating the method of predicting video
playback integrity
provided by another embodiment of the present invention;
[0039] Fig. 3 is a view schematically illustrating a preferred mode of
execution of feature
engineering construction in Step 203;
[0040] Fig. 4 is a view schematically illustrating a preferred mode of
execution of a preset video
playback integrity prediction model provided by an embodiment of the present
invention;
[0041] Fig. 5 is a view schematically illustrating the structure of a device
for predicting video
playback integrity provided by an embodiment of the present invention; and
[0042] Fig. 6 is a view schematically illustrating the structure of a device
for predicting video
playback integrity provided by another embodiment of the present invention.
7
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
DETAILED DESCRIPTION OF THE INVENTION
[0043] To make more lucid and clear the objectives, technical solutions and
advantages of the
present invention, the technical solutions in the embodiments of the present
invention will
be clearly and comprehensively described below with reference to the
accompanying
drawings in the embodiments of the present invention. Apparently, the
embodiments as
described are merely partial, rather than the entire, embodiments of the
present invention.
Any other embodiments makeable by persons ordinarily skilled in the art on the
basis of
the embodiments in the present invention without creative effort shall all
fall within the
protection scope of the present invention. As should be noted, the wordings
"plural/more/a plurality of' mean "two and more" in the description of the
present
invention, unless otherwise definitely and specifically defined.
[0044] In the method of and device for predicting video playback integrity
provided by the
embodiments of the present invention, by modifying the traditional CTR
prediction
method altogether, a video playback integrity indicator is introduced, video
playback
integrities of different users are predicted through a well trained preset
video playback
integrity prediction model, interest data of users more approaching to reality
is obtained
in such important information flow aspect as the watching time duration
through
prediction results of the video playback integrities, the precision in
recognizing user
interests is enhanced, reality relevancy of recommendations is hence enhanced,
and
relatively great enhancement is achieved in the watching time duration and
satisfaction
of users. Accordingly, the method of and device for predicting video playback
integrity
are widely applicable to many network video application scenarios concerning
user
interests mining, user requirements matching or user recommendations.
[0045] Specific embodiments and accompanying drawings are combined below to
describe in
detail the method of and device for predicting video playback integrity
provided by the
embodiments of the present invention.
8
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0046] Fig. 1 is a flowchart illustrating the method of predicting video
playback integrity
provided by an embodiment of the present invention. As shown in Fig. 1, the
method of
predicting video playback integrity comprises the following steps:
[0047] 101 - inputting to-be-tested data of a user video playback feature
vector;
[0048] 102 - calculating through a preset video playback integrity prediction
model; and
[0049] 103 - outputting a video playback integrity value of the to-be-tested
data.
[0050] Different from the traditional user technology in which only few such
collected
information as titles, video categories or click-through-rate is used, the
user video
playback feature vector here at least includes user feature vectors and video
feature
vectors, the user feature includes user poi _______________________________ ti
aits, user historical playback records or other
information relevant to users, and the video information includes video
categories, video
time durations, video times, video playback integrity records or other
information
relevant to releasing videos. Besides user feature vectors and video feature
vectors, the
user video playback feature vector can further include such other information
relevant to
video playback as user client side classification information. In addition,
the preset video
playback integrity prediction model is obtained through training by user video
playback
training data, and the video playback integrity prediction model as
specifically used can
be either obtained by training a corresponding deep learning model designed
and
constructed according to requirements, or obtained by training any possible
deep learning
model available in the art, to which no particular restriction is made in the
embodiments
of the present invention.
[0051] Fig. 2 is a flowchart illustrating the method of predicting video
playback integrity
provided by another embodiment of the present invention. As shown in Fig. 2,
the method
of predicting video playback integrity comprises the following steps.
[0052] 201 - collecting user video playback information data.
9
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0053] Specifically, the user video playback information data containing user
information, user
playback historical information, video information and user client side
information is
obtained.
[0054] This process is a phase for collecting user video playback information
data, the user video
playback information mainly includes user information, user playback
historical
information, video information and user client side information, of which the
user
information mainly indicates user poll" ___________________________________
ait information, including basic attribute
information of a user (gender, age, etc.), the user playback historical
information includes
proportions of historical playbacks on a hourly basis by the user, and
proportions of
various types of videos watched by the user, etc., and the client side
information includes
user equipment types and operator types, etc. In addition to the above, such
contextual
information secondarily associated with videos played back by the user as the
time at
which the user watches each video and the user position information, etc., can
be further
collected for the user video playback information according to requirements.
[0055] As is notable, the process of collecting user video playback
information data in Step 201
can as well be realized by modes other than the mode recited in the
aforementioned step,
and these specific modes are not restricted in the embodiments of the present
invention.
[0056] 202 - screening the user video playback information data, and obtaining
a screening result.
[0057] Specifically, the step of screening the user video playback information
data, and obtaining
a screening result includes: screening the user video playback information
data by
employing a multi-channel recalling mode including user collaboration, user
searching,
a topic model, popular recommendation, a user portrait and a video tag, and
obtaining a
screening result.
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0058] This process is a phase for recalling coarsely screened user video
playback information
data, preferably, the screening is mainly directed to the video information in
the user
video playback information data. Since the video is colossal in scale,
possibly reaching
the order of several millions, direct input of the video into the model for
data
preprocessing would require extremely high cost, and the speed would also be
extremely
slow, so it is possible to coarsely screen out some video information with
higher quality
or in other words possibly more to the taste of users in the recalling phase.
Recalling is
usually embodied as multi-channel recalling, such as through user
collaboration, user
searching, topic models, popular recommendations, user portraits and video
tags, so as to
select certain desirable candidate sets from the massive amount of video.
[0059] As is notable, the process of screening the user video playback
information data in Step
202 can as well be realized by modes other than the mode recited in the
aforementioned
step, and these specific modes are not restricted in the embodiments of the
present
invention.
[0060] 203 - performing a feature extraction on the screening result, and
generating to-be-tested
data of the user video playback feature vector.
[0061] Specifically, the step of performing a feature extraction on the
screening result, and
generating to-be-tested data of the user video playback feature vector
includes:
performing word segmentation on a video title and a video classification tag
in the
screening result by employing a word vector obtained by training a preset
massive corpus
through a word2vec model and IDF weight training, generating a video word
vector,
thereafter performing word vector calculation according to the user playback
historical
information in conjunction with time decay, and generating a user word vector.
The user
word vector and the video word vector here correspond to the aforementioned
user feature
vector and video feature vector.
11
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0062] This process is a feature engineering phase, as shown in Fig. 3,
preferably, a word vector
with 200 dimensions per word is trained through word segmentation and a
word2vec
model on the massive corpus, potential meanings of words are characterized in
a
vectorized form, so as to express the relations between words, and word vector
representation of the video is calculated and obtained through a combination
of word
segmentation process of the video title with the information of IDF obtained
by training.
Word vector representation of the user is calculated according to the word
vector
representation of the user historical playback video in conjunction with time
decay; in the
process of calculating user vector, videos with top 3 tags of the user with a
proportion
exceeding 10% are counted according to video tag categories. As found
according to
playback history analysis of the user, videos to which lower proportional
video tags
correspond are not latent interest points of the user, playback of them is
usually because
they are hotspot videos or due to inadvertent clicking by the user, and they
can be
discarded through feature extraction.
[0063] As is notable, the process of performing a feature extraction on the
screening result, and
generating to-be-tested data of the user video playback feature vector in Step
203 can as
well be realized by modes other than the mode recited in the aforementioned
step, and
these specific modes are not restricted in the embodiments of the present
invention.
[0064] 204 - inputting to-be-tested data of a user video playback feature
vector.
[0065] The preset video playback integrity prediction model is obtained
through training by
inputting the user video playback training data, wherein the user video
playback training
data is an independent variable, while a user watching history video playback
integrity
value is a dependent variable, and the user video playback training data is a
feature vector
combined by a historical user vector and a historical video vector created
according to
the user playback historical information, for training to obtain a desirable
preset video
playback integrity prediction model.
12
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0066] Preferably, the preset video playback integrity prediction model
contains DNNs of three
hidden layers, and input information of the input layer includes word vector
representation of the user (various video word vectors are obtained by a
combination of
word segmentation of the user historical playback video with IDF weight
calculation, and
a word vector with 200 dimensions is subsequently calculated and obtained in
overall
consideration of time decay), basic portrait of the user (gender, age, etc.),
proportions of
videos played back at various time periods (on an hourly basis), and
proportions of
various categories of videos, etc.; word vector (200 dimensions) of the video,
quality of
the video (average playback integrity, video hits, etc.), releasing time of
the video, video
category, equipment type, operator type; region; current time period, etc.
[0067] As is notable, the data content and form of inputting to-be-tested data
of a user video
playback feature vector in Step 204 can as well be realized by modes other
than the mode
recited in the aforementioned step, and these specific modes are not
restricted in the
embodiments of the present invention.
[0068] 205 ¨ calculating through a preset video playback integrity prediction
model.
[0069] 206 - outputting a video playback integrity value of the to-be-tested
data.
[0070] Preferably, the following steps are further included after Step 206:
[0071] sorting video playback integrity values of the to-be-tested data in a
decreasing order,
obtaining topN video sorting results, and recommending the video sorting
results to a
corresponding user according to priority level, wherein N is an integer
greater than 1. As
should be noted, it is also possible to base on requirements to arrange the
step of sorting
video playback integrity values in the calculating process of the preset video
playback
integrity prediction model, as shown in Fig. 4, to which no particular
restriction is made
in the embodiments of the present invention.
13
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0072] Fig. 5 is a view schematically illustrating the structure of a device
for predicting video
playback integrity provided by an embodiment of the present invention. As
shown in Fig.
5, the device for predicting video playback integrity comprises a model
calculating
module 1, and the model calculating module 1 is employed for: inputting to-be-
tested
data of a user video playback feature vector, calculating through a preset
video playback
integrity prediction model, and outputting a video playback integrity value of
the to-be-
tested data, wherein the preset video playback integrity prediction model is
obtained
through training by user video playback training data, and the user video
playback feature
vector at least includes a user feature vector and a video feature vector.
[0073] Fig. 6 is a view schematically illustrating the structure of a device
for predicting video
playback integrity provided by another embodiment of the present invention. As
shown
in Fig. 6, the device 2 for predicting video playback integrity comprises a
data collecting
module 21, a data screening module 22, a vector generating module 23, a model
calculating module 24 and a data recommending module 25.
[0074] The data collecting module 21 collects user video playback information
data. Specifically,
the data collecting module 21 obtains user video playback information data
containing
user information, user playback historical information, video information and
user client
side information.
[0075] The data screening module 22 screens the user video playback
information data, and
obtains a screening result. Specifically, the data screening module 22 screens
the user
video playback information data by employing a multi-channel recalling mode
including
user collaboration, user searching, a topic model, popular recommendation, a
user poi Li ait
and a video tag, and obtains a screening result.
[0076] The vector generating module 23 performs a feature extraction on the
screening result,
14
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
and generates the user video playback feature vector. Specifically, the vector
generating
module 23 performs a feature extraction on the screening result, and generates
to-be-
tested data of the user video playback feature vector, including: performing
word
segmentation on a video title and a video classification tag in the screening
result by
employing a word vector obtained by training a preset massive corpus through a
word2vec model and IDF weight training, generating a video word vector,
thereafter
performing word vector calculation according to the user playback historical
information
in conjunction with time decay, and generating a user word vector. The user
word vector
and the video word vector here correspond to the following user feature vector
and video
feature vector.
[0077] The model calculating module 24 inputs to-be-tested data of a user
video playback feature
vector, calculates through a preset video playback integrity prediction model,
and outputs
a video playback integrity value of the to-be-tested data, wherein the preset
video
playback integrity prediction model is obtained through training by user video
playback
training data, and the user video playback feature vector at least includes a
user feature
vector and a video feature vector.
[0078] The data recommending module 25 sorts video playback integrity values
of the to-be-
tested data in a decreasing order, obtains topN video sorting results, and
recommends the
video sorting results to a corresponding user according to priority level,
wherein N is an
integer greater than 1.
[0079] A preferred mode of execution for the method of and device for
predicting video playback
integrity provided by the embodiments of the present invention is introduced
below.
[0080] Firstly, the word segmentation tool of the present embodiment carries a
lexicon therewith,
entertainment stars, film and TV drama names, sports stars and team
information are
additionally added as supplementary lexicon, Netease news, Baidu encyclopedia
and
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
Wikipedia obtained by a crawler system constitute a massive corpus, and word
segmentation and word vector training are performed with respect to the corpus
to finally
obtain word vector representation of each word (dimensions of the word vectors
are 200
dimensions, as determined by test effect, and normalization is then performed
on the
vectors).
[0081] Under the aforementioned corpus, TF-IDF training is carried out to
obtain IDF values,
normalization is then performed, weight enhancement is subsequently performed
on the
supplementary lexicon as 1, similar to the attention mechanism, more attention
is paid to
these words.
[0082] See the following Table 1 for video information, in which are carried
video IDs, video
title information, classification tags, video tag information, and releasing
times, etc. The
video information is word-segmented, a word vector table of words is searched,
and the
word vector representation of the current video is obtained in combination
with weighted
calculation of an IDF value table (performing normalization).
[0083] Table 1 Video Information
Table
id title cata name tag names release time
3714869 Wu lei has athletics athletics, 2019-08-01
become the European 10:07:32
Mr. Key, Championship,
winning odds Wu lei, sports
show Espanyol lottery, lottery
will bring information
surprise in the station, sports
new season lottery video
collection
[0084] The phase of obtaining user poi Li ait is a process of calculating
user word vector, and the
target user group directed is a group of active users, namely recently active
users (having
16
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
playback records within 7 days lately) with certain volume of playbacks
(exceeding 10
videos) within a period lately (such as the recent 30 days). Word vector
calculation of the
users is particularized according to tag categories, for instance, a user
played 100 videos
within a period lately, of which there are 60 relevant to athletics, 20
relevant to finance
and economics, 15 relevant to amusement, 4 relevant to the society, and 1
relevant to
health; during the process of user poll" __ ait, a poi ____________________ Li
ait is drawn for the user under tag
categories ranking top 3 in the proportions and with the proportions each
exceeding 10%,
through which method it is made possible to obtain the main interest points of
the user,
and to remove few operations by inadvertent clicking and hotspot videos that
cannot
represent the interest points of the user. In this example, athletics occupies
60%, finance
and economics occupy 20%, amusement occupies 15%, society occupies 4%, and
health
occupies 1%, so the portrait is drawn for the user in terms of the three
dimensions of
athletics, finance and economics, and amusement with respect to the current
user, and
word vector representations of the corresponding dimensions of the user are
calculated.
[0085] During the process of calculating user word vectors under different tag
categories of the
user, the time decay factor (for instance, the decay period is 5 days, the
decay coefficient
is 0.95, taking for example a video played back 12 days before the current
date, the video
crosses two decay periods, and should be decayed by 0.95^2) is combined to
calculate
the word vector representations of the user.
[0086] In the feature engineering constructing phase, taken into consideration
are the user word
vector (200 dimensions), video word vector (200 dimensions), proportion of a
category
watched by the user, proportions of historical playbacks of the user on a
hourly basis, user
gender, user age (divided according to the groups of over 20, 2030,¨ 30-
40, 40-50, and
over 50, and on-hot coded), current video classification tag, video time
duration (unit:
second), video releasing time (number of days from the current time), video
average
playback integrity (average playback integrity of videos played by the user
within the
recent 24 hours), hits level (divided into five levels according to the number
of playback
17
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
times, and one-hot coded), the time of the video watched by the user (which
day of the
week, the current time period, one-hot coded), position information (one-hot
coded
according to the Province), terminal type (one-hot coded), and operator type
(one-hot
coded).
[0087] The above features are constructed according to playback records of the
user within a
period lately (such as the recent 30 days), and the deep learning model is
trained in
combination with the video playback integrity of the user.
[0088] The possible playback integrity of videos not played back by the target
user is predicted
through the model with respect to the result set recommended to the user in
the recalling
phase, and inversion is carried out according to the playback integrity to
generate the final
recommended result set.
[0089] As should be noted, when the device for predicting video playback
integrity provided by
this embodiment performs a video playback integrity predicting business, the
division
into the aforementioned various functional modules is merely by way of
example, while
it is possible, in actual application, to base on requirements to assign the
functions to
different functional modules for completion, that is to say, to divide the
internal structure
of the device into different functional modules to complete the entire or
partial functions
described above. In addition, the device for predicting video playback
integrity provided
by this embodiment pertains to the same conception as the method of predicting
video
playback integrity provided by the method embodiment ¨ see the corresponding
method
embodiment for its specific realization process, while no repetition will be
made in this
context.
[0090] All the aforementioned optional technical solutions are randomly
combinable to form
optional embodiments of the present invention, to which no repetition is made
on a one-
by-one basis.
18
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0091] To sum it up, in comparison with prior-art technology, the method of
and device for
predicting video playback integrity provided by the embodiments of the present
invention
achieve the following advantageous effects.
[0092] 1. By modifying the traditional CTR prediction method altogether, a
video playback
integrity indicator is introduced, video playback integrities of different
users are predicted
through a well trained preset video playback integrity prediction model,
interest data of
users more approaching to reality is obtained in such important information
flow aspect
as the watching time duration through prediction results of the video playback
integrities,
the precision in recognizing user interests is enhanced, reality relevancy of
recommendations is hence enhanced, and relatively great enhancement is
achieved in the
watching time duration and satisfaction of users.
[0093] 2. Through vectorized representation of the user portrait, interest
transfer of the user is
reflected in combination with time decay of user behaviors, and hotspot videos
and
inadvertently clicked videos are filtered out in the process of user poll" __
ait, whereby
interference with actual interest of the user is avoided, and the user poi __
hait is made more
precise.
[0094] 3. By collecting such relevant data as user behavior data, video
quality and video
information, etc., vectorized representation of user features and video
attributes is
effectively made, as well as proportions of videos played back at various time
periods,
proportions of various categories and other environment information, different
features
and different data sources are merged in the application of short video
recommendation
sorting model through deep learning modeling and prediction of potential
playback
integrities of videos not watched by users, excellent effect is achieved, and
average
watching time duration of users is enhanced.
19
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0095] 4. By creating such features as user features, video features,
contextual features and client
side classification, deep learning modeling is employed, the playback
integrity prediction
mode is applied in a group of 10% randomly selected users through the AB test,
and such
indicators as CTR, daily average playback volume and user average playback
integrity
are compared through the final report. In the end, user average playback
integrity and
daily average playback volume are enhanced to a greater extent with slight
decrease in
the CTR.
[0096] 5. A TF-IDF algorithm is employed in terms of video recommendation, and
key
information of videos is effectively emphasized through IDF values.
[0097] 6. Reality relevancy of recommendations is enhanced through prediction
of short video
playback integrities, and increase in the time duration of stay of users is
attempted.
[0098] As understandable by persons ordinarily skilled in the art, realization
of the entire or
partial steps of the aforementioned embodiments can be completed by hardware,
or by a
program instructing relevant hardware, the program can be stored in a computer-
readable
storage medium, and the storage medium can be a read-only memory, a magnetic
disk, or
an optical disk, etc.
[0099] The embodiments of the present application are described with reference
to flowcharts
and/or block diagrams of the method, device (system), and computer program
product
embodied in the embodiments of the present application. As should be
understood, each
flow and/or block in the flowcharts and/or block diagrams, and any combination
of flow
and/or block in the flowcharts and/or block diagrams can be realized by
computer
program instructions. These computer program instructions can be supplied to a
general
computer, a dedicated computer, an embedded processor or a processor of any
other
programmable data processing device to form a machine, so that the
instructions executed
by the computer or the processor of any other programmable data processing
device
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
generate a device for realizing the functions designated in one or more
flow(s) of the
flowcharts and/or one or more block(s) of the block diagrams.
[0100] These computer program instructions can also be stored in a computer-
readable memory
enabling a computer or any other programmable data processing device to
operate by a
specific mode, so that the instructions stored in the computer-readable memory
generate
a product containing instructing means, and this instructing means realizes
the functions
designated in one or more flow(s) of the flowcharts and/or one or more
block(s) of the
block diagrams.
[0101] These computer program instructions can also be loaded onto a computer
or any other
programmable data processing device, so as to execute a series of operations
and steps
on the computer or the any other programmable device to generate computer-
realized
processing, so that the instructions executed on the computer or the any other
programmable device provide steps for realizing the functions designated in
one or more
flow(s) of the flowcharts and/or one or more block(s) of the block diagrams.
[0102] Although preferred embodiments in the embodiments of the present
application have
been described so far, it is still possible for persons skilled in the art to
make additional
modifications and amendments to these embodiments upon learning of the basic
inventive
conception. Accordingly, the attached Claims are meant to cover the preferred
embodiments and all modifications and amendments that fall within the scope of
the
embodiments of the present application.
[0103] Apparently, persons skilled in the art can make various amendments and
modifications to
the present invention without departing from the spirit and scope of the
present invention.
Thus, if such amendments and modifications to the present invention fall
within the
Claims of the present invention and equivalent technology, the present
invention is also
meant to cover these amendments and modifications.
21
Date Recue/Date Received 2022-03-07
CA 03153598 2022-03-07
[0104] What is described above is merely directed to preferred embodiments of
the present
invention, and they are not meant to restrict the present invention. Any
amendment,
equivalent replacement and improvement makeable within the spirit and scope of
the
present invention shall all be covered within the protection scope of the
present invention.
22
Date Recue/Date Received 2022-03-07
